Category: Digital Services: Google

Eventually You Run Out of Other People’s Content: How Google May Win the AI Race But Lose on the AI Economics

[This post originally appeared on the True Enough Substack]

The greatest threat to Google isn’t necessarily that OpenAI, Anthropic, or another competitor will defeat it in artificial intelligence. It’s that Google may successfully defend its market position while destroying the extraordinary economics that made its search monopoly so valuable in the first place.

For more than two decades, Google operated one of the most capital-efficient businesses in corporate history. It organized other people’s work product, attracted consumers searching for those brands, sold advertising against their intentions, and generated enormous amounts of free cash flow.

Thanks in significant part to statutory safe harbors that hurt everyone but Google, the underlying content was largely somebody else’s expense. Publishers wrote the articles. Universities funded the research. Songwriters wrote and musicians recorded the songs financed by record companies and music publishers. Businesses produced technical documentation. Google indexed that material and monetized the process of finding it.

Google certainly spent substantial sums developing its search infrastructure, but it didn’t have to bear the cost of creating most of the content that made Search useful. And once the infrastructure and advertising marketplace were established, incremental revenue could be generated at relatively low incremental cost.

Artificial intelligence threatens to reverse that equation.

Instead of directing consumers to content somebody else financed and created, Google is increasingly using expensive computing infrastructure to generate answers itself (scraping content somebody else financed and created). The underlying information may still come from other people’s work, but Google now incurs the computational expense of manufacturing the answer.

That requires significant capital expenditure on data centers, specialized processors, electricity, cooling, memory, model development, and continuous reinvestment in rapidly evolving technology. And that means Google is transitioning from a relatively low-capital-expenditure, high-moat business to a high-capital-expenditure business with potentially weaker competitive advantages.

And its latest financial statements reveal just how expensive that transition has become.

The $53.5 Billion Free-Cash-Flow Collapse

The most revealing number in Alphabet’s third-quarter 2026 financial results isn’t revenue, earnings per share, or even capital expenditures. It’s free cash flow.

During the first nine months of 2025, Alphabet generated approximately $112.3 billion in operating cash flow and spent $63.6 billion on capital expenditures. That left approximately $48.7 billion in free cash flow. During the corresponding period in 2026, operating cash flow actually increased to $127.1 billion.

But capital expenditures more than doubled to approximately $132 billion. The result was negative free cash flow of approximately $4.8 billion.

Think about that.

Alphabet’s operating cash generation improved by nearly $15 billion, yet its free cash flow deteriorated by approximately $53.5 billion. The company’s cash-generating engine was performing better. But the cost of financing its future was consuming more cash than that engine produced.

The quarterly progression is particularly revealing. Alphabet generated positive $10.1 billion in free cash flow in the first quarter of 2026, followed by negative $5.9 billion in the second quarter and negative $9.1 billion in the third.

To be clear, these are Alphabet-wide figures, not just AI. The capital expenditures support Google Cloud and other activities as well as Search, and investments made today may produce substantial future revenue. But that qualification doesn’t eliminate the economic issue. I think it sharpens it quite a bit.

The fundamental question is no longer how much revenue Google can generate. It’s how much capital Google must continually invest to generate that revenue—and how much cash remains for shareholders afterward. Old Google expanded by reinvesting a portion of its extraordinary cash flow. New Google is increasingly consuming that cash flow to build the infrastructure necessary to remain competitive.

That’s not merely a change in spending priorities. It may be a fundamental change in the quality of the business.

The Equity Offering That Changed the Story

In June 2026, Alphabet raised approximately $49.6 billion in net proceeds through a combination of common and mandatory convertible preferred stock, including a private placement with Berkshire Hathaway.

Alphabet expressly identified investments in AI infrastructure and computing capacity among the purposes of the financing. That’s kind of a man bites dog story.

It also established a $40 billion at-the-market equity issuance program, selling another $12.4 billion of common stock during the third quarter. The stated primary purpose of that ATM program was to meet tax obligations associated with employee equity grants, rather than directly finance AI infrastructure.

Meanwhile, Alphabet continued issuing substantial amounts of debt.

The June financing is especially significant because Google historically generated enough cash to finance its expansion internally while repurchasing billions of dollars of its own shares not issuing new shares to finance capex. Now the company has turned to outside capital to supplement that cash generation during its AI infrastructure expansion.

This does not establish that Alphabet was running out of money. It has enormous resources, and raising capital when financing is available on attractive terms may be perfectly rational.

But the reversal is striking.

Google historically generated extraordinary free cash flow without paying dividends, instead returning capital primarily through share repurchases. Now the economics of AI are forcing a reassessment of how much of that cash the company can retain, reinvest, or return to shareholders. That changes the shareholder proposition.

Capital-intensive growth financed partly through new equity must produce sufficient incremental returns not only to justify the investment but also to offset the dilution of existing shareholders.

And a business that repeatedly needs large new investments to maintain its competitive position is economically different from one whose existing infrastructure can support substantial growth with comparatively modest incremental capital.

The critical issue isn’t solvency. It’s whether Alphabet’s legendary ability to convert revenue into cash is becoming a casualty of the AI arms race.

Google Used to Monetize Other People’s Work. Now It Must Manufacture Answers.

Traditional Google Search benefited from an almost ideal division of labor. Google organized access to other people’s work product. That arrangement allowed Google to capture extraordinary value from work it did not itself finance.

A newspaper might spend thousands of dollars investigating a story. Google could index the resulting article and sell advertising to users searching for the subject. An academic might spend years developing specialized expertise. Google could organize access to the resulting research. The creator absorbed the cost of producing the underlying information. Google monetized “discovery.”

AI search changes the relationship. When Gemini generates a detailed answer, Google must supply the computing capacity necessary to produce that response. More sophisticated reasoning and increasingly complex multimodal interactions can require substantially more resources than retrieving and ranking search results.

And Google must make those investments repeatedly as competing AI systems improve. Of course, Google isn’t literally creating all the information contained in its AI answers. Much of that information still originates in third-party content that Google scrapes. Google also has long operated content-intensive businesses, notably YouTube relying on the safe harbors in Section 230 and the DMCA (which is why they push for text and data mining exceptions and why they are litigating over fair use).

But AI search makes Google more directly responsible for producing the answer presented to the consumer, rather than simply organizing access to an answer located elsewhere. That is a crucial distinction and plays directly into the economics.

Google once monetized content it didn’t have to create. Now it must pay to generate answers based substantially on content it still didn’t create.

There’s also a long-term complication. If AI-generated answers reduce visits to the publishers and other creators who supply the underlying information, the economics supporting the creation of that information may deteriorate. Google could end up weakening the very content ecosystem on which both its traditional search engine and its AI systems depend.

The Difference Between Capital Expenditure and Capital Formation

There is another distinction investors should keep in mind. Spending $200 billion on infrastructure does not necessarily create $200 billion in durable economic value. Alphabet expects capital expenditures of approximately $195 billion to $205 billion in 2026, with further substantial increases anticipated in 2027. Some of this spending will produce assets capable of generating revenue for years. But AI infrastructure is not equivalent to a long-lived monopoly franchise like Google’s.

Processors become obsolete. Models improve. Competing systems become more efficient. Hardware must be replaced. Customers may migrate toward cheaper or better-performing alternatives. And today’s enormous investment doesn’t necessarily prevent competitors from making tomorrow’s enormous investment.

That raises the possibility that some portion of AI capital expenditures will become a continuing competitive obligation rather than a one-time investment in growth.

Alphabet’s depreciation on property and equipment already increased from approximately $15.1 billion to $22 billion during the first nine months of 2026. As additional data centers and computing equipment enter service, depreciation and operating costs will continue to flow through future income statements. The cash goes out first. The accounting expense follows.

The investor’s question should be straightforward:

What return will Google earn on its next dollar invested in AI compared with the return it earned on its last dollar invested in traditional Search?

If that incremental return declines substantially, Google can grow revenue and expand its technological capabilities while becoming a less valuable business on a return-on-capital basis.

That is the real risk.

Search Revenue Is Slowing as the Cost of Defending It Accelerates

Back to Google’s earnings, Google Search revenue growth slowed from approximately 19% in the first quarter of 2026 to 17% in the second quarter and 12% in the third. Search remains enormously profitable. Indeed, Google reported strong query growth, suggesting that demand for its search services is not disappearing. But the divergence between rising engagement, decelerating revenue growth, and rapidly increasing infrastructure investment deserves attention.

AI may be expanding the number of queries while changing the economics of monetizing them. Traditional Search benefits from a mature advertising auction that converts commercial intent into revenue. AI search introduces a different consumer experience, in which users increasingly receive synthesized answers rather than navigating a series of links and sponsored results.

Google is developing advertising formats for AI, and those products may ultimately be successful. But the revenue model remains less established than the traditional search auction. And the cost per interaction may be higher.

The danger is that Google could retain users while generating less economic value from serving them. Even if AI search eventually produces comparable advertising revenue per query, it may require substantially more computing expenditure.

That is the cannibalization problem: Google’s innovation may preserve its market share while reducing the cash generation associated with that market share.

A Weaker Moat at a Much Higher Price

Google’s traditional Search moat consisted of several mutually reinforcing advantages: consumer habits, default distribution, search quality, advertiser relationships, scale, and its dominant advertising marketplace.

A competitor couldn’t simply build a good search engine. It had to recreate the commercial ecosystem. AI changes that dynamic.

Consumers can now obtain useful AI answers from Gemini, ChatGPT, Claude, Perplexity, and other AI systems. The products are not identical, and Google retains important advantages in distribution, infrastructure, proprietary chips, data, and advertising.

But conversational AI makes the basic information-retrieval experience more substitutable. The competitive struggle increasingly centers on model quality, speed, cost, and capabilities—areas where rivals can improve rapidly and technological leadership may be temporary.

This doesn’t mean AI has no barriers to entry. Frontier-model development and large-scale inference require enormous resources. But an expensive barrier to entry is not necessarily a profitable moat.

If every major competitor must continually spend billions to maintain technological parity, the result may be an exceptionally capital-intensive industry without the pricing power necessary to justify those investments.

The distinction is between spending capital to establish a durable competitive advantage and spending capital simply to remain in the game. Google’s historical Search business did the former extraordinarily well. AI threatens to make the latter a permanent condition of competition.

The Strongest Counterargument: Google Cloud

There is an obvious counterargument. Alphabet’s AI investments aren’t limited to defending Search. Google Cloud revenue increased 103% in the third quarter of 2026 to approximately $30.7 billion. Alphabet can monetize computing infrastructure through enterprise customers, AI services, subscriptions, and sales of its proprietary TPU systems.

A substantial part of the capital expansion may therefore support genuinely new revenue rather than simply replace existing Search capabilities. That’s a significant counter, and any serious financial analysis must account for it.

Google may ultimately achieve attractive returns on its AI investments across several different businesses. But investors should not assume that rapid Cloud revenue growth establishes that the capital expenditure is earning adequate returns.

Revenue growth is not return on invested capital. The question is whether the combined additional profits from Cloud, AI subscriptions, and AI-enhanced Search will justify the enormous incremental capital committed to obtaining them.

Alphabet management says it expects attractive returns on invested capital above its cost of capital and that aspiration may be realized. But in the meantime, the sharp decline in free cash flow is real, the equity financing is real, and the depreciation burden from new infrastructure is growing.

The financial statements show the cost of the bet. They have not yet demonstrated its ultimate return.

Winning the AI Race May Be the Wrong Goal

Investors are understandably fascinated by whether Gemini can outperform ChatGPT or Claude. This is the normal board room chatter. But those comparisons may distract from the more fundamental economics.

A company can develop excellent technology, maintain its market position, increase revenue, and still deliver disappointing shareholder returns if achieving those results requires too much capital. That’s the real danger Alphabet faces.

Google’s original Search franchise generated extraordinary returns because the underlying business required comparatively little incremental investment relative to the revenue and cash it produced.

The transition to AI is dramatically increasing capital requirements while potentially weakening the competitive barriers that supported those returns.

The first nine months of 2026 illustrate the transformation: operating cash flow increased by nearly $15 billion while free cash flow declined by approximately $53.5 billion.

Alphabet supplemented that cash generation with major equity and debt financing. And management anticipates that capital spending will rise further. That means The stakes are much larger than quarterly earnings or market share.

Google may be exchanging a business that generated extraordinary returns on relatively modest incremental capital for one that requires extraordinary amounts of capital merely to remain competitive.

Conclusion: The Economics Matter More Than the Technology

Google doesn’t have to lose the AI race to lose the economic advantages of its original business. It might remain a leader in artificial intelligence, preserve its enormous customer base, build outstanding products, and continue generating substantial revenue.

And yet shareholders could find themselves owning a company that requires vastly more capital to produce each incremental dollar of free cash flow. That would represent a fundamental deterioration in the quality of Google’s business, even if its technological position remained strong.

The old Google monetized other people’s work through an extraordinarily efficient discovery and advertising system. The new Google is spending hundreds of billions to build the computational machinery necessary to generate answers, defend its distribution advantages, and compete against other well-funded technology companies doing essentially the same thing.

That may be an impressive technological achievement, but the measure of its economic success won’t be how many answers Gemini produces, how large Google’s models become, or whether Google wins the latest AI benchmark.

It will be how much free cash flow remains after paying for the infrastructure required to produce those answers.

Google’s greatest AI risk isn’t losing the race. It’s winning the race and discovering that the victory has destroyed the economics that made Search one of the greatest businesses ever created.

For further reading: Alphabet, Form 10-Q for quarter ended September 30, 2026, Alphabet, Q3 2026 earnings release, and Alphabet, Q2 2026 earnings release.

Bankruptcy Shield as a Sword: The Startup Gets Acquired Without Being Acquired

The proposed sale of Spirit Airlines’ enormous internal dataset to Google exposes a potentially dangerous new loophole in the AI economy. Google won a bankruptcy auction with a $10 million bid for Spirit’s enterprise data and says the material could help improve its products and AI models. The corpus reportedly includes approximately 100 million emails, 500 million Microsoft Teams messages, software code, operational information and decades of business records. (Reuters)

But there is a threshold question that shouldn’t get lost in the excitement over turning bankrupt companies into AI training datasets: Did Spirit actually own everything it proposed to sell? (For BK fans, Spirit’s liquidation is proceeding under Chapter 11, not Chapter 7. See In re Spirit Aviation Holdings, Inc., No. 25-11897 (SHL) (Bankr. S.D.N.Y.). As part of that proceeding, the debtors auctioned a deidentified enterprise dataset on August 14, 2026, selecting Google’s $10 million bid, although approval of the sale remains contested. Chapter 11 isn’t limited to reorganizations. It can be used for an “orderly liquidation,” with § 363 sales disposing of assets while the debtor remains in Chapter 11.)

Possession Is Not Ownership

Springshot, a technology startup whose proprietary airline-operations platform was embedded in Spirit’s technology stack for its final three years, says potentially not. Springshot has objected that the sale agreement defines the dataset so broadly that it may encompass Springshot’s own intellectual property and proprietary information. It wants the bankruptcy court to require a forensic process separating Spirit’s property from property belonging to third parties before anything goes to Google. (Ars Technica)

Springshot founder Doug Kreuzkamp states the basic principle perfectly: “The possession of IP is not ownership.” After all, a debtor’s server is a container, not a chain of title. It may hold the debtor’s property alongside licensed technology, vendor trade secrets, confidential communications, employee information and intellectual property belonging to somebody else. Bankruptcy shouldn’t improve the debtor’s title or expand the debtor’s assets just because someone else’s asset happens to be stored in the wrong place.

Bankruptcy as a Sword

That becomes particularly important with AI because bankruptcy can potentially be transformed from a shield for a distressed debtor into a sword for the purchaser when the debtor’s assets are auctioned off in liquidation. The sequence is straightforward: Startup develops technology → licenses it to customer → technology and know-how become embedded in customer’s systems → customer fails → bankruptcy estate sells its “dataset” → AI company trains on the corpus.

The bankruptcy proceeding has now done something much more consequential than liquidating the debtor’s assets. It may have created an efficient mechanism for aggregating information and transferring it to an AI company for a new use. That’s the loophole.

The Startup Gets Acquired Without Being Acquired

And that’s where this becomes an antitrust problem. Springshot has spent roughly 15 years developing technology used to improve airline operations. Ars reports that Google recently announced a five-year partnership with Ryanair involving airline operational data and Gemini Enterprise—putting Google’s AI efforts in an area where Springshot itself operates. Springshot consequently fears that its proprietary information could end up helping a vastly larger company compete with it.

Think about the alternatives. Google could develop competing technology. It could license Springshot’s technology. It could partner with Springshot. It could potentially acquire Springshot.

Or, potentially: Buy the data estate of Springshot’s bankrupt customer in a bankruptcy auction. That last possibility is profoundly different. Google doesn’t acquire the startup. It potentially acquires what the startup knows. There is no acquisition price paid to the startup. No negotiated license. And no conventional acquisition of the competitor to scrutinize.

As Springshot warns, the precedent could allow large amounts of startup IP to migrate to enormously powerful companies through bankruptcy proceedings. In other words, the startup gets acquired without being acquired. And there may be nothing they can do about it.

In-RE-Spirit-Aviation-Holdings-International-Aero-Engines-8-27-26Download

A Competition Loophole

That makes the bankruptcy loophole a competition loophole. Startups necessarily expose substantial information to enterprise customers. Integration requires documentation, workflows, support communications, operational data and technical knowledge.

Normally contracts, confidentiality obligations, trade-secret law and property rights establish boundaries around that information. But if those boundaries become uncertain when the customer goes bankrupt, the economics of dealing with large enterprise customers change.

A startup could survive while its customer fails—only to discover that the customer’s bankruptcy estate proposes to sell information embodying the startup’s technology to a company with virtually unlimited resources. The customer went bankrupt, so why should the startup’s competitive advantage be liquidated with it?

Procedure Becomes the Product

There is a broader procedural inversion here. Large corporate defendants have long understood that legal procedures nominally directed against them can sometimes produce valuable affirmative rights. A class action, for example, is ostensibly a device for aggregating claims against a defendant. But settlement can also aggregate something enormously valuable for the defendant: releases and finality.

The procedure itself creates economic value. AI may give bankruptcy proceedings a similar secondary function. Bankruptcy ostensibly protects the debtor and organizes an orderly disposition of its property. Or that’s the party line. But to an AI company seeking enormous quantities of authentic enterprise data, bankruptcy can provide something else: Aggregation.

Instead of negotiating separately with every employee, contractor, vendor and technology supplier whose information resides in a company’s systems, the purchaser buys one enormous “dataset” from the estate. What began as the debtor’s shield risks becoming the purchaser’s sword.

De-Identification Doesn’t Fix Title

Google says it will not receive personal information, and Spirit’s dataset is supposed to undergo de-identification. That’s important for privacy, but it doesn’t resolve ownership. Anonymization is not assignment. Removing a person’s name doesn’t transfer a vendor’s trade secrets. And de-identifying employee communications doesn’t necessarily establish that they can be repurposed for AI training. Privacy, ownership, confidentiality and authorization are separate questions.

A Caution for Startups: Treat Customer Data Like Inventory in a Warehouse

There is a useful analogy here to secured lending. A bank lending against inventory does not assume that everything physically sitting in the borrower’s warehouse belongs to the borrower. It identifies the collateral, determines what rights the borrower actually has in it, perfects its security interest, and worries about priority, proceeds, commingling and property belonging to third parties. So the Spirit case highlights the need to think about creditworthiness as a material deal point.

AI startups may want to start thinking about their data in much the same way. A customer’s servers are increasingly the digital equivalent of the warehouse. They may contain customer-owned information alongside licensed technology, vendor trade secrets, proprietary workflows, confidential communications, derived data and other material the customer merely possesses. If the customer enters bankruptcy, the danger is that all of this gets swept into a broadly defined “enterprise dataset” and auctioned as though possession established ownership. It doesn’t.

That means startups should protect themselves before a customer’s financial distress ever arises. Contracts should clearly establish ownership, permitted uses, AI-training restrictions, return and deletion obligations, and what happens upon insolvency. Proprietary information should be identifiable and segregated where practicable, with sufficient provenance to establish what belongs to whom.

There may even be circumstances where traditional secured-credit techniques deserve consideration as additional protection. But the distinction is important: Ownership is the first line of defense: “That isn’t property of the bankruptcy estate.” A security interest is the second: “If it is property of the estate, our rights have priority.”

A startup generally wants to win the first argument. Carelessly taking a security interest in property the startup claims to own could muddy that position rather than strengthen it, so Article 9 protection requires careful drafting.

The larger caution is simple: Bankruptcy courts have spent generations distinguishing what is in a debtor’s warehouse from what the debtor actually owns. AI should not erase that distinction merely by replacing the warehouse with a server. For startups whose technology becomes embedded in customers’ systems, data-title planning may now need to become as routine as IP assignments, confidentiality provisions and cybersecurity. The time to establish ownership and provenance is before the customer’s servers—and everything sitting inside them—show up in a bankruptcy auction.

The Rule Should Be Simple

AI has transformed old corporate records from storage liabilities into valuable assets. Spirit’s data sale shows just how valuable: companies are bidding millions for records of how real humans communicate, solve problems and operate businesses. That makes it essential to get the rule right now. Before a bankruptcy court asks what is this dataset worth?, it should ask: What exactly did the debtor own?

Bankruptcy can transfer the debtor’s assets. It should not manufacture AI rights the debtor never possessed. And it certainly shouldn’t become a mechanism through which a dominant AI company can acquire the accumulated knowledge of a startup without ever acquiring the startup itself.

Ars Technica’s report on the Spirit/Springshot dispute

As AI Infringement Claims move to the C-Suite and Board Room, Plaintiffs should Follow the Money to Wire Fraud, Fiduciary Duty and RICO

AI training litigation is moving from copyright pleadings to governance pleadings. The next discovery fight should follow the money all the way to the C-suite and if necessary, to the board room.

The complaint in Elsevier Inc. v. Meta Platforms, Inc., No. 1:26-cv-03689 (S.D.N.Y. filed May 5, 2026) (the “Elsevier complaint”), alleges that Meta and Mark Zuckerberg personally illegally torrented millions of copyrighted books and journal articles from notorious pirate sites, copied those works repeatedly to train Llama, and did so with knowledge that the conduct violated copyright law. 

2026-05-05-ComplaintDownload

Alternatively, the shareholder derivative complaint in SEIU Pension Plan Master Trust v. Narayen, No. 3:26-cv-03521 (N.D. Cal. filed Apr. 24, 2026) (the “SEIU complaint”), shows the same issue also moving into the boardroom: it alleges that Adobe’s officers and directors adopted and implemented an unlawful AI business strategy by using copyrighted material to develop Adobe’s AI services, exposing Adobe to litigation, reputational harm, and corporate loss. 

Both cases suggest a practical discovery vector for plaintiffs in AI-training cases: if the defendant used Anna’s Archive, LibGen, Z-Library, Sci-Hub, Books3, RedPajama, SlimPajama, or any similar pirate-derived source, plaintiffs should investigate whether the defendant merely downloaded available files or also paid for bulk access, priority access, SFTP credentials, “membership” tiers, “donations,” vendor pass-through datasets, or intermediary transfers. 

SEIU_Pension_Plan_Master_Trust-Shareholders-v.-Narayen-April-24-2026Download

The discovery question: was there a payment trail?

The current Elsevier complaint is framed principally as a copyright and DMCA case, not as a wire-fraud or shareholder derivative complaint. But its allegations make the payment question worth asking because it alleges that Meta downloaded Anna’s Archive, understood Anna’s Archive to be “essentially a bigger libgen” and “a pretty shady website,” acquired more than 81 terabytes of data through Anna’s Archive, and did not disable BitTorrent’s default distribution settings when torrenting from pirate sites. The complaint further alleges that Meta’s logs showed 134.6 TB downloaded and 40.42 TB uploaded through torrenting between April and July 2024.

It also alleges that Meta considered licensing literary works from major publishers in the training data market, discussed increasing its dataset licensing budget from $17 million to $200 million, then stopped licensing efforts after the license-versus-pirate question was escalated to Zuckerberg. Those allegations support discovery into whether Meta, its employees, contractors, affiliates, or data vendors ever communicated with, paid, negotiated with, or obtained credentials from Anna’s Archive or a similar bulk pirate-data provider. 

The public record concerning Anna’s Archive gives that discovery question additional force. As we covered before on MTP, Anna’s Archive reportedly charged tiered “membership” fees for faster download speeds, accepted cryptocurrency or gift cards, and offered AI companies “enterprise-level” high-speed SFTP transfers of a full 1.1-petabyte collection for a reported $200,000 in cryptocurrency. The complaint in Apress Media, LLC et al. v. Anna’s Archive et al., No. 1:26-cv-01850 (S.D.N.Y. filed Mar. 6, 2026) (the “Apress complaint”), similarly alleges that Anna’s Archive invited LLM developers to copy the entire collection for free or make a “donation” for faster download speeds. Publishers Weekly’s account of the Apress complaint states that Anna’s Archive publicly claimed to have provided high-speed access to its illegal collection to companies in China, Russia, and elsewhere, many of them LLMs, and that an email exchange quoted in the complaint offered premium access for $200,000 with payment suggested in cryptocurrency. It is important to note that as much as we may believe that these people are all scumbags, these allegations do not prove that any particular AI defendant paid Anna’s Archive, but they do make paid-access discovery reasonable where a defendant is already alleged to have used Anna’s Archive or comparable pirate repositories. As usual, facts matter.

Why payment alone is not wire fraud

The conditional nature of the theory also matters. Wire fraud is not “copyright infringement plus the internet,” and it is not “payment to a bad actor plus wires.” 18 U.S.C. § 1343 requires a scheme or artifice to defraud, or a scheme to obtain money or property by false or fraudulent pretenses, representations, or promises, plus interstate or foreign wire communications used for the purpose of executing that scheme.  DOJ’s formulation similarly requires voluntary and intentional participation in a scheme to defraud, intent to defraud, reasonably foreseeable use of interstate wire communications, and actual use of interstate wire communications.  DOJ also explains that the fraudulent aspect of a scheme to defraud is measured by nontechnical standards and generally involves wrongdoing in property rights by dishonest methods, trick, chicane, or overreaching. 

That is why a payment to a pirate site is not automatically wire fraud in a formal sense. A company may knowingly pay for unlawful access to copyrighted content, and that payment may be powerful evidence of willfulness, commercial purpose, knowledge, damages, or fiduciary misconduct. But if the buyer and seller both understand that the transaction is a purchase of illicit access and no materially false representation is used to obtain money, data, payment processing, procurement approval, compliance clearance, tax treatment, or concealment from a relevant victim or gatekeeper, the payment itself may not satisfy the fraud element. The discovery target should therefore likely be the deceptive aspect and proves up whether the payment was disguised as a “donation,” “membership,” “research access,” “preservation support,” “lawful dataset,” “vendor service,” “data license,” or other label that misrepresented the transaction’s purpose, legality, source, recipient, or quid pro quo.  And given who we’re dealing with, could even be designed to deceive internal accountants and co-signers depending on corporate check-writing policies.

Deceptive labeling can supply the fraud layer that ordinary infringement lacks. As we covered before on MTP, Anna’s Archive’s alleged “donation” model can be characterized as a commercial SFTP pipeline for stolen works, and the Apress complaint alleges that Anna’s Archive invited LLM developers to make a “donation” for faster download speeds. If discovery in a particular AI case shows that an AI developer, contractor, or data broker used wires, emails, SFTP credentials, cryptocurrency transactions, payment processors, invoices, procurement records, or vendor documentation to disguise a purchase of pirate training data as something lawful or altruistic, the facts may support a wire-fraud predicate theory. If discovery shows only an infringing download or torrent, the same evidence may still matter enormously to copyright liability, willfulness, damages, concealment, and fiduciary duty, but it should not be overstated as wire fraud without proof of deception and, of course, all the wire fraud elements. 

The Potential Civil RICO Angle

Civil RICO is a possible overlay, but not a shortcut and proving it up may be adjacent but separate to other claims. 18 U.S.C. § 1962(c) makes it unlawful for a person associated with an enterprise engaged in or affecting interstate commerce to conduct or participate in the conduct of the enterprise’s affairs through a pattern of racketeering activity (or collection of unlawful debt).  DOJ summarizes an 18 U.S.C. § 1962(c) violation as requiring conduct, of an enterprise, through a pattern, of racketeering activity. RICO’s definition of racketeering activity includes wire fraud under 18 U.S.C. § 1343 and criminal copyright infringement under 18 U.S.C. § 2319. Criminal copyright infringement requires a valid copyright, infringement, willfulness, and commercial advantage or private financial gain. 18 U.S.C. § 2319 provides felony penalties for certain offenses involving reproduction or distribution, including by electronic means, of at least ten copies or phonorecords of one or more copyrighted works with a total retail value of more than $2,500 during a 180-day period. Although criminal copyright infringement cases are not common, one has to ask if AI scraping of millions and millions of works is not criminal infringement, what is?

But RICO also requires relationship and continuity. A “pattern of racketeering activity” requires at least two predicate acts, but the Supreme Court has held that two predicates are not necessarily sufficient because a plaintiff or prosecutor must show that the predicates are related and amount to, or threaten, continued criminal activity. H.J. Inc. v. Northwestern Bell Telephone Co., 492 U.S. 229, 237–39 (1989). Relatedness can be shown where the acts share similar purposes, results, participants, victims, or methods of commission, or are otherwise interrelated and not isolated events.  Continuity can be closed-ended or open-ended, and the DOJ RICO guidance describes continuity as either a closed period of repeated conduct or past conduct that by its nature projects into the future with a threat of repetition.  A civil RICO plaintiff must also show injury to business or property by reason of a violation of 18 U.S.C. § 1962, and 18 U.S.C. § 1964(c) provides treble damages, costs, and attorney’s fees for such injury. 

Judge Mark C. Scarsi’s ruling in Perry v. Shein Distribution Corp. is useful because Judge Scarsi rejected the idea that large-scale copying must be treated as ordinary copyright infringement only. The court allowed independent designers’ civil RICO claims against the “fast fashion” Chinese retailer Shein to proceed, reportedly finding that the plaintiffs plausibly alleged a coordinated enterprise using copyright infringement and mail/wire fraud predicates as part of a broader scheme. The case later settled, so it is valuable as a pleading-stage roadmap.

The connection to Elsevier v. Meta is straightforward. Elsevier and other publishers allege that Meta copied millions of books and journal articles, used pirated libraries and web-scraped datasets, trained Llama on those works, and removed copyright-management information. If plaintiffs can show this was not ad hoc infringement but a coordinated corporate program approved at senior levels including Zuckerberg as they allege, using piracy, concealment, distribution, and monetization, the Shein ruling on RICO supports the argument that the conduct resembles an enterprise scheme rather than isolated copyright violations. The theory is when infringement is systematic, repeated, operationalized, concealed, and tied to enterprise monetization, RICO should not be dismissed merely because copyright is also involved in the defendant’s bad behavior.

Shareholder Derivative Action

A derivative claim does not need to prove that a payment to a pirate site was wire fraud; it can focus on whether fiduciaries knowingly caused or permitted the company to pursue AI profits through unlawful or legally reckless data acquisition. Delaware law does not charter lawbreakers because the DGCL authorizes Delaware corporations to pursue lawful business and lawful acts, 8 Del. C. §§ 101(b), 102(a)(3), and Delaware courts have stated that a fiduciary cannot be loyal to a Delaware corporation by knowingly causing it to seek profit by violating the law. In re Massey Energy Co. Derivative & Class Action Litigation, 2011 WL 2176479, at *20 (Del. Ch. May 31, 2011); see also Metro Communication Corp. BVI v. Advanced Mobilecomm Technologies Inc., 854 A.2d 121, 131, 163–64 (Del. Ch. 2004); Guttman v. Huang, 823 A.2d 492, 506 (Del. Ch. 2003).  

The same payment evidence may be even more immediately useful in shareholder derivative litigation. (Although the derivative theory is not a claim to be stapled onto a copyright-owner complaint. It is a separate governance claim for a different plaintiff and a different injury.). Under Delaware law, the board manages the corporation’s business and affairs, 8 Del. C. § 141(a), and a stockholder derivative action seeks to assert a corporate claim when demand is excused or wrongfully refused. See United Food & Commercial Workers Union & Participating Food Industry Employers Tri-State Pension Fund v. Zuckerberg, 262 A.3d 1034, 1047, 1059 (Del. 2021).

Delaware courts have also explained that directors must make a good-faith effort to implement and monitor systems that keep the board informed about legal-compliance risks, and that personal liability may follow when fiduciaries act in bad faith by utterly failing to implement such systems or consciously failing to monitor them. Stone v. Ritter, 911 A.2d 362, 370 (Del. 2006); Marchand v. Barnhill, 212 A.3d 805, 820–21 (Del. 2019). Delaware officers also owe context-specific oversight duties within their corporate remit. In re McDonald’s Corp. Stockholder Derivative Litigation, 289 A.3d 343, 359–61, 369–70 (Del. Ch. 2023). In that framework, a payment trail to a pirate repository may support allegations of knowing illegality, bad faith, internal-control failure, waste, disclosure failure, or conscious disregard of red flags. 

The SEIU complaint illustrates how quickly AI copyright allegations can become governance allegations. It alleges that Adobe is a Delaware corporation and that the defendants owed Adobe fiduciary duties of care, loyalty, good faith, diligence, fair dealing, and supervision.  The complaint alleges that Adobe used the SlimPajama dataset, which was derived from RedPajama, and that the dataset contained copyrighted works not authorized or approved by authors and copyright holders. It alleges that Adobe told the market that Firefly was commercially safe, trained on licensed content, trained on data Adobe had rights to use, and designed to respect creator rights and avoid infringing third-party intellectual property. It then alleges that those statements were false or misleading because Adobe’s AI products allegedly depended on SlimLM datasets that included pirated materials. 

The SEIU complaint also pleads the red-flag story in governance terms. It alleges that Adobe’s officers were personally involved in Adobe’s AI strategy and dataset choices, that multiple AI copyright lawsuits against competitors put all defendants on notice, and that Adobe was using datasets associated with Books3, RedPajama, the Pile, and SlimPajama. It alleges that copyright holders filed two class actions against Adobe, that Adobe’s stock dropped by more than 25% after those filings, and that Adobe faces litigation costs, potential liability, reputational harm, lost customers, and other corporate injuries. It also alleges that Adobe wasted corporate assets by paying compensation and bonuses, repurchasing shares at allegedly inflated prices, and incurring legal liability and costs associated with the copyright class actions. 

That is exactly why payment discovery matters to derivative plaintiffs. If corporate funds, procurement systems, crypto wallets, reimbursement requests, vendor invoices, or contractor payments were used to purchase pirate datasets, then the derivative theory becomes less abstract. The evidence would speak not only to whether copyrighted works were used, but to who approved the acquisition, how it was booked, what compliance review occurred, what the board or audit committee was told, whether the source was concealed, and whether public statements about “licensed,” “commercially safe,” or “rights-cleared” AI products were misleading.  If a payment was mislabeled as a donation, membership, research support, vendor service, or lawful data license, that same fact could support both a fraud-oriented discovery path and a fiduciary-duty theory focused on bad faith, oversight failure, disclosure controls, and corporate waste. 

What plaintiffs could ask for

Plaintiffs should ask for all communications with Anna’s Archive, LibGen, Z-Library, Sci-Hub, Books3 distributors, shadow-library mirrors, dataset curators, data brokers, contractors, and AI-training data vendors.  They should ask for payment records, cryptocurrency wallet addresses, exchange records, gift-card purchases, reimbursement requests, procurement records, vendor invoices, purchase orders, data-source approvals, security reviews, legal-risk memoranda, audit-committee materials, board presentations, and employee messages about “donations,” “memberships,” “enterprise access,” “SFTP,” “fast downloads,” “bulk transfer,” “shadow libraries,” “pirate datasets,” or “rights-cleared” alternatives. They should also ask for logs showing whether data was received by torrent, direct download, SFTP, cloud transfer, physical drive shipment, contractor delivery, or a repackaged dataset from an intermediary. 

In a copyright case, those materials may bear on copying, distribution, willfulness, CMI removal, concealment, damages, and market substitution. In a wire-fraud/RICO overlay, those materials may bear on whether any payment was part of a deceptive scheme executed through wires, whether the alleged predicates are related and continuous, and whether the plaintiff can show injury to business or property by reason of a RICO violation.  In a derivative suit, those materials may bear on whether officers or directors knowingly caused the company to violate law, ignored red flags, failed to maintain adequate reporting and compliance systems, approved misleading disclosures, or wasted corporate assets. See 8 Del. C. §§ 141(a), 102(b)(7); Stone v. Ritter, 911 A.2d 362, 370 (Del. 2006); Marchand v. Barnhill, 212 A.3d 805, 820–21 (Del. 2019); In re Massey Energy Co. Derivative & Class Action Litigation, 2011 WL 2176479, at *20–22 (Del. Ch. May 31, 2011); In re McDonald’s Corp. Stockholder Derivative Litigation, 289 A.3d 343, 359–61, 369–70 (Del. Ch. 2023). 

The boardroom point

AI malfeasance claims are no longer confined to the question whether a model copied books, music, code, images, or articles. The Elsevier complaint alleges a top-down AI-training strategy in which Zuckerberg, as founder, chairman, CEO, and controlling shareholder, had ultimate control over Llama development and allegedly authorized, directed, and participated in torrenting pirate collections after employees raised legal and ethical concerns. The SEIU complaint alleges that directors and officers allowed an unlawful AI business strategy, made or approved statements about commercially safe AI, ignored red flags from other AI copyright litigation, and exposed Adobe to litigation costs, stock-price decline, reputational harm, and corporate waste. Together, those pleadings suggest that plaintiffs should treat pirate-data acquisition as both an infringement issue and a governance issue. And both state and federal prosecutors should treat it as a potential RICO issue. 

The practical takeaway is simple: follow the data, but also follow the money. If a defendant merely downloaded pirate data, the case may remain principally a copyright, DMCA, and fiduciary-duty case. If discovery shows that the defendant or its agents paid for bulk pirate access, the evidence may sharpen willfulness, commercial-purpose, damages, knowledge, and governance theories. 

If discovery further shows that the payment was deceptively labeled or concealed through wires, invoices, cryptocurrency, credentials, procurement records, or vendor channels, plaintiffs may have reason to evaluate wire fraud as a potential RICO predicate, subject to all the other statutory requirements.

And even where wire fraud cannot be pleaded, the same payment trail may still be highly relevant to a shareholder derivative theory that corporate fiduciaries knowingly used corporate machinery to build AI products through unlawful data acquisition while representing to investors and customers that those AI products were trained lawfully, safely, and with licensed or rights-cleared data. 

Following the money achieves different ends to different plaintiffs. Copyright plaintiffs should use that trail to prove copying, distribution, willfulness, CMI removal, concealment, damages, and, where the evidence supports deception, continuity, enterprise participation, causation, and injury, potential wire-fraud and RICO theories. Stockholders should use the same trail differently: not to staple a derivative count onto a copyright-owner complaint, but to bring a separate governance action if the facts show that corporate fiduciaries used corporate machinery to acquire unlawful training data, ignored red flags, concealed or mislabeled the payments, or represented to investors and customers that AI products were trained lawfully, safely, and with licensed or rights-cleared data.

The point is not to turn every AI-training case into RICO or every copyright case into a derivative suit. The point is that pirate-data acquisition is no longer just a back-end engineering fact; when the data was bought, disguised, approved, ignored, or monetized at scale, it becomes a roadmap to intent, control, concealment, enterprise conduct, and boardroom accountability.

To paraphrase Deep Throat, forget the myths the media has created about Silicon Valley. The truth is these are not very bright guys and things got out of hand. Follow the money.

Deduplication and Discovery: The Smoking Gun in the Machine

WINSTON

“Wipe up all those little pieces of brains and skull”

From Pulp Fiction, screenplay by Quentin Tarantino and Roger Avary

Deduplication—the process of removing identical or near-identical content from AI training data—is a critical yet often overlooked indicator that AI platforms actively monitor and curate their training sets. This is the kind of process that one would expect given the kind of “scrape, ready, aim” business practices that seems precisely the approach of AI platforms that have ready access to large amounts of fairly high quality data from users of other products placed into commerce by business affiliates or confederates of the AI platforms.

For example, Google Gemini could have access to gmail, YouTube, at least “publicly available” Google Docs, Google Translate, or Google for Education, and then of course one of the great scams of all time, Google Books. Microsoft uses Bing searches, MSN browsing, the consumer Copilot experience, and ad interactions. Amazon uses Alexa prompts, Facebook uses “public” posts and so on.

This kind of hoovering up of indiscriminate amounts of “data” in the form of your baby pictures posted on Facebook and your user generated content on YouTube is bound to produce duplicates. After all, how may users have posted their favorite Billie Eilish or Taylor Swift music video. AI doesn’t need 10000 versions of “Shake it Off” they probably just need the official video. Enter deduplication–which by definition means the platform knows what it has scraped and also knows what it wants to get rid of.

“Get rid of” is a relative concept. In many systems—particularly in storage environments like backup servers or object stores—deduplication means keeping only one physical copy of a file. Any other instances of that data don’t get stored again; instead, they’re represented by pointers to the original copy. This approach, known as inline deduplication, happens in real time and minimizes storage waste without actually deleting anything of functional value. It requires knowing what you have, knowing you have more than one version of the same thing, and being able to tell the system where to look to find the “original” copy without disturbing the process and burning compute inefficiently.

In other cases, such as post-process deduplication, the system stores data initially, then later scans for and eliminates redundancies. Again, the AI platform knows there are two or more versions of the same thing, say the book Being and Nothingness, knows where to find the copies and has been trained to keep only one version. Even here, the duplicates may not be permanently erased—they might be archived, versioned, or logged for auditing, compliance, or reconstruction purposes.

In AI training contexts, deduplication usually means removing redundant examples from the training set to avoid copyright risk. The duplicate content may be discarded from the training pipeline but often isn’t destroyed. Instead, AI companies may retain it in a separate filtered corpus or keep hashed fingerprints to ensure future models don’t retrain on the same material unknowingly.

So they know what they have, and likely know where it came from. They just don’t want to tell any plaintiffs.

Ultimately, deduplication is less about destruction and more about optimization. It’s a way to reduce noise, save resources, and improve performance—while still allowing systems to track, reference, or even rehydrate the original data if needed.

Its existence directly undermines claims that companies are unaware of which copyrighted works were ingested. Indeed, it only makes sense that one of the hidden consequences of the indiscriminate scraping that underpins large-scale AI training is the proliferation of duplicated data. Web crawlers ingest everything they can access—news articles republished across syndicates, forum posts echoed in aggregation sites, Wikipedia mirrors, boilerplate license terms, spammy SEO farms repeating the same language over and over. Without any filtering, this avalanche of redundant content floods the training pipeline.

This is where deduplication becomes not just useful, but essential. It’s the cleanup crew after a massive data land grab. The more messy and indiscriminate the scraping, the more aggressively the model must filter for quality, relevance, and uniqueness to avoid training inefficiencies or—worse—model behaviors that are skewed by repetition. If a model sees the same phrase or opinion thousands of times, it might assume it’s authoritative or universally accepted, even if it’s just a meme bouncing around low-quality content farms.

Deduplication is sort of the Winston Wolf of AI. And if the cleaner shows up, somebody had to order the cleanup. It is a direct response to the excesses of indiscriminate scraping. It’s both a technical fix and a quiet admission that the underlying data collection strategy is, by design, uncontrolled. But while the scraping may be uncontrolled to get copies of as much of your data has they can lay hands on, even by cleverly changing their terms of use boilerplate so they can do all this under the effluvia of legality, they send in the cleaner to take care of the crime scene.

So to summarize: To deduplicate, platforms must identify content-level matches (e.g., multiple copies of Being and Nothingness by Jean-Paul Sartre). This process requires tools that compare, fingerprint, or embed full documents—meaning the content is readable and classifiable–and, oh, yes, discoverable.

Platforms may choose the ‘cleanest’ copy to keep, showing knowledge and active decision-making about which version of a copyrighted work is retained. And–big finish–removing duplicates only makes sense if operators know which datasets they scraped and what those datasets contain.

Drilling down on a platform’s deduplication tools and practices may prove up knowledge and intent to a precise degree—contradicting arguments of plausible deniability in litigation. Johnny ate the cookies isn’t going to fly. There’s a market clearing level of record keeping necessary for deduping to work at all, so it’s likely that there are internal deduplication logs or tooling pipelines that are discoverable.

When AI platforms object to discovery about deduplication, plaintiffs can often overcome those objections by narrowing their focus. Rather than requesting broad details about how a model deduplicates its entire training set, plaintiffs should ask a simple, specific question: Were any of these known works—identified by title or author—deduplicated or excluded from training?

This approach avoids objections about overbreadth or burden. It reframes discovery as a factual inquiry, not a technical deep dive. If the platform claims the data was not retained, plaintiffs can ask for existing artifacts—like hash filters, logs, or manifests—or seek a sworn statement explaining the loss and when it occurred. That, in turn, opens the door to potential spoliation arguments.

If trade secrets are cited, plaintiffs can propose a protective order, limiting access to outside counsel or experts like we’ve done 100,000 times before in other cases. And if the defendant claims “duplicate” is too vague, plaintiffs can define it functionally—as content that’s identical or substantially similar, by hash, tokens, or vectors.

Most importantly, deduplication is relevant. If a platform identified a plaintiff’s work and trained on it anyway, that speaks to volitional use, copying, and lack of care—key issues in copyright and fair use analysis. And if they lied about it, particularly to the court—Helloooooo Harper & Row. Discovery requests that are focused, tailored, and anchored in specific works stand a far better chance of surviving objections and yielding meaningful evidence which hopefully will be useful and lead to other positive results.

The Duty Comes From the Data: Rethinking Platform Liability in the Age of Algorithmic Harm

For too long, dominant tech platforms have hidden behind Section 230 of the Communications Decency Act, claiming immunity for any harm caused by third-party content they host or promote. But as platforms like TikTok, YouTube, and Google have long ago moved beyond passive hosting into highly personalized, behavior-shaping recommendation systems, the legal landscape is shifting in the personal injury context. A new theory of liability is emerging—one grounded not in speech, but in conduct. And it begins with a simple premise: the duty comes from the data.

Surveillance-Based Personalization Creates Foreseeable Risk

Modern platforms know more about their users than most doctors, priests, or therapists. Through relentless behavioral surveillance, they collect real-time information about users’ moods, vulnerabilities, preferences, financial stress, and even mental health crises. This data is not inert or passive. It is used to drive engagement by pushing users toward content that exploits or heightens their current state.

If the user is a minor, a person in distress, or someone financially or emotionally unstable, the risk of harm is not abstract. It is foreseeable. When a platform knowingly recommends payday loan ads to someone drowning in debt, promotes eating disorder content to a teenager, or pushes a dangerous viral “challenge” to a 10-year-old child, it becomes an actor, not a conduit. It enters the “range of apprehension,” to borrow from Judge Cardozo’s reasoning in Palsgraf v. Long Island Railroad (one of my favorite law school cases). In tort law, foreseeability or knowledge creates duty. And here, the knowledge is detailed, intimate, and monetized. In fact it is so detailed we had to coin a new name for it: Surveillance capitalism.

Algorithmic Recommendations as Calls to Action

Defenders of platforms often argue that recommendations are just ranked lists—neutral suggestions, not expressive or actionable speech. But I think in the context of harm accruing to users for whatever reason, speech misses the mark. The speech argument collapses when the recommendation is designed to prompt behavior. Let’s be clear, advertisers don’t come to Google because speech, they come to Google because Google can deliver an audience. As Mr. Wanamaker said, “Half the money I spend on advertising is wasted; the trouble is I don’t know which half.” If he’d had Google, none of his money would have been wasted–that’s why Google is a trillion dollar market cap company.

When TikTok serves the same deadly challenge over and over to a child, or Google delivers a “pharmacy” ad to someone seeking pain relief that turns out to be a fentanyl-laced fake pill, the recommendation becomes a call to action. That transforms the platform’s role from curator to instigator. Arguably, that’s why Google paid a $500,000,000 fine and entered a non prosecution agreement to keep their executives out of jail. Again, nothing to do with speech.

Calls to action have long been treated differently in tort and First Amendment law. Calls to action aren’t passive; they are performative and directive. Especially when based on intimate surveillance data, these prompts and nudges are no longer mere expressions—they are behavioral engineering. When they cause harm, they should be judged accordingly. And to paraphrase the gambling bromide, the get paid their money and they takes their chances.

Eggshell Skull Meets Platform Targeting

In tort law, the eggshell skull rule (Smith v. Leech Brain & Co. Ltd. my second favorite law school tort case) holds that a defendant must take their victim as they find them. If a seemingly small nudge causes outsized harm because the victim is unusually vulnerable, the defendant is still liable. Platforms today know exactly who is vulnerable—because they built the profile. There’s nothing random about it. They can’t claim surprise when their behavioral nudges hit someone harder than expected.

When a child dies from a challenge they were algorithmically fed, or a financially desperate person is drawn into predatory lending through targeted promotion, or a mentally fragile person is pushed toward self-harm content, the platform can’t pretend it’s just a pipeline. It is a participant in the causal chain. And under the eggshell skull doctrine, it owns the consequences.

Beyond 230: Duty, Not Censorship

This theory of liability does not require rewriting Section 230 or reclassifying platforms as publishers although I’m not opposed to that review. It’s a legal construct that may have been relevant in 1996 but is no longer fit for purpose. Duty as data bypasses the speech debate entirely. What it says is simple: once you use personal data to push a behavioral outcome, you have a duty to consider the harm that may result and the law will hold you accountable for your action. That duty flows from knowledge, very precise knowledge that is acquired with great effort and cost for a singular purpose–to get rich. The platform designed the targeting, delivered the prompt, and did so based on a data profile it built and exploited. It has left the realm of neutral hosting and entered the realm of actionable conduct.

Courts are beginning to catch up. The Third Circuit’s 2024 decision in Anderson v. TikTok reversed the district court and refused to grant 230 immunity where the platform’s recommendation engine was seen as its own speech. But I think the tort logic may be even more powerful than a 230 analysis based on speech: where platforms collect and act on intimate user data to influence behavior, they incur a duty of care. And when that duty is breached, they should be held liable.

The duty comes from the data. And in a world where your data is their new oil, that duty is long overdue.

David Sacks Is Learning That the States Still Matter

For a moment, it looked like the tech world’s powerbrokers had pulled it off. Buried deep in a Republican infrastructure and tax package was a sleeper provision — the so-called AI moratorium — that would have blocked states from passing their own AI laws for up to a decade. It was an audacious move: centralize control over one of the most consequential technologies in history, bypass 50 state legislatures, and hand the reins to a small circle of federal agencies and especially to tech industry insiders.

But then it collapsed.

The Senate voted 99–1 to strike the moratorium. Governors rebelled. Attorneys general sounded the alarm. Artists, parents, workers, and privacy advocates from across the political spectrum said “no.” Even hardline conservatives like Ted Cruz eventually reversed course when it came down to the final vote. The message to Big Tech or the famous “Little Tech” was clear: the states still matter — and America’s tech elite ignore that at their peril.  (“Little Tech” is the latest rhetorical deflection promoted by Big Tech aka propaganda.)

The old Google crowd pushed the moratorium–their fingerprints were obvious. Having gotten fabulously rich off of their two favorites: The DMCA farce and the Section 230 shakedown. But there’s increasing speculation that White House AI Czar and Silicon Valley Viceroy David Sacks, PayPal alum and vocal MAGA-world player, was calling the ball. If true, that makes this defeat even more revealing.

Sacks represents something of a new breed of power-hungry tech-right influencer — part of the emerging “Red Tech” movement that claims to reject woke capitalism and coastal elitism but still wants experts to shape national policy from Silicon Valley, a chapter straight out of Philip Dru: Administrator. Sacks is tied to figures like Peter Thiel, Elon Musk, and a growing network of Trump-aligned venture capitalists. But even that alignment couldn’t save the moratorium.

Why? Because the core problem wasn’t left vs. right. It was top vs. bottom.

In 1964, Ronald Reagan’s classic speech called A Time for Choosing warned about “a little intellectual elite in a far-distant capitol” deciding what’s best for everyone else. That warning still rings true — except now the “capitol” might just be a server farm in Menlo Park or a podcast studio in LA.

The AI moratorium was an attempt to govern by preemption and fiat, not by consent. And the backlash wasn’t partisan. It came from red states and blue ones alike — places where elected leaders still think they have the right to protect their citizens from unregulated surveillance, deepfakes, data scraping, and economic disruption.

So yes, the defeat of the moratorium was a blow to Google’s strategy of soft-power dominance. But it was also a shot across the bow for David Sacks and the would-be masters of tech populism. You can’t have populism without the people.

If Sacks and his cohort want to play a long game in AI policy, they’ll have to do more than drop ideas into the policy laundry of think tank white papers and Beltway briefings. They’ll need to win public trust, respect state sovereignty, and remember that governing by sneaky safe harbors is no substitute for legitimacy.  

The moratorium failed because it presumed America could be governed like a tech startup — from the top, at speed, with no dissent. Turns out the country is still under the impression they have something to say about how they are governed, especially by Big Tech.

Jim Hood Was First and He Was Right: Japan Serves Google with Anti-Monopoly C&D

Now with added retroactive acrobatics: @DamianCollins calls on UK Prime Minister to stop Google’s “Text and Data Mining” Circus

Damian Collins (former chair of the UK Parliament’s Digital Culture Media and Sport Select Committee) warns of Google’s latest AI shenanigans in a must-read opinion piece in the Daily Mail that highlights Google’s attempt to lobby its way into what is essentially a retroactive safe harbor to protect Google and its confederates in the AI land grab. While Mr. Collins writes about Google’s efforts to rewrite the laws of the UK to free ride in his home country which is egregious bullying, the episode he documents is instructive for all of us. If Google & Co. will do it to the Mother of Parliaments, it’s only a matter of time until Google & Co. do the same everywhere or know the reason why. Their goal is to hoover up all the world’s culture that the AI platforms have not scraped already and–crucially–to get away with it. And as Guy Forsyth says, “…nothing says freedom like getting away with it.”

The timeline of AI’s appropriation of all the world’s culture is a critical understanding to appreciate just how depraved Big Tech’s unbridled greed really is. The important thing to remember is that AI platforms like Google have been scraping the Internet to train their AI for some time now, possibly many years. This apparently includes social media platforms they control. My theory is that Google Books was an early effort at digitization for large language models to support products like corpus machine translation as a predecessor to Gemini (“your twin”) and other Google AI products. We should ask Ray Kurzweil.

There is starting to be increasing evidence that this is exactly what these people are up to.

The New York Times Uncovers the Crimes

According to an extensive long-form report in the New York Times by a team of very highly respected journalists, it turns out that Google has been planning this “Text and Data Mining” land grab for some time. At the very moment YouTube was issuing press releases about their Music AI Incubator and their “partners”–Google was stealing anything that was not nailed down that anyone had hosted on their massive platforms, including Google Docs, Google Maps, and…YouTube. The Times tells us:

Google transcribed YouTube videos to harvest text for its A.I. models, five people with knowledge of the company’s practices said. That potentially violated the copyrights to the videos, which belong to their creators….Google said that its A.I. models “are trained on some YouTube content,” which was allowed under agreements with YouTube creators, and that the company did not use data from office apps outside of an experimental program. 

I find it hard to believe that YouTube was both allowed to transcribe and scrape under all its content deals, or that they parsed through all videos to find the unprotected ones subject to their interpretation of the YouTube terms of use. So as we say in Texas, that sounds like bullshit for starters.

How does this relate to the Text and Data Mining exception that Mr. Collins warns of? Note that the NYT tells us “Google transcribed YouTube videos to harvest text.” That’s a clue.

As Mr. Collins tells us:

Google [recently] published a policy paper entitled: Unlocking The UK’s AI Potential.

What’s not to like?, you might ask. Artificial intelligence has the potential to revolutionise our economy and we don’t want to be left behind as the rest of the world embraces its benefits.

But buried in Google’s report is a call for a ‘text and data mining’ (TDM) exception to copyright.

This TDM exception would allow Google to scrape the entire history of human creativity from the internet without permission and without payment.

And, of course, Mr. Collins is exactly correct, that’s exactly what Google have in mind.

The Conspiracy of Dunces and the YouTube Fraud

In fairness, it wasn’t just Google ripping us off, but Google didn’t do anything to stop it as far as I can tell. One thing to remember is that YouTube was, and I think still is, not very crawlable by outsiders. It is almost certainly the case that Google would know who was crawling youtube.com, such as Bingbot, DuckDuckBot, Yandex Bot, or Yahoo Slurp if for no other reason that those spiders were not googlebot. With that understanding, the Times also tells us:

OpenAI researchers created a speech recognition tool called Whisper. It could transcribe the audio from YouTube videos, yielding new conversational text that would make an A.I. system smarter.

Some OpenAI employees discussed how such a move might go against YouTube’s rules, three people with knowledge of the conversations said. YouTube, which is owned by Google, prohibits use of its videos for applications that are “independent” of the video platform. [Whatever “independent” means.]

Ultimately, an OpenAI team transcribed more than one million hours of YouTube videos, the people said. The team included Greg Brockman, OpenAI’s president, who personally helped collect the videos, two of the people said. The texts were then fed into a system called GPT-4, which was widely considered one of the world’s most powerful A.I. models and was the basis of the latest version of the ChatGPT chatbot….

OpenAI eventually made Whisper, the speech recognition tool, to transcribe YouTube videos and podcasts, six people said. But YouTube prohibits people from not only using its videos for “independent” applications, but also accessing its videos by “any automated means (such as robots, botnets or scrapers).”

OpenAI employees knew they were wading into a legal gray area, the people said, but believed that training A.I. with the videos was fair use. 

And strangely enough, many of the AI platforms sued by creators raise “fair use” as a defense (if not all of the cases) which is strangely reminiscent of the kind of crap we have been hearing from these people since 1999.

Now why might Google have permitted OpenAI to crawl YouTube and transcribe videos (and who knows what else)? Probably because Google was doing the same thing. In fact, the Times tells us:

Some Google employees were aware that OpenAI had harvested YouTube videos for data, two people with knowledge of the companies said. But they didn’t stop OpenAI because Google had also used transcripts of YouTube videos to train its A.I. models, the people said. That practice may have violated the copyrights of YouTube creators. So if Google made a fuss about OpenAI, there might be a public outcry against its own methods, the people said.

So Google and its confederate OpenAI may well have conspired to commit massive copyright infringement against the owner of a valid copyright, did so willingly, and for purposes of commercial advantage and private financial gain. (Attempts to infringe are prohibited to the same extent as the completed act). The acts of these confederates vastly exceed the limits for criminal prosecution for both infringement and conspiracy.

But to Mr. Collins’ concern, the big AI platforms transcribed likely billions of hours of YouTube videos to manipulate text and data–you know, TDM.

The New Retroactive Safe Harbor: The Flying Googles Bring their TDM Circus Act to the Big Tent With Retroactive Acrobatics

But also realize the effect of the new TDM exception that Google and their Big Tech confederates are trying to slip past the UK government (and our own for that matter). A lot of the discussion about AI rulemaking acts as if new rules would be for future AI data scraping. Au contraire mes amis–on the contrary, the bad acts have already happened and they happened on an unimaginable scale.

So what Google is actually trying to do is get the UK to pass a retroactive safe harbor that would deprive citizens of valuable property rights–and also pass a prospective safe harbor so they can keep doing the bad acts with impunity.

Fortunately for UK citizens, the UK Parliament has not passed idiotic retroactive safe harbor legislation like the U.S. Congress has. I am, of course, thinking of the vaunted Music Modernization Act (MMA) that drooled its way to a retroactive safe harbor for copyright infringement, a shining example of the triumph of corruption that has yet to be properly challenged in the US on Constitutional grounds.

There’s nothing like the MMA absurdity in the UK, at least not yet. However, that retroactive safe harbor was not lost on Google, who benefited directly from it. They loved it. They hung it over the mantle next to their other Big Game trophy, the DMCA. And now they’d like to do it again for the triptych of legislative taxidermy.

Because make no mistake–a retroactive safe harbor would be exactly the effect of Google’s TDM exception. Not to mention it would also be a form of retroactive eminent domain, or what the UK analogously might call the compulsory purchase of property under the Compulsory Purchase of Property Act. Well…”purchase” might be too strong a word, more like “transfer” because these people don’t intend to pay for a thing.

The effect of passing Google’s TDM exception would be to take property rights and other personal rights from UK citizens without anything like the level of process or compensation required under the Compulsory Purchase of Property–even when the government requires the sale of private property to another private entity (such as a railroad right of way or a utility easement).

The government is on very shaky ground with a TDM exception imposed by the government for the benefit of a private company, indeed foreign private companies who can well afford to pay for it. It would be missing government oversight on a case-by-base basis, no proper valuation, and for entirely commercial purposes with no public benefit. In the US, it would likely violate the Takings Clause of our Constitution, among other things.

It’s Not Just the Artists

Mr. Collins also makes a very important point that might get lost among the stars–it’s not just the stars that AI is ripping off–it is everyone. As the New York Times story points out (and it seems that there’s more whistleblowers on this point every day), the AI platforms are hoovering up EVERYTHING that is on the Internet, especially on their affiliated platforms. That includes baby videos, influencers, everything.

This is why it is cultural appropriation on a grand scale, indeed a scale of depravity that we haven’t seen since the Nurenberg Trials. A TDM exception would harm all Britons in one massive offshoring of British culture.

AI is Using Forks and Knives to Eat Your Bacon…and then there’s the political clout

The Internet is an extraordinary electricity hog. You know this intuitively even if you’ve never studied the question of just how big a hog it really is. AI has already taken that electricity use to exponentially extraordinary new levels. These hogs will ultimately consume the farm if that herd is not thinned out.

This is nothing new. Consider YouTube. First of all, YouTube has long been the second largest search engine in the world. So there’s that. Reportedly, YouTube’s aggregate audience watches over 1 billion viewing hours per day.

To put that in context, the electricity burned by YouTubers works out to approximately 600 terawatt-hours (TWh) per year. (A terawatt hour (TWh) is a unit of energy that represents the amount of work done by one terawatt of power in one hour. The prefix ‘tera’ signifies 10^12. Therefore, one terawatt equals one trillion (1,000,000,000,000 or 10^12) watts.)

600 TWh is roughly 2.5% of global electricity use, exceeds the combined consumption of all data centers and data transmission networks worldwide, and “could power an American household for about 2 billion years. Or all 127 million U.S. households for about 8 years.”

And that’s just YouTube.

AI’s electricity consumption is a growing concern. Data centers worldwide, which store the information required for online interactions, account for about 1 to 1.5 percent of global electricity use. The AI boom could potentially increase this figure significantly.

A study by Alex de Vries, a data scientist at the central bank of the Netherlands, estimates that by 2027, NVIDIA could be shipping 1.5 million AI server units per year. These servers, running at full capacity, would consume at least 85.4 terawatt-hours (TWh) of electricity annually. This consumption is more than what many countries use in a year.

AI’s energy footprint does not end with training–after training comes users or the “inference” phase. When an AI is put to work generating data based on prompts in the inference phase, it also uses a significant amount of computing power and thus electricity. For example, generative AI like ChatGPT could cost 564 megawatt-hours (MWh) of electricity a day to run.

According to de Vries, “In 2021, Google’s total electricity consumption was 18.3 TWh, with AI accounting for 10%–15% of this total. The worst-case scenario suggests Google’s AI alone could consume as much electricity as a country such as Ireland (29.3 TWh per year), which is a significant increase compared to its historical AI-related energy consumption.” Remember, that’s just Google.

And then there’s bitcoin mining. Bitcoin mining involves computers across the globe racing to complete a computation that creates a 64-digit hexadecimal number, or hash, for each Bitcoin transaction. This hash goes into a public ledger so anyone can confirm that the transaction for that particular bitcoin happened. The computer that solves the computation first gets a reward of 6.2 bitcoins.

Bitcoin mining consumes a significant amount of electricity. In May 2023, Bitcoin mining was estimated to consume around 95.58 terawatt-hours of electricity. It reached its highest annual electricity consumption in 2022, peaking at 204.5 terawatt-hours, surpassing the power consumption of Finland.

By some estimates, a single Bitcoin transaction can spend up to 1,200 kWh of energy, which is equivalent to almost 100,000 VISA transactions. Another estimate is that Bitcoin uses more than 7 times as much electricity as all of Google’s global operations.

Back to AI, remember that Big Tech requires big data centers. According to Bloomberg, AI is currently–no pun intended–currently using so much electrical power that coal plants that utilities planned to shut down for climate sustainability are either staying online or being brought back online if they had been shut down. For example, Virginia has been suffering from this surge in usage:

In a 30-square-mile patch of northern Virginia that’s been dubbed “data center alley,” the boom in artificial intelligence is turbocharging electricity use. Struggling to keep up, the power company that serves the area temporarily paused new data center connections at one point in 2022. Virginia’s environmental regulators considered a plan to allow data centers to run diesel generators during power shortages, but backed off after it drew strong community opposition.

It’s also important to realize that building data centers in states that are far flung from the Silicon Valley heartland also increases Big Tech’s political clout. This explains why Oregon Senator Ron Wyden is the confederate of the worst of Big Tech’s excesses like child exploitation and of course, copyright. Copyright never had a worse enemy, all because of the huge presence in Oregon of Big Tech’s data centers–not their headquarters or anything obvious.

Wyden with his hand in his own pocket.

A terawatt here and a terawatt there and pretty soon you’re talking about a lot of electricity. So if you’re interested in climate change, there’s a lot of material here to work with. Maybe we do this before we slaughter the cattle, just sayin’.

Thinking Outside the Pie: @legrandnetwork Study for GESAC Highlights Streaming Impact on Choking Diversity and Songwriter Royalties

Emmanuel Legrand prepared an excellent and important study for the European Grouping of Societies of Authors and Composers (GESAC) that identifies crucial effects of streaming on culture, creatives and especially songwriters. The study highlights the cultural effects of streaming on the European markets, but it would be easy to extend these harms globally as Emmanuel observes.

For example, consider the core pitch of streaming services that started long ago with the commercial Napster 2.0 pitch of “Own Nothing, Have Everything”. This call-to-serfdom slogan may sound good but having infinite shelf space with no cutouts or localized offering creates its own cultural imperative. And that’s even if you accept the premise the algorithmically programed enterprise playlists on streaming services should not be subject to the same cultural protections for performers and songwriters as broadcast radio–its main competitor.

[This] massive availability of content on [streaming] platforms is overshadowed by the fact that these services are under no positive obligations to ensure visibility and discoverability of more diverse repertoires, particularly European works….[plus]  the initial individual subscription fee of 9.99 (in Euros, US dollars, or British pound) set in 2006, has never increased, despite the exponential growth in the quality, amount of songs, and user-friendliness of music streaming services.

Artists working new recordings, especially in a language other than English, are forced to fight for “shelf space” and “mindshare”–that is, recognition–against every recording ever released. While this was always true theoretically; you never had that same fight the same way at Tower Records.

This is not theoretically true on streaming platforms–it is actually true because these tens of millions of historical recordings are the competition on streaming services. When you look at the global 100 charts for streaming services, almost all of the titles are in English and are largely Anglo-American releases. Yes, we know–Bad Bunny. But this year’s exception proves the rule.

And then Emmanuel notes that it is the back room algorithms–the terribly modern version of the $50 handshake–that support various payola schemes:

The use of algorithms, as well as bottleneck represented by the most popular playlists, exacerbates this. Furthermore, long-standing flaws in the operations of music streaming platforms, such as “streaming fraud”, “ghost/fake artists”, “payola schemes”, “royalty free content” and other coercive practices [not to mention YouTube withholding access to Content ID] worsen the impact on many professional creators….

This report suggests solutions to bring greater transparency in the use of algorithms and invites stakeholders to undertake a review of the economic models of streaming services and evaluate how they currently affect cultural diversity which should be promoted in its various forms — music genres, languages, origin of performers and songwriters, in particular through policy actions.

MTS readers will recall my extensive dives into the hyperefficient market share distribution of streaming royalties known as the “big pool” compared to my “ethical pool” proposal and the “user centric” alternative. As Emmanuel points out, the big pool royalty model belies a cultural imperative–if you are counting streams on a market share basis that results in the rich getting richer based on “stream share” that same stream share almost guarantees that Anglo American repertoire will dominate in every market the big streamers operate.

Emmanuel uses French-Canadian repertoire as an example (a subject I know a fair amount about since I performed and recorded with many vedettes before Quebecoise was cool).

A lot of research has been made in Canada with regards to discoverability, in particular in the context of French-Canadian music, which is subject to quotas for over the air broadcasters which however do not apply to music streaming services. The research shows that while the lists of new releases from Québec studied are present in a large proportion on streaming platforms, they are “not very visible and very little recommended.” 

It further shows that the situation is even worse when it is not about new releases, including hit music, when the presence of titles “drops radically.” It is not very difficult to imagine that if we were to swap Québec in the above sentence with the name of any country from the European Union [or any non-Anglo American country], and even with music from the European Union as a whole, we could find similar results.

In other words, there may be aggregators with repertoire in languages other than English that deliver tracks to streamers in their countries, but–absent localized airplay rules–a Spotify user might never know the tracks were there unless the user already knew about the recording, artist or songwriter. (Speaking of Canada, check the MAPL system.)

This is a prime example of why Professor Feijoo and I proposed streaming remuneration in our WIPO study to allow performers to capture the uncompensated capital markets value to the enterprise driven by these performers. Because of the market share royalty system, revenues and royalties do not compensate all performers, particularly regional or non-featured performers (i.e., session players and singers) who essentially get zero compensation for streaming.

Emmanuel also comments on the imbalance in song royalty payments and invites a re-look at how the streaming system biases against songwriters. I would encourage everyone to stop thinking of a pie to be shared or that Johnny has more apples–when the services refuse to raise prices in order to tell a growth story to Wall Street or The City, measuring royalties by a share of some mythical royalty pie is not ever going to get it done. It will just perpetuate a discriminatory system that fails to value the very people on whose backs it was built be they songwriters or session players.

We must think outside the pie.