The accountability mechanism that eventually broke the tobacco industry open was not a government investigation. It was a filing cabinet.
Tobacco: The Paper Trail
From the 1950s, tobacco companies held internal research confirming that cigarette smoke caused cancer. They suppressed it. The documents took nearly four decades to reach courts. They arrived through state attorneys general subpoenas, whistleblowers inside the industry, and litigation that generated discovery obligations the companies could not refuse. The mechanism was slow. It was not structurally blocked.
The 1998 Master Settlement Agreement, under which the four largest US tobacco manufacturers committed to pay approximately $206 billion to state governments over 25 years, was built on that paper trail (NAAG, 1998). Adult smoking rates in the United States fell from 42.4% in 1965 to 11.6% by 2022 (CDC, 2025). The industry was restructured, taxed, and constrained in ways it had spent decades preventing.
The documents made it possible.
Social Media: The Same Arc, Compressed
In 2021, a single Facebook employee walked out of the company with internal research. Frances Haugen’s disclosure showed that Meta knew its platforms caused psychological harm to teenage girls and had made no substantive change. It created an evidentiary record outside the company’s control.
In March 2026, a New Mexico jury found Meta liable for failing to protect minors from predators on its platforms. A day later, a Los Angeles jury found Meta and YouTube negligent for designing features that caused documented psychological harm to young users (Larson, 2026). The plaintiff in the Los Angeles case had been using the platforms since she was six years old. More than 1,200 lawsuits have since been filed by school districts. Attorneys general in more than 40 states have brought suits modelled on the tobacco state cases (Larson, 2026).
“Apple settled a consolidated fraud suit over inflated artificial intelligence capability claims for $250 million, against a fiscal year 2025 profit of $112 billion.”
Apple settled a consolidated fraud suit over inflated artificial intelligence capability claims for $250 million, against a fiscal year 2025 profit of $112 billion (Larson, 2026; Apple, 2025). The Federal Trade Commission secured a $2.5 billion settlement from Amazon over manipulative subscription practices, which it described as one of the largest settlements in the agency’s history (Larson, 2026). It represented 3.2% of Amazon’s annual profit. No major technology company has been structurally separated as a result of any proceeding.
Tobacco’s long-term accountability came because the cumulative cost of litigation eventually altered industry strategy. At the settlement figures recorded so far, that threshold has not been approached for any major technology platform.
AI: The Evidence Is Owned
Dario Amodei, CEO of Anthropic, has described on the record the technology his company develops as potentially transformative and dangerous in ways that could prove irreversible. Anthropic’s published model specifications, safety research, and red-teaming frameworks document in unusual detail what the company believes about its own product.
The training data on which a model’s capabilities are built is proprietary. Internal evaluations assessing what a model can do, what it will refuse, and where its limits can be circumvented exist inside the companies. Capability assessments conducted before deployment, red-team findings, documented failure modes: none are accessible to external researchers, regulators, or courts without the company’s cooperation.
Tobacco’s internal research could be reached through legal discovery because it existed on paper in filing cabinets. Meta’s internal research became accessible because one employee chose to carry it out. AI’s equivalent has no comparable exposure point.
Meaningful accountability for AI harm would require access to training datasets, model weights, pre-deployment evaluation results, and the internal documentation of what was known and when. None of this is currently reachable through existing legal or regulatory frameworks without the company’s consent. This is analytical inference: no AI accountability case has yet tested it at the frontier.
The European Union’s AI Act, which came into force in stages from 2024, establishes risk categories and imposes obligations on providers of high-risk systems, including requirements for transparency, human oversight, and accuracy (European Parliament, 2024). It does not grant regulators access to training data or model weights. The United States has no equivalent federal legislation.
Global AI infrastructure investment reached approximately $423 billion in 2025 (UBS, 2025). Anthropic’s Series H round, closed in May 2026, raised $65 billion at a post-money valuation of $965 billion (Anthropic, 2026). Tobacco concealed what it knew. The companies building the most capable AI systems have published what they know, in detail, and retained ownership of the evidence that would allow anyone else to verify it.
The tobacco settlement of 1998 changed the industry. It did not require the industry’s cooperation to arrive. The evidence was found.
Signal source: Larson, R. (2026) ‘Is Big Tech Facing Its Big Tobacco Moment?’, Jacobin, 20 June — Apple (2025) FY2025 Q4 Consolidated Financial Statements — Anthropic (2026) Series H Funding Announcement, May — CDC (2025) Adult Cigarette Smoking Data — UBS (2025) Global AI Capital Expenditure Estimates — NAAG (1998) Tobacco Master Settlement Agreement — European Parliament (2024) Regulation (EU) 2024/1689, Artificial Intelligence Act.