The hidden costs of ‘helpful’ AI
Article Date: 31 March 2026
Article URL: https://www.nature.com/articles/d41586-026-00966-2
Article Image: https://media.nature.com/w300/magazine-assets/d41586-026-00966-2/d41586-026-00966-2_52198138.jpg
Summary
Sylvie Delacroix argues that AI systems that appear to help professionals can quietly erode collective professional judgement by narrowing how uncertainties and values are framed and debated. A collaborative chess experiment shows that compatibility with human decision-making — not raw computational power — produces better teamwork. This reframes interpretability: it is not only about understanding AI output but whether humans can act on it productively. In many real-world fields (healthcare, law, education), objectives and values evolve, so AI that fixes uncertainties into probabilistic formats can make ethical and interpretive judgments invisible, effectively de-skilling professions and constraining how communities decide what counts as good practice.
Key Points
- A chess experiment found weaker, human-compatible AIs outperformed superhuman AIs when paired with a human-like partner — showing compatibility matters more than raw power.
- Interpretability should be judged by whether users can act on AI outputs, not merely whether they can read them.
- AI that frames uncertainty solely as probabilities can obscure value-laden, interpretive professional judgements (example: whether to document suspected domestic abuse).
- Even individually interpretable tools can narrow the range of questions and values a professional community considers, leading to collective de-skilling.
- Designing AI for evolving professional practices requires attention to how tools shape debate about values and what counts as good judgement.
Why should I read this?
Look — this piece saves you the bother of sifting through techno-optimism. If you work with or deploy AI in professional settings, it flags a sneaky risk: helpful-looking tools can quietly shrink professional judgement and reshape what counts as acceptable decisions. Important if you care about ethics, safety, or preserving expertise.
Author style
Punchy and direct: Delacroix pushes a crucial point — AI’s usefulness isn’t only technical, it’s social and ethical. If you care about how professions evolve (not just how models perform on benchmarks), read the detail.
Context and relevance
The article matters because AI adoption is shifting decision-making across healthcare, law and education. As systems standardise uncertainty into scores and probabilities, they risk silencing the interpretive, ethical debates that maintain professional standards. This ties into broader debates about AI governance, interpretability and the preservation of human judgement in high-stakes domains.