An analysis of AI-generated content at the Mechanistic Interpretability Workshop
Introduction
Over the past few years, AI tools have become useful for conducting technical AI research. In the early ChatGPT era (~2023–2024), chat assistants were maybe useful as sounding boards for research ideas, or as editors for polishing a paper draft. In the more recent Claude Code era, competent coding agents can code up and run experiments; with enough direction, they can do much of the technical heavy lifting on a PhD-level research project (Schwartz, 2026); and in well-defined settings, they can autonomously try new things and iterate without a human in the loop (Karpathy, 2026). While these tools can unlock researchers to be more productive and widen their ambitions, they can also be abused – anyone can now hand an agent a research prompt, tell it to run the experiments and write up the results in a LaTeX document, and get back an artifact that roughly resembles a conference paper in form. With this steep of a change in the research process, it seems important to study how it is affecting technical research and peer review.
The Mechanistic Interpretability Workshop has run three times in the past two years – at ICML 2024, NeurIPS 2025, and ICML 2026. The most recent iteration felt quite different from the first two – submissions more than doubled from the previous edition (which was only seven months earlier), and a noticeable share of the submissions seemed to resemble “AI slop”. We, the workshop’s program chairs, collectively read through hundreds of abstracts; despite being subject-matter experts in mechanistic interpretability, we often found ourselves reading the same abstract three times and still not understanding what it was claiming or contributing. While we made efforts to desk reject the clearest cases of incoherent papers,1 the prevalence of AI-generated content seeped through to reviewers, and reviewers frequently reported frustration with having to spend time reviewing low-effort AI-generated papers.
In an attempt to understand the scope of AI-generated content in workshop submissions, we decided to do some quantitative analysis. Specifically, we ran every submission and every review through Pangram, an AI-text detector, to estimate how much of each was AI-generated. We then performed some basic analysis to gain insight into the scope and impact of AI-generated content in the workshop.
A brief summary of the most interesting observations:
- The number of submissions has grown rapidly. Submissions more than doubled between iterations, reaching 801 in the 2026 edition – up from 320 in 2025 and 143 in 2024.
- Roughly a quarter of submissions were solo-authored. The share rose from 9% in 2024 to 24% in 2026. Authors were also increasingly likely to first-author multiple submissions: 62 people first-authored at least two papers in the 2026 edition, up from just 4 in 2024.
- AI-generated writing is becoming more prevalent. In the 2024 edition, essentially no submissions were flagged as substantially AI-generated. By 2026, about 33% of papers had a majority of their text flagged as significantly AI-generated.
- The top papers are still mostly human-written. Of the 23 spotlight papers, 91% were categorized as entirely or mostly human-written – compared to about 50% in the overall pool of submissions.
- AI-generated text is prevalent in reviews too. About 50% of all reviews contained at least one passage flagged as AI-generated, and in 17%, every passage was flagged. Interestingly, we observe that heavily AI-generated papers receive higher recommendation scores from AI-generated reviews than from human-written reviews.
Submissions are growing rapidly
Accepted (poster)Accepted (virtual poster)Rejected
The number of submissions has more than doubled between each iteration of the workshop, going from 143 to 320 to 801 in two years.
Amid this growth, the 2026 edition introduced a virtual-poster tier. The workshop was allocated a fixed number of in-person poster slots at the conference venue – roughly 200. Rather than reject papers that area chairs judged to meet the workshop’s quality bar simply because of space constraints, we decided to accept them as virtual posters.
To put the workshop’s growth in context, the chart below plots its submissions alongside those of major AI conferences.
NeurIPSICMLICLRMI Workshop
All of these venues are growing roughly exponentially. Much of this reflects a natural expansion of the AI research community; mechanistic interpretability in particular is a relatively new and fast-growing subfield. AI tools that lower the barrier to producing manuscripts are likely also playing a significant role in recent growth, though we don’t attempt to disentangle these causal factors here.
Note that two more doublings would put the workshop at roughly 3,200 submissions – just about where ICLR was in 2022. At that point, the workshop would be operating at the scale of a small conference.
Individual authors are submitting more papers
Solo-authored papers
About 9% of submissions to the 2024 edition were solo-authored – that is, they listed only one author. By 2026, the share had risen to 24% – roughly 200 of the 801 submissions. The growth in solo papers alone accounts for a meaningful share of the overall submission increase.
We can also examine the number of papers submitted by each individual author. In 2024, about 6% of submissions came from someone who was first author on more than one paper in that edition; by 2026, this had risen to 19%, and we even saw a single individual who was first author on 6 separate submissions. The following plots show the full distributions of solo-authored and first-author papers per author.
Authors by solo-paper count
Authors by first-author count
202420252026
As analyzed below, solo-authored papers and papers from repeat first-authors tend to be more AI-generated than average.
How we measure “AI-generatedness”
To estimate how much of each submission’s writing is AI-generated, we use Pangram (v3.3.2), an AI-text detection tool that identifies linguistic patterns characteristic of LLM-generated text. Note that Pangram measures writing style – it can tell us whether the text reads as AI-generated, but it does not directly tell us whether AI was used more broadly throughout other phases of the research.
Our scoring methodology closely follows the NeurIPS analysis of AI-generated papers in their 2026 position-paper track, using the same detector and thresholds. For each paper, we perform the following steps:
- Text extraction. Each paper’s text is extracted from its PDF – we strip out figures, tables, references, appendices, and template boilerplate.3 This leaves just the running prose of the body.
- Chunk scores. The extracted text is sent to Pangram, which splits it into consecutive ~250–350-word chunks and assigns each chunk a score between 0 (human-written) and 1 (AI-generated). Following the NeurIPS methodology, a chunk is considered flagged if its score is \(\ge 0.75\).
- Paper scores. A paper’s score is the percentage of its chunks that are flagged. A score of 0% means none of the paper’s chunks scored \(\ge 0.75\); a score of 50% means half of the paper’s chunks scored \(\ge 0.75\).
For reviews, the same method is applied to each of the review’s written fields individually (each written field is usually just one or two chunks). A full review is similarly scored by the percentage of its chunks above the 0.75 threshold.
AI-generated writing is becoming more prevalent
The following figure groups papers by the percentage of their text chunks that were flagged, across workshop editions.
Fully human0% flaggedMostly human1–25% flaggedLeans human25–50% flaggedLeans AI50–75% flaggedMostly AI75–99% flaggedFully AI100% flagged
The best papers are still mostly human-written
We can group papers by their outcome (e.g., accepted or rejected) and look at the distribution of AI scores within each of these groups. The spotlight papers are almost entirely human-written: over 90% of the spotlight papers have less than 25% of their chunks flagged. At the other end, rejected and desk-rejected papers are the most heavily AI-generated – and there’s a nice gradient across the tiers in between.
Fully human0% flaggedMostly human1–25% flaggedLeans human25–50% flaggedLeans AI50–75% flaggedMostly AI75–99% flaggedFully AI100% flagged
We can also look at acceptance rates split by AI score. With this analysis, we see that papers that read as heavily AI-generated have a much lower acceptance rate, compared to papers read as human-written.
Acceptance rateAcceptance rate (incl. virtual poster)
Solo-authored and repeat-first-author papers tend to be more AI-generated
Earlier we noted that solo-authored papers and papers from repeat first-authors have both grown as a share of submissions. We can check what the distribution of AI-generatedness looks like within these groups, compared to the overall distribution. Both groups skew noticeably more AI-generated than the baseline.
Fully human0% flaggedMostly human1–25% flaggedLeans human25–50% flaggedLeans AI50–75% flaggedMostly AI75–99% flaggedFully AI100% flagged
Reviews are increasingly AI-generated too
So far we have been analyzing how AI-generated each paper submission is – but we can also measure how AI-generated each review is. We score each review’s written fields using the same Pangram methodology described above. Reviews are much shorter than papers – typically just a few chunks – so the scoring is coarser, but we can see that the overall trend is similar. In the 2024 edition, almost every review read as human-written; by 2026, half of all reviews had at least one flagged chunk, and 17% were entirely flagged.
Fully human0% flaggedMostly human1–49% flaggedMostly AI50–99% flaggedFully AI100% flagged
AI-generated papers receive higher scores from AI-generated reviews than from human-written reviews
Since we have detector scores for both papers and reviews, we can compare how human-written and AI-generated reviews score human-written and AI-generated papers. For this analysis, we include only papers and reviews that the detector classifies confidently, discarding the more ambiguous cases.4 When we look at heavily AI-generated papers, we find that AI-generated reviews rate them more favorably than human-written reviews do, with mean recommendations of 3.82 and 3.08, respectively.
For a slightly more controlled comparison, we do a paired analysis: we take the 24 heavily AI-generated papers that received at least one review confidently classified as AI-generated and one confidently classified as human-written, and compare the two review types within each paper. On average, reviews classified as AI-generated gave these papers recommendation scores 1.38 points higher, and scored them higher on all four component ratings.
Human-written reviewsAI-generated reviews
What do AI-generated papers look like?
We initially wanted to share concrete examples of the heavily AI-generated abstracts we received, so that readers could see exactly what high Pangram detection scores correspond to in this context. We contacted the authors of high-scoring abstracts to ask for permission, but nearly all declined to have their abstracts included, and so we do not include any real abstracts here.
Instead, we examined the highest- and lowest-scoring abstracts and asked Claude Opus 4.8 to look for recurring patterns that distinguish the two groups. Below are the observations we found most salient. (Note that the illustrative phrases in quotes are paraphrases, not quotations – they are intended to preserve the style/texture of the original phrases without identifying any paper.)
Patterns in heavily AI-generated abstracts
A barrage of statistics. Heavily AI-generated abstracts are often packed with precise numerical results, frequently citing statistical quantities such as
p-values
,confidence intervals
, andAUROC
scores. Human-written abstracts more often report one or two headline numerical results; in the AI-generated abstracts, the main result is often difficult to pick out amid all the surrounding numbers. The following might be an extreme, but not so unusual, example sentence:By epoch 6, mid-layer probe accuracy surpassed 79.4%, with an AUROC of 0.817 (Cohen’s \(d = 2.64\); Wilcoxon \(p < 10^{-10}\)).
- Marketing-flavored method names. AI-generated abstracts often give methods catchy, product-like names and introduce them as broad frameworks – something like
We introduce SpectraLens, a unified framework for…
. Human-written method names are more often plain or descriptive, including literal acronyms for what the method does. - Formulaic structure. AI-generated abstracts often follow the same template: broad motivation, an unresolved gap, a named method, a list of results, and a sweeping conclusion. They also overuse stock phrases such as
remains largely unexplained
,not merely X, but Y
, andTaken together, these results…
. Human-written abstracts are more varied in both structure and phrasing.
Musings on research in the age of AI
AI-assisted research is here to stay
It seems that we find ourselves in a world where AI-assisted research is here to stay – there is no going back to the dark ages of writing all our code ourselves. However romantic it may sound, returning to research without AI assistance is not realistic; as researchers and as a community, we should instead embrace the benefits of this new technology and adapt to it.
Viewed as a tool, AI can empower researchers to do more and higher-quality research. It can be useful throughout the research process – for spinning up experiments more quickly, visualizing data more easily, or finding bugs in mathematical arguments or code. It can also help with writing papers – for conducting broad literature reviews for related work, producing clearer and more beautiful figures, or editing text. Writing assistance seems especially useful for researchers for whom English is not their first language, helping them present their ideas more clearly to a research community that, for whatever reason, has largely settled on English as its shared language.
It is also starting to seem as though we need to prepare for a world where AI-generated research, even without a human in the loop, may be worth reading. An OpenAI model recently generated a counterexample to a longstanding Erdős conjecture that was subsequently verified and praised by mathematicians. Terence Tao recently said of other AI-generated proofs: I’m learning from some of the proofs that show up. I enjoy reading them.
Authors should take responsibility for their work
Given how useful and empowering AI assistance can be, we should find ways to use it responsibly. We believe one reasonable guiding principle is that authors should take responsibility for the work they attach their names to. If an author submits a manuscript, they should certainly know what is in it, and also be able to defend its contents.
This broadly aligns with arXiv’s policy on the issue:
“…by signing their name as an author of a paper, [the authors] each individually take full responsibility for all its contents, irrespective of how the contents were generated.
If generative AI language tools generate inappropriate language, plagiarized content, biased content, errors, mistakes, incorrect references, or misleading content, and that output is included in scientific works, it is the responsibility of the author(s).”
How peer review might adapt
The principle that authors should be responsible for the work bearing their names suggests a fairly simple adjustment to peer review that could incentivize authors to take more responsibility for what they submit. At most workshops and conferences, only accepted submissions are eventually deanonymized and made public. A natural alternative would be to deanonymize all submissions, including rejected ones, after double-blind peer review is complete.5 If authors know their names will appear beside their submission on the public internet for all to see, they may feel a greater responsibility for its contents.
In addition to deanonymizing submissions and making reviews public, one could imagine also publishing the results of automated checks – for example, AI-text detection scores or flags for potentially hallucinated references. Over time, researchers who repeatedly submit large volumes of low-quality work would build a public reputation for doing so.
Other approaches being tried include desk-rejecting submissions that violate a venue’s policy on AI-generated writing or contain hallucinated citations (NeurIPS Position Paper Track; ACL), temporarily banning authors who submit unchecked LLM output (arXiv), and limiting the number of papers each author can submit (TMLR).
Acknowledgements
We thank Max Spero, CEO of Pangram, for generously providing the API credits that made this analysis possible. We also thank Tim R. Davidson, Sheridan Feucht, David Bau, and Neel Nanda for helpful discussion and feedback.
Citation information
Please cite this work as:
Andy Arditi, Iván Arcuschin, and Andrew Lee. “An analysis of AI-generated content at the Mechanistic Interpretability Workshop.” 2026. https://www.andyrdt.com/posts/mi-workshop-ai-generated-content.
BibTeX citation
@misc{arditi2026aigenerated,
author = {Arditi, Andy and Arcuschin, Iván and Lee, Andrew},
title = {An analysis of {AI}-generated content at the {Mechanistic Interpretability Workshop}},
year = {2026},
url = {https://www.andyrdt.com/posts/mi-workshop-ai-generated-content}
}
Footnotes
Before the reviewing period began, we ran two automated screens across all submissions. The first flagged abstracts that Pangram scored as heavily AI-generated; the second checked every paper’s bibliography for fabricated references. These two screens yielded candidates for human review – one of us reviewed every flagged paper, and we only desk-rejected papers assessed to have an incomprehensible abstract or egregious citation fabrications. From these screens, 59 of the 801 submissions were desk-rejected before reaching the review process. ↩
Historical conference counts and the raw ICLR 2026 count are from Paper Copilot’s ICML, ICLR, and NeurIPS series; newer points use the NeurIPS program-chairs report (2025) and an official ICML announcement (2026). ↩
We use PyMuPDF and custom post-processing code for text extraction and cleaning. ↩
A review is classified as human-written only if every scored chunk is \(\le 0.25\), and AI-generated only if every chunk is \(\ge 0.75\). A paper is classified as human-written if at least 75% of body chunks score \(\le 0.25\), and AI-generated if at least 75% score \(\ge 0.75\). All other reviews and papers are excluded from this analysis. ↩
Note that ICLR already does this. The 2026 Author Guide specifies this explicitly: “All submitted papers (accepted, rejected or withdrawn) will be deanonymized after the notification. The submissions and reviews will be released to the public.” ↩