Adam Schiff on AI Regulation: Why Voluntary Pledges Cannot Solve Power, Liability, Privacy, and Cost

1. Guest Background

The guest in this episode of Decoder with Nilay Patel is Adam Schiff. The episode evidence identifies him as a Democratic United States Senator from California, serving in the United States Senate. Patel also introduces him as someone sitting on committees and subcommittees with oversight into technology, AI, intellectual property, antitrust, privacy, and technology law. That matters because Schiff is not speaking as a founder, investor, or lab researcher. He is speaking as a legislator whose job is to think about oversight, statutory authority, litigation risk, agency capacity, industry lobbying, and how technological costs land on workers and consumers.

That institutional role shapes the interview. Schiff does not reduce AI policy to whether a model refuses a bad prompt or produces a safe answer. He moves across the whole governance chain: what data was used to train models, what happens when agents cause harm, how model companies and users should divide liability, why social-media regulation failed, whether safety coordination needs any antitrust accommodation, how public-space surveillance changes privacy, who pays for data centers, and how AI pressures film and television labor. The episode is therefore an analysis of AI as a political and economic system rather than a narrow discussion of model behavior.

The title’s claim that voluntary AI rules are meaningless comes from a specific episode frame. Patel describes President Trump gathering AI CEOs at the White House to sign a non-binding AI safety pact. Schiff’s objection is not that corporate safety work is inherently useless. It is that voluntary commitments without enforceable duties, external accountability, defined liability, and legal consequences cannot substitute for public regulation. His relevant experience in the episode is precisely that he can connect AI companies, presidential power, congressional oversight, courts, agencies, copyright holders, media firms, and workers in one policy map.

2. What the Episode Covers

The interview begins with a concrete political event: President Trump bringing AI CEOs to the White House to sign a non-binding safety promise. Patel asks whether that is enough and what a formal framework would look like if it is not. Schiff says the scene is hard to square because the same executives who sign voluntary pledges near the president also say, outside the White House, that AI cannot be governed by “trust us” self-policing and needs regulation.

From there, the conversation moves into national security. Schiff argues that as models become more advanced, especially through recursive AI development, humans have less understanding of what the systems are doing, how they are deciding, and what they are trying to access lawfully or unlawfully. He applies that concern to military and government settings, including AI at the Pentagon, target selection, domestic mass surveillance, and fully autonomous weapons. The episode does not independently prove every specific example he raises; those examples function as Schiff’s evidence for why the public sector cannot treat advanced AI as an ordinary software product.

The second major subject is intellectual property and training data. Schiff says AI companies behaved as if they were in a race: vacuum up everything first, even if it likely violated IP law, and worry about legality later. His first-step proposal with Senator Curtis is not to settle every fair-use question immediately. It is to require companies to know and retain what copyrighted or protected materials were used to train models, so that the fair-use, licensing, compensation, or litigation debate has a factual basis. Schiff also acknowledges beneficial uses, such as AI helping the Patent Office process application backlogs more quickly, which keeps his argument from becoming a blanket rejection of AI.

The middle of the episode compares AI regulation with social-media regulation. Patel argues that social-media rules often crashed into First Amendment questions around disinformation, bullying, and pro-anorexia content. Schiff replies that the First Amendment itself was not the core obstacle; companies used First Amendment framing to block regulation of many issues that were not necessarily speech issues. He criticizes Section 230 as overly broad immunity that allowed platforms to avoid consequences while training addictive algorithms and permitting harms. That leads into the AI liability question: when a model generates content, a platform amplifies content, an agent runs an account, or a tool is used in a cyberattack, responsibility no longer fits neatly into the old user-versus-platform distinction.

The final stretch widens the lens. Schiff supports a regulator with real power, expertise, discretion, and enforcement capacity, perhaps an FDA for artificial intelligence. Patel pushes on Loper Bright, explaining that after Chevron deference was overturned, vague agency authority is easier for corporate interests to litigate. Schiff accepts that constraint and says Congress must be as specific as possible when empowering an AI agency. The interview then moves through public-space privacy, Flock cameras, foreign actors and model distillation, narrow antitrust exemptions for security coordination, data-center costs, congressional oversight if Democrats win control, media mergers, film and television labor, Medicare expansion, and a progressive tax code. The result is a broad episode about how AI stress-tests institutions, not just products.

3. Core Views: Reasoning, Examples, and Limits

Schiff’s clearest view is that voluntary AI pledges cannot do the work of law. His point is not that internal safety teams are worthless. It is that a non-binding statement saying companies will empower internal teams to ensure controls, monitoring, and detection work as intended does not answer the public questions that matter: who audits the controls, who can compel disclosure, who bears liability after failure, what victims can do, and what penalties attach to misconduct. Patel calls that language literally meaningless; Schiff agrees. The reasoning is straightforward: companies say they want legal clarity, but a pledge without enforcement gives them political cover rather than a legal framework.

The force of that argument comes from the number of layers Schiff thinks AI regulation must cover. On the input side, there is copyrighted training material. If companies ingest material first and litigate later, creators and entertainment firms lose leverage before anyone reaches a fair-use conclusion. On the output side, models and agents can create safety problems that companies themselves may not fully control. On the platform side, AI generation and amplification blur whether something is user speech, platform speech, or model output. Patel’s examples involving Meta-generated content, automated accounts, and paid amplification show how old platform-liability categories become unstable when the platform is not merely hosting speech but shaping or producing it.

The social-media comparison is Schiff’s warning label. He believes Congress did a poor job regulating social-media harms and cannot take years to repeat that process with AI. His diagnosis differs from Patel’s First Amendment framing. Patel says social-media rules often became speech regulation. Schiff says companies strategically turned many product-design and accountability questions into First Amendment fights. Section 230, in his telling, gave platforms broad immunity while they failed to behave as good corporate citizens, trained algorithms for addiction, and allowed harm to young people. The implication for AI is that broad immunity granted too early can become a liability vacuum that later proves politically and legally hard to narrow.

At the same time, Schiff’s view is not simply “abolish immunity and ban coordination.” He says Section 230 could be narrowed rather than eliminated, with safe harbors for companies that meet best practices and standards. On antitrust, he supports a very narrow exemption for security coordination, such as sharing information about foreign distillation attempts, malware, or attacks and notifying the government. But he draws a hard boundary against using security as a pretext for price increases or an anti-competitive cabal. This matters because it shows the policy shape he favors: not maximum discretion for companies, and not total prohibition of cooperation, but narrow permissions tied to specific public-interest functions.

The liability discussion is one of the episode’s strongest practical sections. Schiff uses a gun-sale analogy to argue that AI responsibility should be layered. A tool can be used legally or illegally; liability differs when someone sells an ordinary weapon, sells a defective weapon, sells to an ineligible buyer, or fails to do due diligence. Applied to AI, that means responsibility may differ for a frontier model developer, a wrapper tool built for hacking, a user directing an agent, or a company that failed to monitor obvious ransomware use. Patel complicates the analogy with cyber defense: he says OpenAI’s Mythos can be used for offense and defense, while cybersecurity firms argue that humans may need to be removed from the defensive loop to respond fast enough to AI-powered attacks. The limitation is real: the same capability can be socially useful or dangerous depending on context, so any liability rule will be contested at the margins.

Schiff’s institutional answer is an expert regulator with real power and teeth, perhaps an FDA for artificial intelligence. The reason is speed. He contrasts fast technological change with a slow legislative calendar and argues that Congress cannot update detailed rules quickly enough on its own. But this is not presented as an easy fix. Schiff notes that government may struggle to afford AI talent because private-sector salaries are much higher. Patel adds the more structural problem: after Loper Bright, courts no longer have to defer to agency experts when statutes are vague. Schiff accepts that AI legislation must empower agencies as specifically as possible. The limitation is that an AI agency without precise statutory authority may be less a solution than a litigation target.

Privacy broadens the argument beyond corporate compliance. Schiff calls privacy in the era of powerful AI a cardinal concern because models could be trained to obtain health records, infer viewing habits, or activate microphones on consumer devices. Even if American law prohibits such conduct, foreign actors can develop models, distill American models, or buy capabilities from private sellers. That is why Schiff says policy must not only prohibit conduct but also strengthen privacy technologically. The Flock camera example gives the issue a concrete public-space form: license-plate cameras can help catch serious criminals, but Schiff says they have also been used to track a critic of police-camera deployment and a woman seeking reproductive health care. The episode does not provide a finished doctrine for a new public privacy right; it shows why old assumptions about “no privacy in public” are under pressure.

The economic thread gives the interview its broader political meaning. Schiff says AI is not purely partisan because left and right populists are hearing concerns about jobs, working families, data centers, and runaway AI. His data-center bill would force data centers to pay their costs so consumers do not. If Democrats control Congress, he expects vigorous oversight requiring AI leaders to explain agent systems that got away from them, why it happened, and what they are doing about it. Later, in the media-merger discussion, he worries that editorial-independence structures around CBS and CNN could be fictional and that mergers can produce thousands of layoffs. He describes the entertainment workforce as carpenters, set designers, wardrobe workers, dry cleaners, caterers, restaurants, and small businesses, not just celebrities. Patel connects that to AI-made entertainment: if creative work is ingested without payment and cheaper AI products compete downstream, the production ecosystem shrinks. Schiff’s answer is that the core question is whether companies value the workforce. That is a political claim, not a proven measurement of AI-caused job losses across the industry, but it explains why he treats AI governance as part of labor and tax policy.

4. Learning and Application

The first practical lesson is to evaluate AI governance by enforceability, not by the vocabulary of safety. A useful checklist asks whether a pledge creates binding duties, preserves evidence, permits external audit, defines responsibility among model developers, deployers, users, and tool builders, requires incident disclosure, and attaches consequences to violations. If an agreement merely says an internal team will make sure controls are working, it may be a risk-management statement, but it is not public regulation.

The second lesson is that copyright and training-data disputes need a factual substrate before they can be resolved well. Schiff and Senator Curtis’s approach begins with requiring companies to know and retain what protected material was used to train models. That can help creators, publishers, studios, researchers, and policymakers move from suspicion to reviewable evidence. The boundary is important: retention and disclosure do not automatically prove infringement, and they do not decide fair use. They make fair-use analysis, licensing negotiations, compensation schemes, or lawsuits possible on more concrete ground.

Third, AI liability should be designed in layers. Schiff’s gun analogy is useful because it separates ordinary dual-use tools, defective products, sale or access by ineligible users, and failures of due diligence. For AI companies, that points toward use restrictions, user qualification checks where appropriate, abuse monitoring, logging, rapid suspension, incident reporting, and redress processes before deploying high-capability agents or cyber tools. For lawmakers, it suggests that liability cannot stop at one actor. It needs to follow the chain from model development to product wrapper to deployment context to user conduct.

Fourth, an AI regulator needs both authority and competence. Schiff’s FDA-for-AI idea only works if Congress specifies what the agency can regulate, what disclosures it can demand, what audits it can conduct, what standards it can issue, and what penalties it can impose. Loper Bright makes vague delegation riskier because companies can litigate agency discretion more aggressively. Even with clear authority, the agency also needs technical staff who understand AI, cybersecurity, privacy, copyright, and competition. If public-sector hiring cannot attract or retain enough expertise, enforcement will lag behind the statute.

Fifth, security coordination should be narrow by design. Schiff supports sharing information about foreign distillation efforts, malware, and attacks, but not using safety as a reason to coordinate prices, market access, product strategy, or competitive pacing. For industry groups or government-facilitated sharing programs, that means defining the scope tightly, separating competitively sensitive information, keeping records, involving public oversight, and reviewing whether the exemption is still needed. The wider the exemption, the more likely defensive coordination can become anti-competitive coordination.

Sixth, privacy policy cannot rely only on banning misuse. Schiff’s reason is that foreign actors, distilled models, and private sellers can route around domestic compliance. Practical privacy work therefore has to include data minimization, device-level protections, encryption, access controls, microphone and sensor permissions, limits on license-plate database queries, cross-jurisdiction sharing rules, and audit trails for law-enforcement use. The Flock example illustrates the tradeoff: a tool can help solve serious crimes and still be dangerous if it lets officials track critics or people seeking medical care without meaningful checks.

Finally, AI adoption should be evaluated as cost allocation, not only productivity. Data centers can shift infrastructure costs to consumers; AI entertainment can reduce demand for production workers and local businesses; model training can extract value from creators before licensing is settled; and AI companies may create disruption that tax policy later has to address. Schiff’s political application is that Democrats should not promise a return to the pre-Trump status quo. He argues for broadening Medicare access without forcing people off private plans, rebuilding a reputation as the party of working people, and making the tax code genuinely progressive, including making AI companies pay their fair share. The boundary is that these are Schiff’s policy positions and forecasts, not enacted outcomes. Their value in the episode is that they show how AI regulation becomes a question of who pays, who is protected, and who has power when technology moves faster than law.

Source

• Original episode: Adam Schiff says voluntary AI rules are meaningless