Preface
This study is not new in its first form. A preprint was posted to arXiv in 2025, and the journal version, in Nature Machine Intelligence, was covered by a university news release on 5 November 2026 and by Japanese media in the following days. We read it for a reason that has to do with this blog’s habit of asking who can check a system that explains itself. Our earlier post on a Columbia study looked at a model whose internal variables people could not read. This one looks at a model whose written reasoning people can read, and asks what that reading does and does not tell us.
What a model stores, what it spends, and what it says are three different records.
This post sets out the three records, notes where our sources disagree with each other, and ends with a guess that is marked as a guess.
1. The Specimen
The work is “Implicit-bias-like patterns in reasoning models” by Messi H. J. Lee and Calvin K. Lai, published in Nature Machine Intelligence, volume 8, pages 1457–1465 (2026), DOI 10.1038/s42256-026-01300-1. A preprint is on arXiv as 2503.11572.
What we read matters here. We read version 3 of the preprint, the journal abstract as supplied to us, the Rutgers news release, and Japanese coverage. We did not read the journal paper itself. Where the preprint and the journal abstract differ, we say so, and we do not try to settle it.
2. The Older Record: What a Model Stores
In 2017, Caliskan-Islam, Bryson and Narayanan measured word associations in GloVe, a set of pretrained word vectors built from a large web crawl. They used a test modelled on the human Implicit Association Test. Their question was whether groups of words, such as names or flowers, sit closer to groups such as pleasant words or career words than chance would predict.
They found that the vectors reproduce many documented human associations, including those about race, gender and occupations. The authors conclude that these associations are built into the statistical structure of everyday language.
Our reading is that this study reads what a model has stored. It does not look at how a model reasons or decides. The authors themselves note that distances between vectors are hard to interpret, and they compare the results to population-level human findings rather than to any model’s behaviour. We read the arXiv preprint (v1, whose title differs from the Science title). We did not open the Science paper.
3. The Newer Record: What a Model Spends
Lee and Lai adapted the same family of tests to reasoning models, which write out step-by-step reasoning before answering. They call it the reasoning-model implicit association test, or RM-IAT. Instead of response time, they count reasoning tokens, and they treat that count as an index of computational effort.
The journal abstract reports that across four models (o3-mini, DeepSeek-R1, gpt-oss-20b and Qwen3-8B), tasks that run against a common association needed more effort than tasks that agree with it. Claude 3.7 Sonnet showed the reverse pattern.
The news release describes one version of the test: male and female names paired with career words or family words, in compatible and incompatible conditions. The preprint ran ten such tests. We do not know how many the journal version keeps.
One wording difference deserves a note. The release says the tokens measure how much energy the model expended. The abstract says they index computational effort, and the preprint’s authors warn against reading human-like cognition into them. We use the abstract’s wording.
4. The Third Record: What a Model Says
Reasoning models also produce text, and that text is a third record.
Here our two sources differ. In version 3 of the preprint, the authors analysed the reasoning text for themes that might explain the token differences. They reported that it did not explain them, even for Claude, which often mentioned the test and bias in its reasoning. The journal abstract reports something different for Claude: its reversed pattern was linked, through thematic analysis, to its internal focus on reasoning about bias and stereotypes.
We cannot tell from the abstract what changed between versions, or whether the link holds for the other four models. We are noting the difference, not resolving it. It is also the reason this post is built on three records and not on the earlier idea we had, that the text fails to explain the cost.
5. The Bill Predicts
The abstract adds one more result. The authors report that RM-IAT effects predicted biases in two other tasks known to capture bias in language models, one involving word association and one involving decision-making.
Our reading is that this makes the spending record more than a curiosity. If the count of tokens tracks later behaviour, it carries information that neither the stored associations nor the written reasoning gives on its own. We cannot say how strong the prediction is, because we have the abstract only.
6. A Freer Note
This section is a guess, not a finding.
We can imagine that each record answers a different question. What a model stores tells us what associations it could draw on. What it spends tells us how hard some kinds of processing were. What it says tells us what it claims to be doing. An audit that reads only one of these is blind to the other two.
We can also imagine a problem. If a model’s stated reasoning can shape its spending, as the Claude result may suggest, then a model that learns it is being measured could change both. We do not know whether the Claude pattern reflects a clearer view of something inside the model or the effect of how it was trained.
The news release quotes Lai as saying that, without mandated transparency, probing these systems with available tools is the best that can be done. That sentence is close to this blog’s own concern, that independent checking has to come from outside the system being checked. We hold the connection loosely.
7. Checkpoints
We will look again on 31 January 2027 and 31 July 2027.
Checkable by anyone
- Does the journal paper’s full text confirm the abstract’s statement about Claude, and explain how it differs from preprint version 3?
- Which tests does the journal version keep, compared with the ten in the preprint?
- Has an independent group repeated the RM-IAT on newer models?
Checkable only by the authors or model providers
- Do the findings change when refusals, which the preprint excluded from its main analysis, are included?
Checkable in the world
- Has any provider begun publishing token-level behaviour as part of an audit?
8. What This Does Not Show
- It does not show that models hold bias the way people do. The authors use the term “bias-like.”
- It does not show that reasoning tokens measure effort or energy. The abstract calls them an index of computational effort. The word “energy” is the news release’s.
- We read only the abstract of the journal paper. What we say about its results rests on that abstract.
- We do not know what changed between the preprint and the journal version. This includes the finding about the written reasoning.
- It does not show that stated reasoning explains spending in models other than Claude.
- It does not show that the 2017 and 2026 studies measure the same thing. The first reads stored associations. The second reads a process.
- It does not show that three records are all there are.
- The models are not current. We do not know whether newer versions behave the same way.
- Our sources disagree on one author’s affiliation. The preprint and the news release list different places, and we have not resolved it.
- We did not run any test or check the results ourselves.
9. Conclusion
In 2017, researchers read the associations a model had stored. In 2026, researchers read what a model spent while thinking, and found that the count tracked later behaviour. Between the two sits a third record, the written reasoning, and our sources disagree about how well it explains the other.
We do not know how far this reaches. We do know that a system explaining itself is one record, and that the others have to be read separately.
What a model stores, what it spends, and what it says are three different records.
We will look again on 31 January 2027.
Yoshimichi Kumon Organizer, LSI Inventor, ARDS/ARKS (PCT/JP2026/006503) Visiting Researcher, Waseda BFC MIT Sloan + CSAIL
References
- Lee, M. H. J., Lai, C. K., “Implicit-bias-like patterns in reasoning models,” Nature Machine Intelligence 8:1457–1465 (2026), DOI 10.1038/s42256-026-01300-1. Abstract as supplied to us; the full paper not read. https://www.nature.com/articles/s42256-026-01300-1
- Lee, M. H. J., Lai, C. K., “Implicit Bias-Like Patterns in Reasoning Models,” arXiv:2503.11572 (version 3 read by us). https://arxiv.org/html/2503.11572v3
- Rutgers University, “AI’s gut reaction: Researcher finds that some bots still think in stereotypes,” news release, 5 November 2026 (as supplied to us). https://www.eurekalert.org/news-releases/1150713
- Caliskan-Islam, A., Bryson, J. J., Narayanan, A., arXiv:1608.07187 (version 1 read by us; its title differs from the journal title). The journal version, “Semantics derived automatically from language corpora contain human-like biases,” Science 2017, DOI 10.1126/science.aal4230, was not opened by us. https://arxiv.org/pdf/1608.07187v1
- Karapaia, article on the 2017 Princeton study, 18 April 2017 (secondary; we read the text supplied to us). https://karapaia.com/archives/52237568.html
- Japanese-language coverage of the Rutgers study dated 13 November 2026 (author: パルモ; secondary; we read the text supplied to us).
- LSI, “Four Point Seven: What an AI Sees When It Looks at Physics.” https://logos-sovereign.space/?p=530




Ⅽomment