Former OpenAI Researcher Warns AI Is Not Loyal

Daniel Kokotajlo, founder of the AI Futures Project and a former OpenAI researcher, gave a full interview to Business Insider in which he said, "AI is not loyal to us," according to Business Insider. The interview, published May 12, 2026, covers Kokotajlo's explanations of AGI and superintelligence, his view that AI agents could mark a turning point for autonomous capability, and his warnings about risks if the AI development race proceeds without stronger safeguards, per Business Insider. The piece reports that Kokotajlo also outlined steps he thinks governments and companies can take to reduce the chance of losing control, as described in the interview.

Industry-pattern observations: Discussions that foreground AI agents typically focus on three technical vectors practitioners watch: persistent state and long-horizon planning, automated action in the real world, and scaling of reward-seeking behavior. These vectors raise engineering questions about safe reward specification, robustness-to-distribution-shift, and monitoring of emergent subgoal-seeking behavior.

Editorial analysis: The concerns Kokotajlo voices mirror recurring themes in AI safety and governance discourse: alignment gaps between model objectives and human values, the governance challenge of competitive pressure, and the practical difficulty of implementing enforceable safety controls at scale. Public warnings from former researchers and safety advocates have repeatedly pushed regulators and firms to consider verification, red-teaming, and deployment gating as mitigations.

Editorial analysis: Observers should track three measurable indicators: adoption of standardized evaluation suites for agentic behavior, company disclosures about deployment safeguards and incident reporting, and any regulatory proposals that mandate external audits or operational safety checks. Progress on these indicators will affect how engineering teams prioritize runtime monitoring, access controls, and interpretability tools.

Editorial analysis: For engineers and ML teams, the interview reinforces the value of investing in robust monitoring, adversarial testing for agentic behaviors, and clear operational limits on action-taking systems. These are generic best practices drawn from industry experience with safety-critical systems, not claims about any single organization's internal roadmaps.