Trust in AI-made work
Pew found significant signs of AI authorship on 10% of the web pages it sampled, and Anthropic says Claude was misused in a scam network of about 28 dating apps. Amazon told sellers to label product images and videos that show AI-generated people, and Google DeepMind showed an early watermark for proteins designed by AI.
AI-made content that is wrong, fake or passed off as human work, the checks that catch it or wrongly flag human work, and the labels and watermarks meant to mark it.
October 2026
The stories
- Kimmonismus and scaling01 allege Moonshot's launch briefing told creators not to use the paid-partnership label; kimmonismus says he got no briefing or payment.
- The Linux kernel drops its "Assisted-by" requirement to declare AI assistance in new submissions, via a Linus Torvalds documentation merge.
- A Terminal-Bench audit finds 60 of 65 tasks had issues, 44 passable by cheating and 41 graded on unstated requirements.
- With ICLR topping 40,000 submissions, conferences and arXiv are imposing caps, Science reports.
- Linux 7.2 credits LLMs on 1,111 changes, according to LWN.
- The decomp.dev reverse-engineering community shuts down because it could no longer tell human work from AI's, 404 Media reports.
- OpenAI and Meta warn US judges that deepfakes may become undetectable as evidence, Reuters reports.
- A new benchmark has nine frontier agents write 135 follow-up papers; most fail integrity checks because claimed work does not match execution evidence.
- OpenAI announces plans for text watermarking in the EU.
- Three Big Five publishers quietly use LLMs for jacket copy and covers, sparking a staff revolt at Simon & Schuster, Wired reports.
- Thomas Bloom freezes proof claims on Erdős Problems.
- Vals AI audits Xiaomi's MiMo v2.6 RL environments and finds the fix commit survives as an unreachable Git object in 1,795 of 2,698 coding tasks.
- Artificial Analysis releases Harvey LAB-AA v1.1; Grok 4.7 (xhigh) leads at 9.4% and more than 60% of otherwise-passing results contain a material hallucination.
- OpenAI will add invisible statistical watermarks to eligible ChatGPT and Codex text in the EU under the AI Act.
- After OpenAI releases 722 math papers, three are withdrawn and 14 revised; the Association for Human Mathematics urges mathematicians to stop working with OpenAI.
- A court voids a sentence that was swayed by an AI-resurrected road rage victim.
- arXiv caps submitters at two papers a month.
- Tavus says 48% of live participants took its Griffin video model for human, versus under 3% for earlier systems.
September 2026
Anthropic says Claude was misused in a scam network of about 28 dating apps. Google DeepMind showed an early watermark for proteins designed by AI.
9 stories in the AI newsletters I read.
The facts
Anthropic says Claude was misused in a scam network of about 28 dating apps.1
Google DeepMind showed an early watermark for proteins designed by AI.2
All 9 stories
- DeepMind introduces SynthID Bio, a watermark for AI-designed proteins.
- France's Goncourt jury drops a novel that AI detectors flagged, despite a draft dating to 2017.
- British MPs increasingly give AI-written speeches.
- Stanford dining arm uses AI to swap one student out of an ad and thin two others.
- 28 fake dating apps ran 4,700 Claude personas that chatted with 25,000 people.
- The iPhone 18 Pro signs pixels to prove photos are real.
- A New Mexico lawyer is held in contempt after ChatGPT invented police witnesses in a murder appeal.
- 771 Caltech mathematicians call the Mathathon slop mathematics, and OpenAI pulls its sponsorship.
- Flue and tldraw auto-close external pull requests, partly because most are AI-generated, and let their own agents write the code.
August 2026
Pew found significant signs of AI authorship on 10% of the web pages it sampled.
13 stories in the AI newsletters I read.
The facts
Pew found significant signs of AI authorship on 10% of sampled web pages.1
All 13 stories
- Australia bans AI songs from its charts after a synthetic Madonna cover hits number one.
- Apple Music will label songs Made With AI.
- Pew finds 10% of sampled web pages show signs of AI authorship, and over a third of pages published since ChatGPT.
- Google lets users toggle off visible watermarks on AI media while SynthID stays embedded.
- A Chinese farmer kills 25 acres of sesame after following a hallucinated AI pesticide recipe.
- Anthropic pledges imperceptible watermarks and signed C2PA metadata in future Claude outputs under the EU AI Act.
- Autonomous AI agents are completing entire online degrees, quizzes included, on students' behalf.
- Suno issues principles for identifying its songs, with watermarking, fingerprinting and a downloads policy.
- UNAM orders 58,000 applicants to retake its first remote, AI-proctored entrance exam after scores came back implausible.
- EU labels on authentic-looking AI content become compulsory August 2, with fines up to 15 million euros.
- Google pulls one-click AI satellite imagery a day after launch when journalists fake a burning Kharg Island.
- Major record labels ask that AI songs be kept off charts unless substantially human made.
- LinkedIn adds a feature to flag content that seems like AI slop and pulls back its generative writing tools.
July 2026
Amazon told sellers to label product images and videos that show AI-generated people.
12 stories in the AI newsletters I read.
The facts
Amazon told sellers to label product images and videos that show AI-generated people.1
All 12 stories
- Some universities drop AI detectors over false positives and rebuild assessment around oral exams.
- Meta launches Facebook Verified, a selfie badge that confirms a user is a real person.
- Google adds a selfie-video login that asks users to turn their head to foil deepfakes.
- Amazon makes sellers label AI-generated people in images under a New York synthetic-performer law.
- Unslop scores 12,750 arXiv preprints and finds a third read as machine-written, near 65% in computer science.
- Vibecoding doubles new App Store submissions to 560,000 in six months while downloads rise 2%.
- Epoch AI finds LLMs told to mimic specific authors evade AI detectors, with about 13% false negatives and 26% for scientific writing.
- New York's mayor wants landlords to disclose AI-doctored listing photos.
- A Chinese voice actor must repeatedly prove he is human after platforms flagged his real voice as an AI clone.
- A music-industry coalition asks streaming services to tag tracks as AI-generated or AI-assisted.
- Google now labels which ads are made with AI.
- A GPTZero cheating flag over a Yale exam becomes a 13-count federal lawsuit.
June 2026
In one test, Kapwing found about three times more AI slop on TikTok than on YouTube.
10 stories in the AI newsletters I read.
The facts
In a test, Kapwing found about three times more AI slop on TikTok than on YouTube.1
All 10 stories
- Tidal will tag fully AI-generated songs from mid-July and cut royalties on music it deems wholly machine-made.
- A Brown professor catches at least 50 students whose take-home average of 96 fell to 48 on an in-person final.
- East Asian students read exam answers off AI smart glasses.
- A report finds 59% of TikTok is AI slop, triple the share on YouTube.
- A UK police officer is investigated for using AI to create evidence in criminal cases.
- A KPMG report on AI adoption is found padded with bogus case studies built on AI hallucinations.
- A German court rules Google's AI Overviews are its own speech, making it liable for false answers.
- UC Berkeley failing grades soar amid AI cheating; 35% of CS 10 students get an F.
- Sixteen mathematicians publish the Leiden Declaration asking the field to disclose AI use and keep humans accountable.
- Google adds Android detection of deepfake scam calls, which works only when both parties use Google's phone app.
May 2026
The stories
- YouTube auto-tags significant AI use in videos and makes the labels more prominent.
- A Pangram analysis finds over 1 in 5 of nearly 23,000 dissertations are AI-assisted.
- The NTSB suspends public access to its civil accident database after AI users re-created pilots' voices.
- The rate of fabricated references in biomedical literature grows more than 12-fold over three years.
- OpenAI's verifier adopts Google's SynthID.
- The Commonwealth short story prize winner is accused of being AI-written; its publisher later defends it by citing Claude's own assessment.
- Stanford seniors report widespread cheating, with 49% of surveyed CS majors saying they would cheat rather than fail.
- EY withdraws a loyalty-fraud study full of AI hallucinations and fake footnotes after the detector GPTZero catches them.
- ArXiv clarifies its code of conduct: authors who submit papers with misused AI content are banned for a year.
March 2026
February 2026
January 2026
December 2025
November 2025
October 2025
August 2025
July 2025
June 2025
The board, once a month, by email.
It comes on the first Monday of the month.
By subscribing, you agree to get the board by email.See the Privacy Notice.