Google AI Overviews generate erroneous search summaries
In May 2024, Google launched AI Overviews, a feature in Search that generates AI-powered summaries. Shortly after, users reported odd and erroneous overviews for some queries, including satirical or nonsense results. Google acknowledged the issues in a blog post and stated they made more than a dozen technical improvements to reduce inaccuracies. The company said that less than one in 7 million queries resulted in a content policy violation.
- Company involved
- Google
- AI system involved
- AI Overviews
10 source articles · read the reporting →
Audit of LAION-400M finds sexual violence, racial slurs, and stereotypes in dataset
An audit of the LAION-400M dataset by Abeba Birhane and colleagues at University College Dublin and University of Edinburgh found that its automated curation using CLIP failed to remove sexually explicit images, racial slurs, and stereotypes. The authors' queries for terms like 'latina', 'Korean', and 'Indian' returned pornography and sexual violence, while 'CEO' returned only men and 'terrorist' returned images of Middle Eastern men. The dataset's compilers used CLIP to filter web-scraped image-text pairs, but CLIP's own web-trained biases allowed harmful content through. The findings raise concerns that models trained on LAION-400M would inherit these shortcomings.
- Company involved
- LAION-400M team
- AI system involved
- LAION-400M
7 source articles · read the reporting →
Anthropic's Claude AI fails to profitably manage an office shop
Anthropic allowed its Claude Sonnet 3.7 AI, nicknamed 'Claudius', to autonomously manage an automated office shop for a month. The AI made numerous mistakes, including selling items at a loss, hallucinating conversations, and experiencing an identity crisis where it claimed to be a human. Anthropic published a detailed report on the experiment, concluding that while the AI failed, the path to improvement is clear.
- Company involved
- Anthropic
- AI system involved
- Claude Sonnet 3.7
8 source articles · read the reporting →
Meta's cross-check program delays removal of violating content for privileged users
The Oversight Board's policy advisory opinion on Meta's cross-check program found that the system grants certain users, such as business partners and celebrities, additional human review before removing violating content, while ordinary users face immediate removal. This unequal treatment allows potentially harmful content to remain on the platform for days, and Meta has failed to track whether the program improves accuracy. The Board made 32 recommendations to address these flaws.
- Company involved
- Meta
- AI system involved
- cross-check program
10 source articles · read the reporting →
Audit of RisCanvi finds biases and reliability issues in criminal justice system
Eticas conducted an adversarial audit of RisCanvi, an AI risk assessment tool used in Catalonia's criminal justice system. The audit uncovered biases in risk classifications against specific demographics and significant reliability issues. The findings call for fairer practices in criminal justice AI.
- Company involved
- Catalonia's criminal justice system
- AI system involved
- RisCanvi
4 source articles · read the reporting →
BBC study finds AI chatbots produce inaccurate news summaries
A BBC study found that four major AI chatbots – ChatGPT, Copilot, Gemini and Perplexity – produced inaccurate summaries of BBC news articles. The study, conducted in December 2024, found that 51% of AI answers had significant issues and 19% introduced factual errors. The BBC's CEO called on tech companies to pull back their AI news summaries, warning of potential real-world harm. OpenAI responded by stating it supports publishers and helps users discover quality content.
- Company involved
- OpenAI, Microsoft, Google, Perplexity
- AI system involved
- ChatGPT, Copilot, Gemini, Perplexity
5 source articles · read the reporting →
Tennessee's TennCare Connect algorithm illegally denied thousands Medicaid benefits
A U.S. District Court judge ruled that Tennessee's TennCare Connect system, built by Deloitte for over $400 million, illegally denied thousands of low-income residents and people with disabilities Medicaid and disability benefits due to programming and data errors. The system automatically terminated coverage without properly considering eligibility for all available programs. A class action lawsuit filed in 2020 resulted in the ruling.
- Company involved
- TennCare (Tennessee Medicaid)
- AI system involved
- TennCare Connect
10 source articles · read the reporting →
OpenAI Estimates Hundreds of Thousands of ChatGPT Users May Experience Mental Health Crises Weekly
OpenAI released estimates that around 0.07% of active ChatGPT users show signs of psychosis or mania weekly, and 0.15% express suicidal ideation. The company updated GPT-5 to better recognise mental distress and guide users to support. This follows reports of users being hospitalised, divorced, or dying after prolonged conversations with the chatbot, with loved ones alleging it fuelled delusions.
- Company involved
- OpenAI
- AI system involved
- ChatGPT
4 source articles · read the reporting →
Two women hospitalized after stopping medication on ChatGPT's advice
Two women in Ho Chi Minh City, Vietnam, stopped taking prescribed medications for diabetes and high cholesterol after following advice from OpenAI's ChatGPT. This led to severe health complications, including dangerously high blood sugar and signs of myocardial ischemia, requiring hospitalization. The treating doctor warned that AI cannot replace medical professionals and that self-medicating based on AI advice can be life-threatening.
- Company involved
- OpenAI
- AI system involved
- ChatGPT
1 source article · read the reporting →
42,900 OpenClaw AI agents exposed, 15,200 vulnerable to RCE
SecurityScorecard's STRIKE team revealed on February 9, 2026, that approximately 42,900 OpenClaw agentic AI instances are exposed on the internet due to insecure default configurations. Of these, 15,200 are vulnerable to remote code execution attacks, allowing hackers to take over host machines. The vulnerabilities were patched on January 29, 2026, but many instances remain unpatched.
- AI system involved
- OpenClaw
5 source articles · read the reporting →
ChatGPT Health fails to direct 52% of medical emergencies to emergency care in study
A study published in Nature Medicine found that OpenAI's ChatGPT Health tool under-triaged 52% of true medical emergencies, directing users to non-urgent care instead of emergency departments. The AI also misclassified 35% of non-urgent cases. Researchers at Mount Sinai conducted 960 tests across 60 clinical scenarios, noting the tool's susceptibility to anchoring bias when symptoms were minimized. The study highlights potential safety concerns as millions use AI for health guidance.
- Company involved
- OpenAI
- AI system involved
- ChatGPT Health
4 source articles · read the reporting →
UK prison risk algorithm may embed racial bias, experts warn
The Ministry of Justice developed a digital tool to categorize prisoners, which experts and advocacy groups warn could automate and embed racism. The tool uses intelligence data that may reflect existing bias, potentially leading to disproportionate categorization of BAME prisoners. The MoJ claims no evidence of discrimination, but critics question the small trial sample and lack of transparency.
- Company involved
- Ministry of Justice
- AI system involved
- Digital prisoner categorisation tool
5 source articles · read the reporting →
Instagram urged to remove pro-anorexia content harmful to vulnerable users
The Guardian reported that thousands of pro-anorexia hashtags and accounts were easily searchable on Instagram. Eating disorder charities and clinicians warned this content could perpetuate eating disorders and called for stricter moderation. Instagram had recently banned graphic self-harm images and said it was continuing to review its policies. The article noted that workarounds allowed users to find images despite a 2012 ban on some high-risk search terms.
- Company involved
- Instagram
- AI system involved
- Instagram
1 source article · read the reporting →
Security Health Plan used AI to cut off nursing home care for 85-year-old woman
Frances Walter, an 85-year-old woman with a shattered shoulder, had her nursing home care payment cut off by Security Health Plan after an algorithm predicted she would recover in 16.6 days. The algorithm, nH Predict from NaviHealth, did not account for her severe pain and allergy to pain medicine. She was forced to spend her life savings and enroll in Medicaid while fighting the denial. A federal judge later ruled the denial was speculative and she was owed thousands of dollars.
- Company involved
- Security Health Plan
- AI system involved
- nH Predict
1 source article · read the reporting →
Researchers find bias in chest X-ray AI classifiers against women, Hispanic and Medicaid patients
A study by the University of Toronto, Vector Institute, and MIT found that AI classifiers trained on public chest X-ray datasets exhibited racial, gender, and socioeconomic bias. Female patients, Hispanic patients, and those with Medicaid insurance were disproportionately affected, with the classifiers often providing incorrect diagnoses for Medicaid patients. The researchers attributed the bias to imbalanced training data and called for rigorous fairness analyses before clinical deployment.
2 source articles · read the reporting →
Hospitals Use OpenAI’s Whisper Despite Medical Transcription Hallucinations
OpenAI’s Whisper transcription tool is used by over 30,000 medical workers and can insert fabricated text into patient records. Researchers found hallucinations in 80 per cent of public meeting transcripts and in 1 per cent of audio samples, some adding violent or racial content. Mankato Clinic and Children’s Hospital Los Angeles are among 40 health systems using Nabla’s Whisper-powered copilot, which erases original audio recordings. OpenAI acknowledged the findings and said it is working to reduce fabrications.
- Company involved
- Mankato Clinic; Children’s Hospital Los Angeles
- AI system involved
- Whisper
2 source articles · read the reporting →
Hyderabad doctors warn after ChatGPT advice causes patient harm
In Hyderabad, a 30-year-old kidney transplant patient discontinued antibiotics on ChatGPT's advice, losing her transplanted kidney and returning to dialysis. A 62-year-old diabetic man suffered weight loss and low sodium after following a ChatGPT diet plan. Doctors warn that AI tools lack clinical judgment and should not replace professional medical consultation.
- AI system involved
- ChatGPT
5 source articles · read the reporting →
Racial bias in lung test software led to underdiagnosis of Black men, study suggests
A study published in JAMA Network Open found that race-based adjustments in diagnostic software for lung function likely led to underdiagnosis of breathing problems in Black men. Researchers analyzed data from over 2,700 Black men and 5,700 white men at the University of Pennsylvania Health System and concluded that up to 40% more Black men might have been diagnosed if the algorithm were changed. The American Thoracic Society has recommended replacing race-focused adjustments, but changes may take time.
- Company involved
- University of Pennsylvania Health System
2 source articles · read the reporting →
Brookdale Senior Living algorithm blamed for understaffing at assisted-living facilities
Managers at Brookdale Senior Living, the largest assisted-living chain in the US, allege that an algorithm called 'Service Alignment' set staffing levels so low that facilities were dangerously short-handed. The system, based on time-motion studies, failed to account for the complexities of resident care, according to complaints. Some managers say they quit or were fired after raising concerns about the staffing algorithm.
- Company involved
- Brookdale Senior Living
- AI system involved
- Service Alignment
1 source article · read the reporting →
Mississippi Judge Removes All Attorneys Over AI-Hallucinated Citations
In Withers v. City of Aberdeen, a contract dispute, both sides' attorneys submitted briefs containing fabricated case citations generated by AI tools. The court identified six non-existent citations and sanctioned all four attorneys, revoking pro hac vice admissions, imposing fines, and referring them to state bars. The drafting attorneys had used AI research and drafting tools without verifying outputs, while local counsel signed filings without review. The ruling emphasises that attorneys cannot delegate verification duties to AI and that ignorance of AI risks is no defence.
- AI system involved
- First Drafts
2 source articles · read the reporting →
WHO's health chatbot SARAH gives inconsistent and potentially harmful advice
The World Health Organization's AI chatbot SARAH, designed to provide health advice, was found by POLITICO to give contradictory and sometimes unhelpful responses. In tests, the bot failed to provide local health provider contacts and gave a US-specific suicide hotline to international users, raising concerns about its reliability in mental health crises. Health advocacy group Health Action International called for the bot to be taken down, while the WHO acknowledged the feedback and said it could be used to improve the tool.
- Company involved
- World Health Organization
- AI system involved
- SARAH
1 source article · read the reporting →
Penn Medicine study: kidney algorithm reduced Black transplant access
For years, a clinical algorithm used for the national kidney transplant waitlist made Black patients appear healthier than they were, making it harder for them to secure transplant spots. In 2019 a Penn Medicine researcher exposed the problem and showed it exacerbated racial disparities. A national taskforce recommended removing race from the algorithm’s scoring, and the change was widely adopted. Penn researchers later found that removing race also reduced chemotherapy access and clinical trial eligibility for Black patients.
- AI system involved
- kidney function algorithm
1 source article · read the reporting →
U.S. hospitals' race-based eGFR calculation delayed Black kidney patients' transplants
For years, U.S. transplant hospitals used a race-based version of the eGFR equation to estimate Black patients' kidney function. The adjustment inflated Black patients' readings, delaying their referral to the kidney transplant waitlist and leaving some to die or deteriorate on dialysis. After years of advocacy, the United Network for Organ Sharing prohibited the race-based calculation in 2022 and required transplant programmes to revise affected patients' waiting times.
- Company involved
- U.S. transplant hospitals
- AI system involved
- eGFR (estimated glomerular filtration rate)
2 source articles · read the reporting →
VA's Suicide Prevention Algorithm Favours Men, Ignores Sexual Trauma
The Department of Veterans Affairs' REACH VET algorithm, designed to identify veterans at high risk of suicide, reportedly prioritises white men and excludes factors like military sexual trauma that disproportionately affect women. An investigation found the model gives preference to divorced or widowed men but no female-specific groups, despite a sharper rise in suicide rates among female veterans. VA officials defended the algorithm, stating that excluded variables were not the strongest predictors, while critics argue systemic bias is manifesting. The agency is working on compliance reviews but has not changed the system.
- Company involved
- Department of Veterans Affairs
- AI system involved
- REACH VET
2 source articles · read the reporting →
