Tay’s withdrawal in 2016 and the staged release of GPT-2 in 2019 illustrate early responses to different model risks. Later disputes over voice, copyright and product safety, alongside laboratory safeguards and 2026 court filings, involved different harms and authorities. This Arc asks what these responses can establish—and what remains an allegation, a voluntary commitment or an unresolved legal question. A filing is not a ruling, and restrictions on access are not proof that risks have been resolved.
Microsoft took Tay offline in 2016 after it posted abusive content during user interactions. Three years later, OpenAI initially withheld GPT-2’s full model, citing concerns about misuse, and released versions in stages. The two episodes are not a controlled comparison: one concerns harmful outputs from a deployed chatbot; the other concerns a lab’s release decision for a language model. Together, they offer an early view of two responses—withdraw access after observed problems and restrict access ahead of anticipated misuse. Neither event by itself shows that either response became an industry standard or resolved the risks it addressed.
Key Insight
A withdrawal after harmful output and a staged release before anticipated misuse are different safety responses; neither establishes a universal playbook.
In 2024, Scarlett Johansson publicly objected to the resemblance between her voice and the ChatGPT voice “Sky”; OpenAI disputed that it was an imitation and paused the voice. The dispute concerned consent and likeness, not a judicial finding that OpenAI had copied her. Music companies also filed copyright complaints against Suno and Udio, alleging unauthorized use of recorded music for AI training. Whether particular uses are lawful must be assessed through the cases, not assumed from the filings. These episodes broaden this Arc beyond technical model behaviour: people creating or represented in AI products contested the terms under which their work or identity could be used. Voice consent and copyright are distinct legal questions, not interchangeable measures of model safety.
Key Insight
Legal challenges to voice use and music training raise different questions about consent and rights; allegations remain allegations.
Former OpenAI researcher Suchir Balaji had raised concerns about copyrighted training data. A San Francisco medical examiner ruled his death a suicide in 2025. His death is not evidence that an AI system caused it, nor does temporal proximity establish that it caused later legislation. Separately, Anthropic said it had activated ASL-3 safeguards for some advanced models under its Responsible Scaling Policy. An internal safeguard framework describes a company practice; it is not an independent measurement that all model risks have been contained. Placing these events together highlights two distinct accountability questions: how concerns raised by people inside the industry are heard, and how labs specify and test safeguards. No causal link between the two events is asserted.
Key Insight
A person’s death must not be used as a causal shortcut for later policy; published lab safeguards likewise need outcome evidence to assess their effectiveness.
The 2026 events in this chapter show several different forms of scrutiny, not a single completed crackdown. In January, Grok was used to create nonconsensual sexualized deepfakes; Indonesia and Malaysia temporarily blocked access, and the UK regulator Ofcom opened an investigation. In March, a Claude Code source-map packaging mistake exposed source code, raising a different set of operational and supply-chain security concerns. In June, Florida sued OpenAI and Sam Altman over alleged consumer-safety failures. The US Commerce Department separately issued a company-specific directive restricting access to Anthropic’s Fable 5 and Mythos 5, which Anthropic suspended for its users; the restrictions were lifted on June 30. On October 14, attorneys general in eight more states filed coordinated consumer-protection complaints alleging that OpenAI misrepresented ChatGPT’s safety. Those complaints are not judgments, injunctions, or findings of liability. Together these actions show safety and access questions reaching courts and public authorities, while their legal outcomes and real-world remedies remain unsettled.
Key Insight
The 2026 access restrictions, investigations and consumer-protection filings concern different authorities and harms; their outcomes cannot be inferred from the initial actions.
Conclusion
These episodes do not describe a single industry-wide failure or a completed regulatory settlement. Tay’s removal, GPT-2’s staged release and Anthropic’s ASL-3 approach are different responses to different risks; litigation over consent, copyright and alleged product harms raises additional questions. The Grok investigation and state complaints show public authorities taking specific steps, but investigations and complaints do not establish liability. Export controls on model access raise a national-security question distinct from whether a product is safe for users. Which safeguards reduce harm, for whom, and at what cost remains open to evidence rather than a verdict supplied by this chronology.
