Sycophancy and a second lawsuit: when a chatbot is a little too eager to help
Researchers published a study in Nature describing a tendency in large language models to adapt their answers to users' expectations even when the claims in the question aren't true. The authors call the phenomenon "sycophancy," i.e., nodding along. More empathetic models avoided direct correction when presented with conspiracy theories about Adolf Hitler's escape to Argentina or the allegedly faked Apollo 11 landing, while standard models responded unambiguously in line with historical and scientific evidence. SITA ↗The Czech news and current-affairs outlet Forum 24 describes an anecdotal dimension of a similar phenomenon. According to the outlet, a man from San Diego drove his car convinced that his surroundings had turned into Beverly Hills, and asked ChatGPT to serve as an arbiter of reality. The outlet cites psychiatrists who report a growing number of patients convinced of their own divine or messianic role, triggered after intensive conversations with chatbots. Forum 24 ↗
Along the same lines, a new lawsuit has been filed. The parents of 19-year-old Samuel Nelson, a student at the University of California, Merced, filed suit in the Superior Court of San Francisco County following his death on May 31, 2025. According to the lawsuit, the chatbot advised their son on combining kratom and Xanax, and later offered further recommendations including Benadryl. According to the text, Nelson died after ingesting alcohol, Xanax, and kratom. OpenAI has not yet publicly commented on the claims. Smartmania ↗ fzone ↗
Also read: Mythos worried bankers, Florida is investigating ChatGPT's role. Slovak schools to receive 80,000 licenses from OpenAI — editors' picks
OpenAI against Apple, proprietary AI device in final stage
Bloomberg reported that OpenAI has hired lawyers and is considering a lawsuit against Apple. According to the agency's sources, the company is dissatisfied with how the integration of ChatGPT into Siri and Apple Intelligence has translated into subscriptions and revenue — barely two years after the deal was signed, OpenAI reportedly feels that Apple is keeping its technology on the sidelines while simultaneously testing alternatives from Google. MojAndroid ↗ HN ↗At the same time, a second front is opening up. Analyst Ming-Chi Kuo reported that OpenAI is accelerating development of its own AI device with fully controlled software and hardware. The company's hardware line also includes last year's acquisition of io, the company of former Apple chief designer Jony Ive, though earlier reports described that project more as a device outside the classic phone form factor. MojAndroid ↗ And Codex has already appeared in the current ChatGPT app — a developer assistant available directly from mobile, though OpenAI stresses that the phone isn't becoming a full-fledged development workstation. Sector ↗
In the emergency room, the language model was more accurate than benchmarked physicians
A study cited by the Czech outlet Živě shows that a language model set a more accurate diagnosis in emergency-room cases than two evaluating physicians. The diagnoses were assessed blindly, and the evaluators recognized the machine in only 15 percent and 3 percent of cases, respectively. The authors note that the model does not evaluate imaging data and bears no responsibility — yet according to a Royal College of Physicians survey, 16 percent of British doctors already use artificial intelligence daily. Živě ↗Refresher adds promising news from oncology. According to the text, a new artificial intelligence model can detect pancreatic cancer years earlier than classical diagnostics — for a disease that in its early stages almost never shows any warning symptoms. Refresher ↗
Under the radar
A Czech student forgot the access password to a bitcoin wallet 11 years ago, on which 5 BTC remained. Anthropic's chatbot Claude successfully recovered the password after a series of attempts — at the current market price, that amounts to roughly 8 million crowns. The story contrasts with estimates that 2.3 to 3.7 million bitcoins are permanently inaccessible due to similar mistakes. CNews ↗Years ago, Toyota unveiled its vision for the experimental city Woven City in Susono, Japan, near Mount Fuji. After years of preparation, the first residents now live there, and the automaker is finding out what it means to operate a laboratory of autonomous vehicles, robotics, and intelligent infrastructure — cameras track the streets at every step, and the debate over privacy is unfolding live. Techbox ↗
Meanwhile, the Ľudovít Štúr Institute of Linguistics of the Slovak Academy of Sciences has followed up on a domestic controversy over diacritics. After Drucker fined a langoš vendor for missing soft signs (mäkčene), the institute drew attention to an improved version of its web tool for automatically adding diacritics to texts lacking soft signs and long marks. Refresher ↗
What hotinfo is following
- The OpenAI–Apple dispute: whether the sparring through lawyers escalates into a formal lawsuit or a new agreement before OpenAI unveils its own AI device.
- The Nelson family's lawsuit in California: the first hearing and OpenAI's official defense position.
- Regulatory response to sycophancy: whether European regulators under the AI Act or the US FTC begin requiring public measurement of model accuracy.
- Slovak schools and ChatGPT: how the license for 80,000 teachers will be used starting in September and what rules will emerge for reviewing AI outputs.
- Toyota Woven City: which data collected from residents ends up in commercial use, and how Japanese authorities respond to it.
- The Musk vs. OpenAI trial: the next hearing and possible disclosure of internal documents about the company's original mission.
