Alignment Is 'Not on Track': A Rare Candid Signal From Inside Safety Discourse
A named editorial warning that AI alignment is failing shifts the safety conversation from technical debate to institutional credibility crisis.
1. Alignment Is "Not on Track": A Rare Candid Signal From Inside Safety Discourse
Import AI 461, published June 2026, flags a named-source claim that AI alignment is "not on track." The framing is notable for what it is not: a speculative blog post or an anonymous forum complaint. Import AI, edited by Jack Clark, co-founder of Anthropic, carries institutional weight inside the safety research community. The newsletter does not typically editorialize in those terms. That makes this a signal worth tracking separately from its technical content.
The strategic read here is about credibility, not just capability. Anthropic has built its entire market position on the claim that safety-first development is both possible and commercially viable. OpenAI rebuilt its public narrative around safety governance after the November 2023 board crisis. Google DeepMind publishes alignment research as a trust signal to regulators in Brussels and Washington. If "not on track" language from a credible insider voice gains traction, it does not just pressure safety teams. It pressures the regulatory framing that has protected frontier labs from harder legislative constraints. The EU AI Act's risk tiers, NIST's AI RMF, and the UK's frontier AI commitments all assume that alignment progress is measurable and directional. A public credibility fracture complicates that assumption in ways that are harder to contain than a benchmark failure.
Watch for how Anthropic, OpenAI, and DeepMind respond in their next round of safety publications or policy testimony. If the "not on track" framing spreads beyond Import AI into congressional briefings or EU technical working groups, the political cost of the current self-regulatory posture rises sharply. The labs have been ahead of regulators by publishing their own evals. That advantage shrinks when their own community questions whether the evals mean anything.
Source: Import AI 461: "Alignment is not on track"; FrontierCode; and synthetic research interns