What happened
- Science Advances published online on 28 October 2026 the paper “Machine psychology for language models”, signed by eight researchers.
- The authors include Thilo Hagendorff of the University of Stuttgart, Ishita Dasgupta of Google DeepMind in London and Marcel Binz of the Helmholtz Institute for Human-Centered AI in Munich.
- The paper proposes a “machine psychology”: studying large language models (LLMs) with experiments inspired by those psychology uses to understand human cognition and behaviour.
- It summarises theoretical perspectives, experimental paradigms and computational analysis techniques, reviews work that already uses this approach and suggests good practices.
Why it matters
- Most model evaluations today measure performance: how many questions they get right. The authors argue that this is not enough to understand how models reason or to discover abilities that appear without being explicitly trained.
- Psychology experiments seek to explain behaviour, not just score it. Applied to a model, they can reveal biases, shortcuts or patterns that a multiple-choice test does not show.
- The paper also warns that methods designed for people do not transfer automatically to machines, and devotes a section to those caveats.
The number
8 authors. The paper brings together researchers from German universities, an AI research institute and Google DeepMind.
Context
The authors argue the approach will become more useful as models grow more powerful, more opaque and multimodal, and are built into agent settings that act in the real world.
The idea connects with recent work that measures model behaviour beyond accuracy. Mamífero reported how Microsoft measured agents’ judgment with a taste test, how a Swedish study assessed the political neutrality of six models across 148,000 texts and how another benchmark measured where an ad bothers people least inside a chat. On safety, OpenAI proposed halting a training run if nobody answers an alert.
What’s next
- The paper does not present a new experiment. Its value lies in organising a fast-growing field and proposing shared criteria.
- The open question is whether results from these tests hold across versions of the same model, which change every few months.
- For teams evaluating models in companies or the public sector, the text offers a repertoire of methods already tested in psychology.
Bottom line
If language models are going to make decisions with real effects, knowing how many answers they get right is not enough. We need to understand why they answer the way they do. The authors propose using a century of psychology methods for that, without forgetting that a machine is not a person.
Sources
- Hagendorff, T., Dasgupta, I., Binz, M. et al., “Machine psychology for language models”, Science Advances, published online 28 October 2026, DOI 10.1126/sciadv.aec1912. Date and abstract confirmed on Crossref; the journal page could not be read because automated access was blocked.
- Cover image: Wellcome Collection, CC BY 4.0, via Wikimedia Commons
Edited by Rodrigo Cornejo. How we select and verify: who writes these notes.
