by

Markus Schaffner, Senior Developer/Architect for Java and SAP Commerce | Data Scientist, Coop, Switzerland

Benno Torgler, School of Economics and Finance & ARC Training Centre for Behavioural Insights for Technology Adoption (BITA), Queensland University of Technology

Your News Explainer on “AI scientists” (see Callaway, Nature 10 July 2026) shows AI moving beyond prose into research workflows. Papers, code and figures increasingly emerge from human–machine systems, yet assessment still treats finished artefacts as evidence of individual competence. Generation is becoming abundant; accountable judgement remains scarce.

Minsky’s layered account in The Emotion Machine offers a useful analogy: critics and selectors activate higher-level ways of thinking. As AI absorbs routine scholarly operations, assessment should shift towards choosing consequential questions, recognizing failure and transferring arguments to altered conditions.

But this upward move alone is unsustainable. AI can also generate critiques, rationales and plausible defences too. Most standardised higher-order task can become another signal to imitate. Evaluation should probe responses to changed assumptions, contradictory evidence and failed replication.

This sharpens Feynman’s (1965) Nobel lecture warning that polished papers conceal blind alleys and gives our earlier “4R” call for revelation new meaning (see Pagan and Torgler, Nature 522, 34, 2015). In AI-mediated research, revelation should be auditable in the workflow, not merely narrated in the paper: contemporaneous records of consequential choices, rejected alternatives, human overrides and validation decisions.

AI does not abolish the need for researchers’ excellence; it relocates it from output to accountable judgement.

Illustration:

AI-generated illustration of scientific judgement in an AI-mediated research environment. Created using OpenAI’s image-generation tool in ChatGPT, based on the authors’ concept and creative direction, 2026.