Eric Schmidt Said Startups Could ‘Steal’ Content—The Law Offers No Blank Check

Eric Schmidt’s 2024 remarks about copying content and hiring lawyers remain a historical controversy, not a newly announced policy or a reliable description of copyright law. As of August 13, 2026, there is still no blanket rule allowing an AI startup to take music, writing, images or user data merely because its product is experimental or unsuccessful.

What has changed is the legal record surrounding that argument. The Copyright Office’s current AI initiative page still lists its May 2025 analysis of generative-AI training as a pre-publication report, with a final version expected without substantive changes to its conclusions. Those conclusions describe fair use as a fact-specific inquiry—not permission to copy first and treat liability as a problem for later.

What Schmidt actually told the Stanford audience

The disputed passage came from a Stanford University talk that became public in August 2024. Schmidt, Google’s CEO from 2001 to 2011, imagined students asking a large language model to produce a TikTok competitor quickly if the service were banned in the United States.

In that hypothetical command, he told the model to copy TikTok, take its users and music, launch immediately and abandon the attempt if it failed to spread. He then suggested that a successful Silicon Valley entrepreneur could hire lawyers to “clean the mess up.” Fortune’s contemporaneous account also records his qualification that he was not advocating the illegal theft of everyone’s music and reports that the Stanford video was removed at his request.

That qualification matters, but it does not erase the premise of the example: legal exposure was treated as something a growing company might address after proving demand. The remarks therefore support a narrower conclusion than the most sensational version of the story. Schmidt did not establish that AI companies are legally free to absorb copyrighted material; he described a provocative startup tactic and immediately acknowledged that it could create a legal mess.

Why failure does not make copying irrelevant

A product’s popularity is not the switch that determines whether copying is lawful. Under US copyright law, questions can arise when protected expression is reproduced, distributed, displayed, adapted or incorporated into a competing product. Whether a particular use is excused as fair use depends on its purpose, the kind of work copied, the amount used and its effect on the market.

The commercial result can still affect practical decisions. A rights holder may consider the reach of a product, the evidence available and the cost of enforcement before bringing a claim. A startup may also attract more scrutiny once it gains users or funding. Those realities explain the business logic behind Schmidt’s remark, but they do not turn an unsuccessful release into automatically lawful conduct.

His example also bundled together legally distinct actions. Reproducing music implicates copyright; misleading users or importing account information can raise contractual, privacy, computer-access or consumer-protection questions; imitating an interface may involve a different mixture of protectable expression and unprotected functionality. Calling every element “content” conceals those differences.

Later rulings made the picture more specific, not simpler

The strongest update since the remarks is that courts and federal analysts have begun distinguishing model training from the acquisition and storage of training material. The Congressional Research Service’s July 2025 legal analysis summarizes two federal decisions that found particular uses of books to train generative models fair. Yet one of those decisions separately rejected maintaining a central library of books obtained from pirate sites, while the two judges took different approaches to unauthorized downloads.

Those decisions were not universal licenses for AI training. They turned on the records developed in specific cases, including the purpose of the copying, how the books were obtained and the evidence of market harm. Other plaintiffs, datasets, outputs and business models can produce different results.

The Copyright Office’s broader assessment points in the same direction. It expects some training uses to qualify as fair use and others not to qualify. Noncommercial research that does not reproduce protected passages in outputs sits toward the more defensible end of the spectrum; commercially copying expressive works from illicit sources to generate competing material sits toward the other end. Many real products fall between those poles.

What the episode means for creators and AI founders

For creators, the key issue is not whether an executive used the word “steal.” It is whether a company copied a protected work, had authorization or another lawful basis, retained the material, generated substantially similar output, and harmed an existing or reasonably developing market. Evidence about provenance, licensing terms and outputs is therefore more useful than a general accusation that a model learned from the internet.

For founders, rapid prototyping does not require treating provenance as an afterthought. A team can identify which assets come from users, commercial licensors, public-domain collections, permissive licenses or material generated inside the company. It can also record the permissions attached to each dataset and separate temporary evaluation data from material retained for training or product delivery.

That work is not only defensive compliance. It tells a company which features it can continue operating if a supplier changes terms, a creator withdraws permission or a court narrows an expected fair-use argument. A product whose core output depends on unverifiable copies carries a different operational risk from one built on documented licenses and controlled inputs.

Schmidt’s remarks endure because they captured a genuine power imbalance: a successful company may be better able to fund litigation than the people whose work it used. The subsequent legal developments have not validated that strategy. They have instead made source, purpose, market substitution and licensing more consequential—precisely the details that “launch now, clean up later” leaves unresolved.

Also read:

• Uber CEO Dara Khosrowshahi Just Said What Every Other Executive Is Too Scared to Admit: AI Will Replace 70-80% of Human Work
• Tech Companies Furious at New Law That Would Hold Them Accountable When Their AI Does Bad Stuff
• Google’s Wake-Up Call: Quantum Computers Could Crack Bitcoin With 20x Fewer Resources Than Anyone Thought