NBER working paper puts AI-generated text at 29.4% of early-2026 US STEM dissertation abstracts
NBER researchers using the Pangram detector flagged AI-generated text in 29.4% of US STEM dissertation abstracts filed from January to May 2026. The unreviewed paper also finds that graduates with flagged abstracts who did their undergraduate degrees in the US were 17% less likely to work in academia.
Reality
- Evidence50
- Adoption30
- Hype gap+15
- Incentives
- Not scored
- Confidence50
Caine Prize reviews its winner after an AI detector scored the story almost entirely machine-written
Caine Prize is reviewing Simbarashe Kundizeza's winning story after The Atlantic reported that Pangram software scored it as almost completely AI-generated. Any ruling would rest on a tool that, by the critic's own account, has little independent testing behind it.
Reality
- Evidence30
- Adoption
- Not scored
- Hype gap+35
- Incentives
- Not scored
- Confidence40
buildOne report1 publisher Same-model rewording costs SlopShape's structural AI detector 0.9 points of F1
Sitefire's SlopShape spots AI-written company blog posts at 97.0% macro-F1 from page structure alone, and at 96.1% after each model rewords its own output. Paraphrasing generated copy hides little from it, though the test reworded text without reordering any page's sections.
Reality
- Evidence40
- Adoption
- Not scored
- Hype gap+25
- Incentives45
- Confidence35
Graphite counted 13,000 phrases that AI models use at least twice as often as human writers, with a different set for every model version. Editors cleaning AI drafts need a phrase list tied to the model that wrote them, rebuilt at each release.
Reality
- Evidence45
- Adoption
- Not scored
- Hype gap+25
- Incentives55
- Confidence50
Superhuman's purchase of AI detector GPTZero returned at least 10 times the $13.5 million it raised, a source told Fortune. Treated as a price, that puts a floor of $135 million under a deal with undisclosed terms.
Reality
- Evidence50
- Adoption65
- Hype gap+20
- Incentives55
- Confidence55
buildOne report1 publisher Wild AI text, about 31% of web tokens, gets a harm term in a new scaling law
Pangram Labs and UMass Amherst researchers put AI-generated text at about 31% of August 2026 web tokens and fit a scaling law for when it starts to hurt. For teams that pretrain on web crawls, AI text becomes a measured line item in the data budget, with a curve for its marginal value.
Reality
- Evidence35
- Adoption
- Not scored
- Hype gap+40
- Incentives45
- Confidence40
Dartmouth's inquiry into its own provost tests whether campus AI rules reach the top
Dartmouth College is investigating AI use in provost Santiago Schnell's writing after a detector scored his essay on AI cheating 100% AI-written. In his apology he names disclosure as the rule he broke, and a college can apply that rule to its provost and its students alike.
Reality
- Evidence45
- Adoption
- Not scored
- Hype gap+30
- Incentives
- Not scored
- Confidence40
France's Académie Goncourt dropped Thélyson Orélien's novel from its first selection, citing plagiarism and analyses that AI most likely wrote most of it. For anyone who sells AI checks, the case shows a likelihood score counting once a plagiarism record sat beside it.
Perspective Coverage
3 publishers
- Builder
- Builder 35%
- Operator
- Operator 48%
- Investor
- Investor 17%
Reality
- Evidence58
- Adoption
- Not scored
- Hype gap+20
- Incentives55
- Confidence60
AI detectors are flagging executives' posts as AI-written, and not everyone agrees with the verdict
LinkedIn asks users to flag suspected AI posts, Substack labels them and Anthropic watermarks Claude's text, while detector Pangram claims 99.7% accuracy. Readers can now check anything an executive signs. That makes a written rule on AI drafting necessary.
Reality
- Evidence35
- Adoption40
- Hype gap+30
- Incentives60
- Confidence40
Anthropic says it is watermarking Claude text to satisfy the EU AI Act Transparency Code and plans a detection API. The exposure sits with teams shipping generated work as their own.
Perspective Coverage
7 publishers
- Builder
- Builder 37%
- Operator
- Operator 44%
- Investor
- Investor 19%
Reality
- Evidence48
- Adoption45
- Hype gap+20
- Incentives62
- Confidence55
buildConfirmed2 publishers A detector run over 490,000 Common Crawl pages puts a number on synthetic text in the crawlable web, and the number is uneven enough to change how you build a training set.
Reality
- Evidence62
- Adoption40
- Hype gap+15
- Incentives35
- Confidence58
Researchers counted at least 40,000 people in Nairobi writing other people's coursework early this decade. The trade is close to gone, and the work that still pays is rewriting model output by hand until the software cannot tell.
Perspective Coverage
3 publishers
- Builder
- Builder 20%
- Operator
- Operator 50%
- Investor
- Investor 30%
Reality
- Evidence55
- Adoption40
- Hype gap+25
- Incentives30
- Confidence60
buildOne report1 publisher Which prompt built the test set decides Pangram's AI-assisted false-positive rate
Pangram's own 4.0 testing put the false-positive rate on AI-assisted documents at 0.01 percent, 4 percent or 7 percent depending on how the test text was produced, and the website shows the smallest of the three.
Reality
- Evidence48
- Adoption30
- Hype gap+20
- Incentives60
- Confidence42
Reviewers policing OpenAI's training contractors are barred from using AI detection tools
404 Media found contractors who read real ChatGPT conversations have been fired for handing in model-generated work, and the internal documents show the sanctioned check on them is a human judging the overall pattern.
Reality
- Evidence64
- Adoption58
- Hype gap+12
- Incentives62
- Confidence66
The words that flag a chatbot are turning up in ordinary human writing, Northeastern researchers say
Three Northeastern researchers describe chatbot vocabulary and sentence habits moving into human prose. The measurement behind that description is a preprint on how people and ChatGPT copy each other mid-conversation.
Reality
- Evidence38
- Adoption
- Not scored
- Hype gap+15
- Incentives40
- Confidence45
Self-published genre fiction on Amazon added titles twice as fast as revenue
Two detection studies put machine-written text across large slices of Amazon's self-published shelf, and the sales figures underneath one of them show a market that grew in dollars while thinning out per title.
Reality
- Evidence45
- Adoption55
- Hype gap+20
- Incentives60
- Confidence45
Substack hands its readers a button that scans your post for AI text
Fast Company reports that Pangram beats rival detectors in independent testing, without naming an error rate, and Substack, NewsGuard and an exam platform are already running it on writing whose authors never opted in.
Reality
- Evidence42
- Adoption58
- Hype gap+32
- Incentives66
- Confidence48
buildOne report1 publisher StealthGPT claims 89% Pangram bypass rate in own benchmark, separately reports 18% factual-consistency failure rate
Super exists because Pangram V4 stopped working against StealthGPT's old rewriter, and the launch report puts the evasion number next to a factual-consistency score of 82 percent on the same 100 samples.
Reality
- Evidence32
- Adoption27
- Hype gap+41
- Incentives86
- Confidence57
Custom detectors find AI-assisted text in almost one in ten parliamentary motions
Chalmers researchers trained Swedish and English detectors on 2,800 pre-chatbot motions and then watched flagged texts climb from near zero to almost one in ten by 2025-2026, with no member declaring the help.
Reality
- Evidence45
- Adoption30
- Hype gap+20
- Incentives60
- Confidence40
AI detectors hand content teams a percentage where they wanted a decision
Pangram scored a Wall Street Journal guest column as entirely machine-written, the author cheerfully agreed, and the paper answered with a test about the author's standing rather than the text. Semafor puts the flagged share of op-eds at about 3%.
Reality
- Evidence38
- Adoption45
- Hype gap+22
- Incentives55
- Confidence46