Opinion: When AI can find you, ‘legal’ isn't enough

Opinion

Imagine that an employee anonymously records a confidential company presentation on her laptop. She keeps her name and face out of the video. A plane passes overhead, too faint to demand her attention.

In our companion story, an investigator uses that sound, flight records and a list of employees to identify her. The scene is fictional: no one has shown that this complete investigation happened. But it shows a question that applies to every AI system, not only this one: is it enough for an AI model to stay within the law?

Our answer is no. Legality is a necessary baseline, but it is too thin a standard for systems that can combine ordinary details into deeply personal knowledge, and for systems that are changing faster than the rules that govern them.

The case that law is enough

The argument for relying on existing law has influential supporters, and it deserves a fair statement.

Venture capitalist Marc Andreessen made it in a 2023 essay, “Why AI Will Save the World.” Fears of AI misuse, he wrote, can be met with laws already on the books: “We can simply focus on preventing those crimes when we can, and prosecuting them when we cannot.” He added: “We don’t even need new laws.” If a new harmful use appears, he wrote, “we ban that use.” He was also dismissive of the people raising the alarm. Of some AI-risk advocates, he wrote: “They are paid to be doomers, and their statements should be processed appropriately.” 1

His manifesto from that October went further, listing “trust and safety,” “tech ethics,” “social responsibility” and “risk management” among the bad ideas of a six-decade “mass demoralization campaign.” It said that “any deceleration of AI will cost lives.” 2

Parts of this argument are reasonable. Elected governments, not companies, should make the final rules. Safety demands can be used by incumbents to shut out competitors, a danger Andreessen and others rightly flag. And existing law does cover a great deal.

The same view has shaped US policy. In January 2025 President Trump revoked President Biden’s 2023 executive order on AI, which had required developers of the most capable models to report their safety-testing results to the federal government 3, and set a goal of “sustain and enhance America’s global AI dominance.” 4 A month later Vice President JD Vance told the Paris AI summit: “The AI future is not going to be won by hand-wringing about safety.” He did not say safety no longer mattered, but he said he had come to talk about opportunity, and warned that excessive regulation could kill a transformative industry. 5 In December 2025 another order, which describes the earlier revocation as ending a predecessor’s “attempt to paralyze this industry,” created a Justice Department task force to challenge state AI laws and told agencies to consider conditioning grants on states not enacting laws that conflict with a “minimally burdensome” national policy. The same order says the national framework should protect children and safeguard communities, and it exempts child-safety laws. 6

The Pentagon applies the same logic to its own contracts. On January 9, 2026, Secretary of War Pete Hegseth issued a memo instructing that an “any lawful use” provision be written into any department contract through which AI services are procured. 7 The memo also told the department to use models “free from usage policy constraints that may limit lawful military applications.” 8 That is a position about government procurement, not a claim that every use of AI by a private person should be allowed. But it rests on the same premise: if it is lawful, the developer should not stand in the way. When Anthropic refused to drop two restrictions, on lethal autonomous warfare and mass surveillance of Americans, the government designated it a supply chain risk. A federal judge later found that the record showed those actions were unlawful retaliation. 7

Where “lawful” runs out

Anthropic chief executive Dario Amodei refused to remove his company’s limits on mass domestic surveillance and fully autonomous weapons. His reason is the weakness at the center of the law-is-enough view. Amodei argued that mass surveillance is legal today only because the law “has not yet caught up with the rapidly growing capabilities of AI.” The government can buy detailed records about Americans without a warrant, he said, and AI can assemble them into comprehensive profiles. 9

He also accepted the Pentagon’s point that the military, not a private company, makes military decisions. The disagreement was narrower and more interesting: what should a developer refuse to do even when no law forbids it?

OpenAI’s answer was in between. Its Pentagon contract says the department “may use the AI System for all lawful purposes,” but OpenAI also describes red lines against mass domestic surveillance, directing autonomous weapons and high-stakes automated decisions, and says it keeps control of its own safety stack rather than relying on usage policies alone. Asked what happens if the government changes the law, it said its contract refers to the relevant laws and policies as they exist today. 10 That is a company building a boundary that does not move with the law.

Lawfare’s Jessica Tillipman raised a related problem. Under an “any lawful use” rule, the government is often the first and only interpreter of what is lawful, and a contractor’s remedies arrive after the fact. She concluded that these questions deserve answers from Congress and the courts, not from procurement negotiations. 8

The same gap appears when ordinary people use AI. Researchers at ETH Zurich showed that language models could infer personal details from Reddit posts, reaching up to 85 percent top-1 accuracy at about one-hundredth the cost and one-two-hundred-fortieth the time of human investigators. 11 Researchers at Privacy International and several universities found that vision-language models can work out where a photograph was taken even when its GPS data has been removed. 12 None of this requires secret data. What changes is the effort needed to assemble public clues into a private conclusion.

Philosopher Helen Nissenbaum’s idea of contextual integrity helps explain the problem. Information flows within settings, with expectations about who receives it and why. Hearing a plane in the background of a video is one thing. Combining that sound with flight records and a list of home addresses to identify the recorder is another. 13

What a better standard looks like

A better standard does not mean banning everything. Investigative journalists, researchers and people looking for a missing relative all have legitimate reasons to examine public evidence, and a blanket ban would sacrifice valuable work. It means asking more than “is this legal?”

AI developers should consider what someone is trying to accomplish and whose privacy is at stake, not merely whether each individual data source is public. An assistant asked to find the home of a private person who has not consented should not treat that request like trivia. Providers can limit help in identifying or precisely locating private people from incidental clues, test whether their systems can combine sources to get around those limits, and be cautious about presenting an inferred identity as fact. Organizations that act on an AI’s answer should require human review and a way to challenge a mistaken identification. A confident answer can accuse the wrong person.

Andreessen is right that rules should be accountable and should not simply protect incumbents. That is a reason to make safety standards public, tested and open to challenge, not a reason to leave the question to whatever the law happened to say before the technology arrived.

We do not have to wait for an AI that can find everyone, everywhere. The time to decide what an assistant should help reveal is while its capabilities are still changing, not after the expectation of privacy has quietly become a relic.

References

References

1. Why AI Will Save the World. (n.d.). https://a16z.com/ai-will-save-the-world/
2. The Techno-Optimist Manifesto. (n.d.). https://a16z.com/the-techno-optimist-manifesto/
3. Safe, Secure, and Trustworthy Development and Use of Artificial Intelligence (Executive Order 14110). (n.d.). https://www.federalregister.gov/documents/2023/11/01/2023-24283/safe-secure-and-trustworthy-development-and-use-of-artificial-intelligence
4. Removing Barriers to American Leadership in Artificial Intelligence (Executive Order 14179). (n.d.). https://www.whitehouse.gov/presidential-actions/2025/01/removing-barriers-to-american-leadership-in-artificial-intelligence/
5. Read JD Vance’s full speech on AI and the EU. (n.d.). https://spectator.com/article/read-jd-vances-full-speech-on-ai-and-the-eu
6. Ensuring a National Policy Framework for Artificial Intelligence (Executive Order 14365). (n.d.). https://www.federalregister.gov/documents/2025/12/16/2025-23092/ensuring-a-national-policy-framework-for-artificial-intelligence
7. Anthropic PBC v. U.S. Department of War: Order on cross motions for summary judgment. (n.d.). https://storage.courtlistener.com/recap/gov.uscourts.cand.465515/gov.uscourts.cand.465515.250.0.pdf
8. Military AI Policy by Contract: The Limits of Procurement as Governance. (n.d.). https://www.lawfaremedia.org/article/military-ai-policy-by-contract--the-limits-of-procurement-as-governance
9. Statement from Dario Amodei on our discussions with the Department of War. (n.d.). https://www.anthropic.com/news/statement-department-of-war
10. Our agreement with the Department of War. (n.d.). https://openai.com/index/our-agreement-with-the-department-of-war/
11. Beyond Memorization: Violating Privacy Via Inference with Large Language Models. (n.d.).
12. Nowhere to Hide? Privacy Risks and Policy Implications of AI Geolocation. (n.d.).
13. Privacy as Contextual Integrity. (n.d.). https://digitalcommons.law.uw.edu/wlr/vol79/iss1/10/