Stuart Armstrong Proposes Value Generalisation as Neglected Alignment Research Direction

Stuart Armstrong published a three-part series on the Alignment Forum on July 29, 2026, arguing that 'value generalisation' — the ability to reliably extrapolate human values to novel situations — is a structurally missing capability in current AI systems that scaling alone will not deliver [2][1]. He contends that LLMs pattern-match rather than genuinely generalise, making them untrustworthy for autonomous delegation even where raw capabilities exist [1]. Armstrong proposes a commercial program to build 'pre-aligned AIs' whose moral capabilities grow architecturally alongside their empirical capabilities, and is soliciting collaborators and funding [3][1].

If Armstrong's diagnosis is correct, current AI systems cannot be safely delegated autonomy even when they appear capable, because the gap is structural rather than a matter of insufficient scale. His proposed commercial path — treating alignment as a market differentiator rather than an academic publication target — represents a distinct strategy whose success or failure would be directly legible in adoption outcomes.

Stuart Armstrong, co-founder of Aligned AI and a longtime alignment researcher, published a three-post series on the Alignment Forum on July 29, 2026, arguing that 'value generalisation' is a neglected and structurally critical alignment problem [1]. His central claim is that current AI systems — including state-of-the-art LLMs — cannot extrapolate human values reliably to situations outside their training distribution. They pattern-match confidently, but their outputs in novel territory are not reliably what users want [1]. Armstrong frames this not as a gap that more scale or better prompting will close, but as a capability that must be deliberately engineered.

In the second post, Armstrong introduces 'strong generalisation' — a cluster of human cognitive abilities including situational awareness, adaptive world modeling, and genuine out-of-distribution reasoning — and argues LLMs lack it entirely [2]. He offers an explanation for a known pattern: subject-matter experts extract far more value from LLMs than novices because experts supply the generalisation capability the model lacks, serving as the missing cognitive layer [2]. He also argues that benchmark saturation on long-horizon tasks reflects pattern-matching on benchmark structure rather than genuine generalisation, meaning benchmark performance does not reliably predict real-world autonomous capability [2].

The third post describes Armstrong's proposed solution: 'pre-aligned AIs' that bind moral and empirical concepts architecturally during learning, so that increased capability automatically produces stronger ethical alignment rather than enabling misalignment [3]. Armstrong argues this inverts the standard capability-alignment tradeoff: ethical users get the full benefit of a learning AI, while bad actors must deliberately cripple theirs to prevent the system from recognising harmful tasks [3]. The commercial logic follows from this — organizations that let the AI learn freely get better outcomes, creating a market incentive for alignment that does not depend on regulatory pressure.

Armstrong is explicit that he prefers commercial deployment over academic publication as the mechanism for adoption, arguing that alignment-relevant findings published academically tend to be ignored or mined only for capability-relevant components while alignment components are discarded [1]. As of the first synthesis pass, no other named voices have responded publicly to the series, leaving it as a proposal seeking engagement rather than an established debate.