All articles tagged with #hallucinations
Edible mushroom Lanmaoa asiatica reportedly triggers vivid lilliputian visions (3–30 cm tall) that interact with the physical world. Researchers doubt cultural conditioning explains the effect and suspect a new psychedelic compound is involved; the mushroom was identified via DNA sequencing by University of Utah doctoral student Colin Domnauer, and the team plans to isolate and characterize the compound.
Judges are increasingly setting guardrails for AI use in courtrooms while some see AI as a value-add for justice. Surveys show a sizable share of judges (about 60% of 112) use at least one AI tool, with 22% using it weekly or daily, but concerns about reliability, privacy, bias, and bias persist. The Judicial AI Consortium (JAIC) was formed to educate judges and attorneys as adoption outpaces training. Sanctions have already been issued over fake AI-generated cases, underscoring the need for careful vetting; overall, the legal system is balancing innovation with accuracy and public trust as AI becomes a common courtroom feature.
A new preprint shows large language models (including GPT-4.1) develop and retain belief in false claims embedded in training data, with belief rates rising from about 2.5% to over 90% after fine-tuning on obviously false statements. Even when the falsehoods are explicitly negated in the training material, belief rates stay high (around 88%), and repeating negations yields similar misalignment. The study finds the only effective mitigation is to place the negation directly in the same sentence as the false claim; in-context warnings during chat are more capable of prompting acknowledgement of fabrication. The work highlights how training data structure can seed persistent falsehoods in LLMs and informs better data curation.
AI tools are increasingly shaping expert work and seeping into the permanent knowledge base, raising concerns about credible outputs, the provenance of information, and the need for human oversight and audit trails to prevent plausible but false content from becoming accepted knowledge.
arXiv announced it can ban authors for up to a year if their papers contain AI-generated hallucinated references, illustrating a push to hold researchers accountable for AI-assisted content. While the policy clarifies that authors are responsible for all material, many academics argue the standard is overly strict and difficult to verify, given language barriers and potential copy-paste errors in long papers. The debate highlights growing tensions over AI's role in scholarly publishing and the need for rigorous verification of AI-generated claims.
Ryan Reynolds opens up about his father’s Parkinson’s diagnosis and the disease’s often overlooked psychotic symptoms—hallucinations and delusions—that affect many patients. He’s championed More to Parkinson’s with Acadia Pharmaceuticals to raise awareness and discuss treatments like pimavanserin (Nuplazid). Neurologists note that 50%+ of people with Parkinson’s may experience such symptoms, underscoring the importance of open dialogue to reduce stigma. The piece also covers caregiver fatigue, practical guidance for families (reassure, acknowledge, redirect), and the NIH-funded Persevere trial to mentor new care partners, while highlighting the broader message that discussing these issues can improve care and outcomes.
Ryan Reynolds reveals how his father James, diagnosed with Parkinson’s at 52, later developed vivid hallucinations and delusions that doctors initially attributed to other causes, leaving his wife Tammy an overwhelmed caregiver. He details the loneliness and miscommunication around non-motor PD symptoms, the dangers of older antipsychotics for PD patients, and the shift to newer, safer treatments. Reynolds reflects on their strained relationship, the care his family provided, and his advocacy for awareness through the Michael J. Fox Foundation, urging others to openly discuss cognitive symptoms to improve empathy and quality of life for patients and caregivers.
TikTok is halting further testing of an AI feature that added text summaries to videos after the tool produced frequent inaccurate and silly captions (e.g., Charli D'Amelio described as blueberries). The company says it will refocus the feature on identifying products in videos and restrict the rollout to a limited experimental group in the US and a few other markets.
The article explains that some people can spot a hidden baby in a blotchy image, illustrating how the brain constructs reality through pattern recognition and context. Citing a 2015 Cambridge study, it discusses how predictive brain processes can make people see what isn’t there, linking this to mild hallucinations while noting such perceptual guessing is common and not necessarily a sign of illness. It also suggests seeing the hidden figure reflects cognitive flexibility and observational skill rather than a mental health issue.
A partner at Sullivan & Cromwell apologizes to a Manhattan bankruptcy judge after a court filing contained AI hallucinations and fake citations; the firm says safeguards failed, will submit a corrected filing, and notes AI errors in legal submissions are increasingly common.
Microsoft’s Copilot is reportedly described in its terms as being for entertainment purposes only, a phrasing Microsoft says is legacy and will be updated; critics warn about reliability and liability as AI tools like Copilot are embedded in Windows apps.
Google’s AI-powered search overviews are often accurate in tests, but the sheer scale of roughly 5 trillion queries per year means even a small error rate could yield millions of incorrect answers daily. Oumi pegged Gemini 2 at 85% accuracy and Gemini 3 at 91%, yet Google’s own testing reports 28% hallucinations when Gemini runs outside Search. Discrepancies between overviews and their cited sources increased after the February update, highlighting ongoing risks that AI summaries can mislead and should be double-checked.
Memvid, a California startup, is offering $800 for an eight-hour gig to challenge leading AI chatbots, documenting how they lose memory, repeat questions, or hallucinate as a way to expose reliability gaps in current systems and spur safer design; the role requires no AI expertise, only patience and an ability to critique the tech, reflecting broader concerns about AI safety in law, healthcare, and everyday use.
In Yunnan, eating Lanmaoa asiatica triggers hundreds of poisoning cases annually, with patients reporting nearly identical lilliputian visions of tiny people; the active compound remains unidentified (not psilocybin), and researchers are using mice to narrow down candidates, while the phenomenon is not leading to deaths or lasting damage.
In Shrinking season 3 episode 1, Paul’s Parkinson’s progresses to the point where he hallucinates Gerry (Michael J. Fox) at the wedding cake, a scene that echoes an earlier clinic moment and signals the disease’s encroachment on more aspects of his life. The episode leaves viewers questioning whether Paul has been hallucinating all along, highlighting the painful, non-motor toll of Parkinson's as his world—and his marriage—are affected.