It’s ‘more likely than not’ humanity loses control: Former AI insiders testify safety fixes may be ‘duct tape that will fall off later’

Former OpenAI and Anthropic researchers warned New York City Council that humanity is more likely than not to lose control of advanced AI, citing weak safety evaluations and risks from AI-written code; a former Google DeepMind researcher put the chance of an AI takeover at roughly one in three. The council is weighing bills requiring independent validation and human shutdown controls for AI systems sold or deployed in the city, with fines of $25,000 per violation, alongside whistleblower protections and liability provisions. The hearing also highlighted that no company representatives raised a hand when asked whether their companies insured against catastrophic AI risks.

Analysis

The investable signal is not a near-term change in AI demand; it is a possible shift in who bears deployment risk. If New York’s proposals advance, independent validation, human shutdown controls and whistleblower incentives could add recurring compliance costs and slow product launches. Those costs may favor scaled incumbents over smaller labs over time, but could also expose large platforms to more visible enforcement and litigation. The city-level proposal is not yet a national regime: scope, enactment, enforcement and potential legal challenges remain key uncertainties.

For Alphabet, the hearing adds a company-specific reputational and policy overhang because testimony criticized its handling of a proposed Pentagon deal. Treat that account as testimony, not independently established evidence of corporate liability. Meta was not singled out in the same way in the testimony described, but would still face sector-wide compliance exposure if rules apply broadly.

Days: headline risk is likely to be sentiment-led unless the bills progress. Over 1–3 months, track committee action, bill language and whether other jurisdictions adopt similar requirements. Over 6–18 months, binding external audits or liability standards could raise fixed costs, constrain high-risk deployments and advantage firms able to fund compliance—while reducing the value of speed-to-market. The contrarian point: dramatic catastrophe testimony is not itself evidence of imminent regulation or lost commercial demand; the market may overprice the near-term policy impact while underpricing cumulative compliance costs if rules diffuse. Falsifiers include stalled or materially narrowed bills, limited applicability to major platforms, or no observable increase in safety-related spending or launch delays.

Key Decisions for Investors

• No high-conviction directional trade on this hearing alone. Avoid adding to Alphabet solely on the assumption that this is transient reputational noise; reassess if the council advances binding requirements or the Pentagon-contract controversy produces further company-specific scrutiny.
• Set an event alert for NYC committee action and the final bill text. Verify which systems, providers and deployments are covered, the validator and shutdown requirements, and whether enforcement could reach major platforms; do not price a broad compliance burden before scope is clear.
• For a relative-value watch, compare Alphabet with Meta rather than shorting the AI complex: Alphabet has the more direct company-specific testimony overhang, while both retain broad regulatory exposure. Consider a modest long-Meta/short-Alphabet expression only if subsequent policy developments remain sector-wide and Alphabet-specific scrutiny persists; exit the thesis if the bills stall or the company-specific overhang fades.