Anthropic Researcher Argues AI 'Overcorrection' Could Address Historical Injustices

Summary

Amanda Askell, a philosopher hired by Anthropic to develop the company's AI ethics framework, has surfaced in recent news coverage advocating that artificial intelligence systems could intentionally 'overcorrect' or apply positive discrimination toward marginalized groups to combat historical inequities. In a 2023 research paper co-authored with other AI researchers, Askell argued that companies might benefit from training models to exhibit bias in favor of historically disadvantaged groups, though she emphasized this approach would require deliberate human input to modify AI responses.

The paper presents experimental evidence showing how training methodology affects bias outcomes in large language models. A 175-billion parameter model discriminated against Black students by 3% when trained without human corrections, but showed 7% positive discrimination toward Black students when trained with additional human feedback and chain-of-thought reasoning. Notably, the paper explicitly states that 'positive discrimination in favor of black students may be considered morally justified'—contingent on alignment with local laws and proper governance.

Askell's research highlights a critical tension in modern AI development: how to build systems that abandon harmful human biases while potentially introducing intentional corrective ones. The resurfacing of this paper reflects growing industry debate over what 'ethical AI' truly means. Anthropic has positioned Claude as the ethical choice, with a constitution designed to ensure 'good, wise and virtuous' decision-making, yet internally grapples with complex questions about whether AI systems should actively discriminate to advance historical justice.

• Askell emphasizes that overcorrective approaches require explicit human oversight and alignment with applicable laws to be ethically justified

Editorial Opinion

Askell's argument that AI systems should sometimes deliberately discriminate in favor of marginalized groups is philosophically provocative but deeply consequential. While the intent to address historical injustices is admirable, embedding intentional bias into AI systems—even 'positive' bias—risks normalizing algorithmic discrimination and could paradoxically undermine public trust in AI's impartiality. The critical insight isn't that overcorrection is always justified, but that humans must remain actively involved in deciding when it is. This puts the burden on regulators, companies, and society to establish clearer ethical guardrails around when and how AI systems should be designed to make discriminatory decisions.