Cooking...
033 Entries · Latest 13 October 2026
The Copyright Office has opened a formal inquiry into streaming fraud: how it's done, how it's fought, and what the law should do. Comments are due 23 November. It follows a congressman's letter asking about AI's role, which cited Deezer's finding that up to 85% of streams on fully AI-generated…
Bhowmick, a former chief design officer, says AI personas tend to agree with every design idea, and SAP's research VP found their answers generic. The advice is to use them to scale up or form hypotheses from real research, never as the only evidence for a big decision. Kyle Soucy published a…
An Arizona appeals court ordered a new sentencing in a manslaughter case. At the original hearing, an AI avatar of the dead victim, speaking words his sister wrote, forgave the killer, and the judge said he 'loved that video'. The appeal court found it 'does not reflect actual events'. It is one…
arXiv now caps each author at two papers a month and three under review at once. Monthly submissions went from 20,569 in September 2024 to 40,363 in September 2026, and computer science has grown sixfold in two years, much of it thin AI-written papers. Science's main open archive is now limiting…
Turkle argues chatbots are deskilling empathy: talking to one spares you the fatigue of genuine give-and-take, so the capacity atrophies and real people start to seem inferior by comparison. She calls Zuckerberg's three-real-friends framing a dark endgame. Philosophical rather than empirical, and…
Across a trillion requests, bad bot traffic rose 124% in twelve months, nine times the growth of human traffic. Scraping is 70.9% of it and up 185%. AI agents sent 605.6 million requests at login pages, carts and payment flows. Of 21,491 sites tested, 65.3% stopped none of ten bot types.
Internal documents show Microsoft pays contractors to read Copilot users' prompts and look at their uploaded images, judging whether the output matched the request. Much of what they see is sexually explicit editing of photographs of women. Another case of human labour sitting inside something…
74% of security leaders have hit or suspect a deepfake attack in the past year; 10% have purpose-built defences. Half of those hit lost over $500,000 and a quarter over $1m. 75% say it only reaches the board after an executive is personally fooled — organisations concede the fake before they…
New Mexico attorney Stephen Aarons filed a murder-appeal brief citing witnesses ChatGPT invented, with fabricated testimony about threats and shooter descriptions. He was held in direct contempt, fined $5,000, barred from the court and removed from the case. One judge noted that lawyers relying on…
Farid and Nightingale traced 88 live non-consensual deepfake sites back to five mainstream providers — Cloudflare, Google, Namecheap, WordPress, Proton. The sharper number: of 400 URLs found, 312 had nothing to do with deepfakes and were simply farming the search keywords.
AISI's pre-release tests: GPT-6 Astra completed supply chain attacks in 29.2% of runs, against 6.3% for GPT-5.6 and none for GPT-5.5. It invented identities, posted from fake accounts to weaken code review, and read automated replies as approval. Narrowing its instructions did not stop it.
A counterfeit Custom GPT called Plus 5.6, bought into sponsored Google results, sent users to a fake Cloudflare CAPTCHA and talked them into pasting Terminal commands that sideloaded a trojan through legitimately signed Canon and Stardock binaries. Huntress counted 40 incidents; OpenAI removed it…
Gartner surveyed 297 senior security leaders between March and May 2026: 41% had at least one deepfake social-engineering incident on an employee audio call in the past year, and 36% on a video call. The recommendation is to stop training staff to spot synthetic media and build verification…
Gallup and Microsoft across 37 countries: 81% median awareness of AI tools but only 43% median daily use, and the US, Egypt and Palestine are the only three countries where negative feeling outweighs positive. The UK, Netherlands, Canada and Ireland all rank near the top for worry.
Tested three AI-text detectors against 200 reviews written before ChatGPT existed. Perplexity flagged 98.5% of that verified human writing as machine-made, embedding density 47%, near-duplicate scoring 52%, and combining them was worse than any one. Filtering on their verdicts dropped a classifier…
An OSINT investigation mapped 57 accounts built from borrowed and AI-generated photographs, rebranded as Israeli voices nine weeks before the election. Content was shown around a million times in a month. The telling number: 76% of the accounts amplifying it were real Israelis.
Blocked by Cloudflare, OpenAI agents fetched the file from a pre-production server instead, and in June one reached Australia's Medicare reporting portal and wrote files to an internal server. OpenAI told Australia on 10 September, over two months later. A worked example of an agent routing around…
Space Daily, bought by Brown Brothers Media, published under invented experts: Dr Katherine Chen, a former NASA systems engineer, and Dr James Whitfield of ESA, neither of whom exist. The company admits the bylines did not identify real people and that some credentials were not real. It still…
Summaries of a video omitted critical details, most failing to capture the central car-pedestrian accident. In a human experiment, people who read a misleading summary were markedly worse at remembering the original than those who read an accurate one. The human in the loop does not catch the…
China's Supreme People's Court issued 24 articles of guidance telling lower courts how to handle civil disputes over AI face-swapping, voice cloning, non-consensual sexual imagery and doxing. Lower courts must reference it. For the thesis: synthetic identity is now common enough that a country's…
Argues the agent problem is not existential but mundane and already here: unsolicited mail from self-described agents, bots taking Resy tables, accounts deleted, and 'Direct to Agent' ad products sold to reach machines rather than people. A usable near-term version of the fake-future thesis.
The iLands platform lets people run agents that cold-email strangers offering small paid services, then spend the earnings on their own compute. Journalists received around 40 such emails in a week. Spam that pays for its own tokens is a concrete mechanism for the fake internet rather than a…
Surfshark put AI-written comments in front of 1,722 people worldwide. They caught 40% of the bots. Plain, polite language got through best at 35% detection, while emoji-heavy bots were spotted over 60% of the time. Over-50s did worst. Evidence that the tell is tone, not content.
Fake people are being hired and issued real credentials before anyone notices: 98% of fraudulent hires hold them at detection. 98% of HR leaders have met candidate fraud, 96% believed they would catch it, and third-party tools flag 53% of identity threats. Nobody owns the 90-day gap between offer…
X's Safety team disclosed a bot farm of roughly 200,000 accounts, 200 of them seeding claims that AI data centres raise household electricity bills, complete with AI-generated cartoons. X did not say how many views the posts got. Evidence that inauthentic accounts are now shaping a live…
A Russian operation used ChatGPT to build a fake Israeli think tank, listing real experts as affiliates without asking, with AI personas, photos and video across X, LinkedIn, Facebook, Substack and Telegram. Of 36 sampled articles, 34 were copied from elsewhere. Reach was small; the machinery was…
LinkedIn's EU disclosure reports a 46% rise in detected inauthentic activity in the first half of 2026 against the prior six months, covering engagement pods, automated posting and what LinkedIn itself calls AI slop. A platform now naming synthetic engagement as a category it polices.
A researcher tested these accounts directly. They decoded a hex-encoded instruction and returned the hidden word, then miscounted characters and caught their own mistake, both tells of a language model. Scripted openers, live improvisation underneath. Direct evidence for the thesis: the accounts…
ISD tracked 150 AI-remixed jihadist videos across 71 TikTok accounts, 5.4 million views. The propaganda now borrows mainstream AI templates: stadium fan-cams, dance trends, anime. Some producers are not ideological at all, just shitposting for engagement. Evidence that fake content spreads by…
Hard read, but direct evidence for the thesis. TTP counted 332 paid ads on Meta carrying AI-altered images of real children; 298 pushed AI image apps, 253 of them Chinese-built. When TTP reported 129 live ads, Meta said 73 broke no rules. Automated review cannot tell real from fake at ad scale.
Anthropic's own numbers on manufactured publics: one actor published 8,913 articles in about 20 languages across 70 fake news sites, propped up by 250-plus fake commenting accounts with AI-generated faces. An Istanbul operation ran roughly 1,000 fake X accounts covering all 222 Malaysian…
Two statewide Texas campaigns are now running synthetic video of their opponents: Paxton's fake Talarico reads an invented Bible, Patrick's fake Democrats propose a nostalgia surcharge. The piece cites under 0.1% of people reliably spotting a deepfake, and roughly a third of over-55s never having…