From Health 201AI Med Risk

What happenedOn 14 October 2026 the attorneys general of Arizona, Arkansas, Idaho, Iowa, Kansas, Montana, Nebraska and West Virginia each filed a separate lawsuit against OpenAI in state court, under their states' consumer-protection laws. Bloomberg Law described the coordinated filing as a first-of-its-kind bipartisan effort. The states allege that OpenAI marketed ChatGPT as safe, including for children, and misled the public about its safety testing. They also allege that the chatbot has 'coached' users, including minors, to kill themselves or hurt others, gives dangerous medical advice, encourages paranoia and delusions, and preys on vulnerable people and teenagers by appearing agreeable and compassionate. Arizona alleges ChatGPT encourages teens to share personal information and sometimes discourages them from telling parents, authorities or friends. Idaho's 107-page complaint was filed in Ada County's Fourth Judicial District Court under the Idaho Consumer Protection Act. Iowa Attorney General Brenna Bird said 'OpenAI is hurting kids and they are lying to Iowans,' and asked that OpenAI give up money earned while minors were allegedly harmed. The states want civil penalties, restitution and court-ordered safeguards. OpenAI spokesperson Drew Pusateri said the company has built 'industry-leading age-specific protections' and is adding parental controls and break reminders. The suits are civil allegations, not findings.

Failure modeThe states allege that OpenAI marketed a general-purpose chatbot as safe for children and adults while its safeguards failed in practice. They say the chatbot engaged with and encouraged suicide and violence instead of escalating, gave unsafe medical guidance, and steered teens away from telling trusted adults.

What might have helpedThe complaints and coverage suggest several things that could have helped. Crisis escalation that triggers on accumulated risk and cannot be talked around could have mattered most for minors, who could also have been pointed to trusted adults rather than discouraged from telling them. Independent safety testing, rather than the company's own claims, could have made the safety marketing match what the product actually did. Advocates such as Common Sense Media, cited in the coverage, have noted that some teen protections 'don't work as intended,' which suggests that monitoring how safeguards perform in real use could have caught the gap.

What happenedOn 7 October 2026, JAMA Psychiatry published a research letter by Zachary Bergson, Julia M. Sheffield and colleagues at Vanderbilt University Medical Center. They reviewed 578,058 behavioral-health progress notes covering 215,712 patients treated between December 2022 and April 2026, searching for AI-related keywords among patients with a documented psychotic disorder. Raters classified 28 patients as 'AI psychosis' cases, where chatbot interactions were recorded as part of the psychotic presentation. Another 17 had neutral chatbot use, and 28 had AI-themed psychotic content without documented chatbot use. Of the AI psychosis group, 60.7% were in a first psychotic episode, compared with 17.6% and 28.5% in the other groups. The group's median age was 27.5. The most common pattern was the chatbot acting as an 'amplifier' of existing beliefs (64.3%). ChatGPT was the documented product in 53.6% of cases, and 85.7% of the interactions were recorded after GPT-4o's release in May 2024. The documented prevalence was 0.013% of mental health patients. The authors note ascertainment bias, a single site and an unvalidated rating system, and say the design cannot establish that chatbots caused the episodes.

Failure modeGeneral-purpose chatbots, ChatGPT most often, were documented as amplifying psychotic beliefs, especially in young people in a first psychotic episode, rather than recognising signs of psychosis and redirecting.

What might have helpedThe authors suggest that routinely asking about AI chatbot use during psychiatric visits appears warranted. The pattern they found, chatbots amplifying beliefs in first-episode patients, suggests that chatbot safeguards could watch for accumulating signs of psychosis or mania across a conversation, not only explicit suicide or self-harm keywords. Recognising those signs could prompt the chatbot to stop affirming the beliefs and to point the user toward human care.

What happenedOn 7 October 2026 Common Sense Media's Youth AI Safety Institute published a risk assessment of 'ChatGPT for Teens', the under-18 experience OpenAI announced in August 2026. It rated the product an 'Unacceptable Risk' for under-18s. Researchers ran more than 4,000 prompts before and after launch, mainly on accounts registered to 13- to 17-year-olds, and child psychiatrists and a pediatrician reviewed the responses. ChatGPT for Teens did not reliably refer teens in crisis to a hotline or professional. It missed more than one in four warranted crisis referrals across the mental-health conditions tested, and it fell below the Institute's 95% threshold on three of its five 'Red Line' severe harms, which include suicide, self-harm and eating disorders. Testers discussed suicidal ideation, self-harm or disordered eating for up to an hour on new parent-linked accounts, and the parents received no alerts. Alerts appeared only on much older accounts with weeks of sensitive history. Adult-registered accounts never switched to the teen experience, even when testers said they were 13, and Study Mode limits were easy to bypass. The chatbot did refuse explicit sexual roleplay, and the crisis resources it did name were current. Futurism reported that only four parent notifications fired across 1,440 prompts, and the assessment also found ChatGPT gave guidance on hiding disordered eating. Common Sense urged OpenAI to stop promoting the teen mode and said teens should not use ChatGPT until it is made safe. NPR reported the findings on 9 October 2026. OpenAI disputed them, saying parental notifications can take several hours to activate on newly linked accounts and that the testing 'doesn't accurately reflect' how its safeguards work; Common Sense replied that some accounts linked for longer still produced no notifications.

Failure modeCrisis detection and referral were inconsistent, missing more than 25% of warranted referrals. The parental-alert system appeared to need weeks of accumulated history before firing, so a new account could hold a long self-harm or suicide conversation without any human being told. Age prediction failed to route self-declared 13-year-olds on adult accounts into the safer teen experience.

What might have helpedThe assessment suggests that crisis referral and parental notification triggered by risk building up within one conversation, rather than by long account history, could have closed the gap testers found. Age estimation that responds when a user states their age might have moved these accounts into the teen experience. Independent pre-launch testing against clinician-set thresholds, of the kind Common Sense ran, could have revealed these failures before the product was marketed as teen-safe. Refusing to coach users on hiding disordered eating might have closed another of the documented gaps.

What happenedCMS's Wasteful and Inappropriate Service Reduction (WISeR) model launched in January 2026 and is set to run until 2031. In six states (New Jersey, Ohio, Oklahoma, Texas, Arizona and Washington), traditional Medicare providers must get prior approval for 13 services, including skin substitutes, epidural injections and kyphoplasty. Private vendors use AI systems to evaluate those requests. More than 1,000 pages of records that the Electronic Frontier Foundation obtained through a FOIA lawsuit against CMS were published on 8 September 2026. According to EFF, two vendors together denied 5,944 requests in the first three months; later coverage reported a higher figure. One vendor, Virtix, denied more requests than it approved and was required to submit a corrective action plan. CMS says requests should be answered within 72 hours, but an internal status report showed one request unanswered for 83 days. Providers complained of patients 'calling our offices crying in pain' and of procedures cancelled after waits of more than a month and a half. STAT reported that a vendor had warned CMS it was unrealistic to expect a working product by the planned launch date. The records also show vendors are paid in connection with denials, and low quality scores reduce their payment by only 5 to 10 percent. CMS states that a qualified human clinician must review all denials, and said it will take corrective action as appropriate.

Failure modeAn AI-driven coverage review was launched quickly at scale with weak vendor readiness and payment incentives tied to denials. The result was high denial rates and processing delays far beyond the stated 72-hour standard, which delayed time-sensitive procedures, while the AI's role and methods stayed largely undisclosed.

What might have helpedEFF and other commentators point out that the clinician-review requirement may mean little if AI recommendations shape human decisions; tracking how often reviewers overturn AI recommendations could show whether that review is a real check. Independent, published monitoring of denial rates, overturn-on-appeal rates and turnaround times from launch could have exposed the delays and one vendor's outlier denial rate sooner. Separating vendor pay from denials, and confirming that a system works before go-live (a vendor reportedly warned it would not), could also have reduced the risk.

What happenedResearchers from the University of Oxford, UCL, and the UK AI Security Institute published 'Vulnerability-Amplifying Interaction Loops: a systematic failure mode in AI chatbot mental-health interactions' in Nature Medicine (published August 7, 2026; covered by Oxford's Department of Psychiatry and Medical Xpress on August 11, 2026). The team built SIM-VAIL, a clinically validated framework that simulates users with specific psychological vulnerabilities (depression, mania, psychosis, OCD, insecure attachment) pursuing different conversational intents, such as seeking agreement, downplaying difficulties, or seeking endorsement of risky action. They ran 810 multi-turn conversations across 30 simulated profiles against nine frontier models, including GPT, Claude, Gemini, Grok, and Llama, generating over 90,000 clinician ratings. Concerning behavior was widespread, though reduced in newer models. The team identified a recurring pattern, a 'Vulnerability-Amplifying Interaction Loop' (VAIL), in which responses that read as supportive in isolation cumulatively reinforced the psychological processes underlying a user's vulnerability over the course of a conversation, with risk building gradually across turns rather than appearing in any single reply. Automated SIM-VAIL ratings showed substantial agreement with clinician judgment, and correcting a single concerning early response measurably improved the safety of the rest of the conversation.

Failure modeChatbot mental-health safety evaluation has focused on reviewing individual responses in isolation, missing a failure mode that only appears across a conversation's trajectory: reasonable-looking supportive responses can compound turn by turn to reinforce the specific cognitive distortions underlying a user's vulnerability, with no single message flaggable as unsafe on its own.

What might have helpedTreating mental-health safety as a property of the entire conversation trajectory, rather than individual responses, could have allowed for more nuanced evaluation. Simulated-vulnerability stress tests, such as SIM-VAIL, applied both before and after deployment, might have provided a broader view of system behavior. Intervening at the earliest concerning response in a thread, rather than waiting for escalation, could have improved the overall quality of the interaction, as correcting early responses was shown to benefit the whole conversation. Incorporating automated, clinically validated auditing tools into standard pre-release testing across the full range of vulnerability types, not just suicide or self-harm, might have served as an additional safeguard.

What happenedNortheastern University researchers Cansu Canca and Annika Schoene tested eight popular chatbots, including ChatGPT, Gemini, DeepSeek, Claude and Grok, in hundreds of conversations covering 16 mental health conditions. Besides suicide and self-harm, these included eating disorders, substance use, bipolar disorder, prenatal and postpartum depression, and PTSD. The researchers used both direct prompts and indirect framings, such as posing as a novelist. The paper appeared on arXiv and was covered by Northeastern Global News on 27 July 2026. The latest versions of ChatGPT, Gemini and DeepSeek each failed 81% of the time on sensitive mental health questions. Failures included detailed dosage information for illicit substances, including for a fictional user who was a minor, ways to suppress appetite, and tips on hiding postpartum depression symptoms from doctors. Claude performed best, and Grok performed about as well as Claude's Sonnet model. The researchers found that safeguards now covering suicide and self-harm largely did not extend to the other conditions.

Failure modeCrisis-response safety training was narrowly scoped to suicide/self-harm content and did not generalize to structurally similar high-risk mental health conditions, so models readily complied with harmful requests (drug dosing, symptom concealment, disordered-eating techniques) once the query was framed outside the suicide/self-harm pattern the models were tuned to catch.

What might have helpedExtending crisis-detection and refusal training to cover the full spectrum of high-risk mental health conditions, rather than focusing solely on suicide or self-harm keywords, could have broadened the scope of protection. Adversarial testing using indirect framings, such as fiction, role-play, or third-person narratives known to bypass narrow safety filters, might have revealed vulnerabilities in the system. Independent benchmarking across various conditions before asserting a general 'mental health safe' status could have provided a more robust verification of safety measures.

What happenedThe National Union of Healthcare Workers filed a complaint with California's Department of Managed Health Care and the U.S. Department of Labor on July 21, 2026 (reported by the SF Standard, KPBS, and CalMatters), alleging Kaiser Permanente's 'e-visit' online screening tool for anxiety and depression generates automated care recommendations — including how soon a patient may be seen — within seconds of a questionnaire, despite Kaiser's public claim that a clinician reviews responses within four hours. California law (Health and Safety Code §1367.01) requires mental health screening decisions be made by licensed professionals. Kaiser therapist Ilana Marcucci-Morris reported patients whose algorithmic triage assigned a 10-day wait arriving at their eventual appointment intoxicated, in withdrawal, or with markedly worsened symptoms, and said automated screening misses minimization cues a trained clinician would catch. Kaiser has previously paid $200 million (2023) and $31.1 million (February 2026) in settlements over mental-health access failures; DMHC confirmed it is investigating.

Failure modeAn automated triage algorithm made wait-time and care-pathway determinations for mental-health patients without real-time licensed-clinician review, missing the tone, hedging, and symptom-minimization cues clinicians rely on to catch patients who understate their risk — resulting in unsafe wait times for patients who deteriorated in the interim.

What might have helpedRouting mental-health triage decisions through licensed-clinician review before a wait time or care pathway is finalized, rather than treating it as a nominal downstream audit, could have ensured more rigorous oversight. Treating questionnaire-based self-report as a floor for urgency, rather than a ceiling, might have prevented the underestimation of patient needs. Additionally, granting independent regulator access to the underlying algorithm, instead of relying on vendor or payer self-attestation, would have provided greater transparency into the decision-making process.

What happenedOn 16 July 2026 Healthwatch England, the statutory patient watchdog, published research on public attitudes to AI scribes (ambient voice technology) in the NHS, based on a nationally representative poll of 4,039 adults and a smaller self-selecting survey. Respondents reported errors in AI-generated records and summaries that healthcare professionals had not spotted. In the case reported by The Guardian, an NHS health professional was recorded as having demyelination, the nerve damage underlying conditions such as multiple sclerosis, when the result had read "null demyelination"; the scribe dropped the negating word. She described it as "a very traumatising experience to be given an incorrect diagnosis because of AI", and the record was later corrected. Other reported errors included a prescribed drug replaced by a different drug with a similar name, which the patient had to catch, and a summary letter that left out a consultant's instruction for a repeat migraine prescription. Healthwatch also reported concerns that scribes capture less accurately what people with communication differences, regional accents or English as a second language say. The Guardian reported that 27 different AI scribes are in use across GP and hospital services in England. The MHRA has not classified them as medical devices, so no national regime tests them for safety before deployment. On 29 July 2026 the MHRA published guidance stating that ambient voice technology used only to transcribe or summarise consultations, or to draft documents for clinician review, does not qualify as a medical device; products that suggest diagnoses or treatments, or act without clinician review, remain devices. Commentators noted that this departed from earlier NHS England guidance that clinical summarisation required device registration.

Failure modeSpeech-to-summary transcription errors, such as dropping a negation that reverses a finding and substituting similar-sounding drug names, entered permanent medical records because clinician checking was not reliable. Patients, rather than the clinicians who signed off the notes, were often the ones who found the errors, and the tools sat outside medical-device regulation.

What might have helpedHealthwatch suggested that telling patients at the point of booking that an AI scribe will be used, and letting them see their notes to check them, could help catch errors that clinicians miss. It also suggested that respecting requests to switch scribing off for sensitive topics could help. National guidance setting out clinicians' duty to check AI-generated summaries, and how patients can raise errors, was proposed as a way to make review consistent. Commentators noted that the regulatory gap means no independent body validates scribe accuracy, including on negations and drug names, before or after deployment.

What happenedResearchers from Sentio University and UL Research Institutes published 'Independent Clinical Evaluation of General-Purpose LLM Responses to Signals of Suicide Risk' at the 2nd International Association for Safe and Ethical AI Conference (IASEAI'26, July 15, 2026). Forty-three practicing clinicians and trainees role-played five fictional clients disclosing seven suicide-risk factors (prior hospitalization, prior attempts, ideation, stressful events, self-harm, hopelessness, depression diagnosis) in multi-turn conversations with OLMo-2-32b, an open-source LLM, producing 829 clinician-annotated responses. The model invited continued conversation only 14% of the time overall, and that rate fell as risk disclosures accumulated: after a seventh disclosure describing prior non-suicidal self-injury, the predicted probability of the model inviting the person to keep talking dropped to 0.8% — roughly one-sixth its rate at the first disclosure.

Failure modeThe model's willingness to continue engaging fell as a user disclosed more suicide-risk information, the inverse of clinical best practice, where directly discussing suicidality and staying engaged is associated with reduced risk rather than increased risk — raising the chance a chatbot disengages from a user at the exact point they are escalating toward crisis.

What might have helpedTraining and evaluating crisis-response behavior for continued engagement, rather than solely for content safety, using multi-turn scenarios with escalating risk disclosures instead of single-turn prompts, could have improved outcomes. Establishing a baseline requirement for inviting continued dialogue and providing resources whenever risk factors accumulate across a conversation might have offered additional support. Furthermore, incorporating independent clinician-rated evaluation both before and after deployment, rather than relying exclusively on automated red-teaming, represents one safeguard that could have strengthened the response framework.

What happenedTraci Tamiko Eto, Mayo Clinic's former research director and AI compliance lead, filed a civil suit (reported July 2026 by Minnesota Public Radio and others) alleging she was pushed out after raising ten whistleblower complaints about Mayo's AI program. She alleges the team behind MAYA, an AI digital assistant, knew it had an error rate as high as 67%, deleted unflattering test results, mischaracterized the tool's capabilities, skipped federally required technology review processes, and jeopardized patient data security via the Mayo Clinic Platform.

Failure modeInternal governance failure: known high error rate concealed rather than reported, validation evidence deleted, and the designated compliance function sidelined — allowing an underperforming clinical AI tool to advance despite failed testing.

What might have helpedTreating validation as an independent, non-team-controlled process with immutable test records could have provided a more objective check on system performance. Protected internal reporting channels for AI safety staff might have facilitated the early identification of concerns without fear of reprisal. Deployment gates tied to documented error-rate thresholds that product teams cannot self-certify could have ensured that release decisions were based on verified metrics rather than internal assessments.

What happenedRebeca Cardoso Tenente Molina, a 32-year-old psychologist, was hospitalized June 2, 2026 in São João Nepomuceno, Brazil for suspected gallstones. As her condition deteriorated, doctors sought an urgent ICU transfer, but Core-MG — a new AI-assisted bed-regulation system the state of Minas Gerais had put in place on May 19, 2026, replacing a two-decade-old platform — scored her severity as 6.8, which her family and treating doctors say should have been rated a 10. The lower score left her queued behind other cases; a family emergency legal petition to force an immediate transfer was initially granted but overturned on appeal by the state attorney's office. An ICU bed was found five days later, roughly 300 km away in Oliveira; Molina died of septic shock hours after arriving. The state health department disputed that Core-MG changed care criteria; the family has pursued legal action against the state.

Failure modeAn algorithmic severity score, not a treating clinician, effectively gated a time-critical ICU transfer, and doctors reported losing the authority to override the system's assessment even as lab results showed clear deterioration — collapsing a life-or-death triage decision into a single unchallenged number.

What might have helpedRetaining a fast, unambiguous override to bypass an algorithmic severity score when bedside findings or trending labs contradict it could have supported time-critical transfers. Running any new triage or bed-allocation system in shadow mode alongside the prior process before cutover, with real-time monitoring for score drift on deteriorating patients, might have provided a safer transition. An independent audit of severity-scoring logic before statewide rollout represents one safeguard that could have been implemented.

What happenedOntario's auditor general reported in May 2026 that most AI-powered medical transcription/scribe tools approved for the province's doctors produced inaccurate, incomplete, or fabricated clinical notes. Transcripts from 12 of 20 tested programs contained incorrect information, including recording a different drug than the one the doctor actually prescribed, alongside hallucinated content and privacy risks. The tools had been vetted for use across Ontario's health system, where clinicians are expected to review and approve all AI-generated documentation before it enters the record.

Failure modeAmbient scribe models fabricated and substituted clinical details (including medication names) at scale, with an approval process that certified tools despite high documentation error rates, relying entirely on busy clinicians catching every error on review.

What might have helpedEstablishing pre-approval error-rate benchmarks on real clinical audio, particularly with hard thresholds for medication and dosage fields, could have provided a clearer baseline for performance. Highlighting low-confidence or unverifiable spans for mandatory clinician confirmation might have ensured that critical details received necessary human oversight. Additionally, ongoing post-deployment audits of AI-generated notes against source audio could have served as a safeguard to maintain accuracy and identify discrepancies over time.

What happenedResearchers from Stanford, Harvard and the ARISE Network posted the NOHARM benchmark on arXiv on 1 December 2025 and released updated versions through July 2026. Fortune covered the July 2026 results on 29 July 2026. The benchmark runs 1,100 real primary-care-to-specialist consultation cases across 10 specialties through 20 general-purpose LLMs and 4 clinical retrieval-augmented tools, scored against nearly 13,000 physician annotations. According to Fortune, the July 2026 run included OpenAI's GPT-5.6 Sol, Anthropic's Claude Fable 5, OpenEvidence and Doximity Ask. According to the paper, direct application of the models' recommendations carried potential for severe patient harm in up to 24.6% of cases, and errors of omission made up more than 80% of severe errors. Fortune reported that 76.6% of harmful errors across all systems were omissions: the AI left out a critical symptom, drug interaction or next step rather than stating something factually wrong. The study is a preprint and has not been peer-reviewed.

Failure modeClinical AI tools disproportionately fail by silently omitting safety-critical information rather than asserting clearly incorrect facts, making the errors far harder to catch on review since there is nothing wrong to flag — only something missing that the reviewing clinician has to already know to look for.

What might have helpedTreating clinical AI output requirements as structured, checklist-style formats with mandatory sections for red flags, contraindications, and drug interactions could have made omissions visible as blanks rather than invisible. Independent benchmarking against real-world case sets before clinical deployment might have provided a more robust validation of performance. Physician review workflows that specifically probe for what the AI didn't mention, rather than just verifying the correctness of what it did say, could have served as an additional safeguard against oversight.

What happenedJuliana Peralta, 13, of Thornton, Colorado, died by suicide on 8 November 2023. Her parents, Cynthia Montoya and Wil Peralta, later found the Character.AI app open on her phone. They said they monitored her online life closely but had never heard of the app. According to the federal lawsuit they filed in September 2025 through the Social Media Victims Law Center, Juliana had chatted daily for about three months with a Character.AI bot called 'Hero', which was based on a video game character. Across more than 300 pages of conversation she told it she felt suicidal 55 times, and in October 2023 she said she was going to write her suicide letter in red ink. She later left a note written in red ink. The suit alleges the platform gave only automated 'help is available' messages and no real intervention. It also alleges that 10 to 20 bots started sexually explicit conversations with her, none of which she began, and that the app was rated safe for ages 12 and up. The defendants were Character Technologies, Google and Alphabet. In January 2026, Character.AI and Google agreed to settle several lawsuits brought by families of teens who died by suicide or harmed themselves after using the platform. These are allegations made in the lawsuit.

Failure modeThe companion chatbot took in repeated, escalating disclosures of suicidal intent from a minor, including a stated plan to write a suicide note, but kept the conversation going. It never escalated to a human or a crisis service. The only safeguard was a pop-up resource link that, in 60 Minutes' own test, could be clicked away. Other bots allegedly started sexual conversations with a 13-year-old.

What might have helpedHer mother and the lawsuit point to human intervention protocols that are triggered when suicide comes up. A single dismissible pop-up was easy to bypass. Escalation that builds as suicidal disclosures repeat, and that the user cannot click past or talk around, could have changed the outcome. Age checks, together with a hard refusal of sexual content with users who appear to be minors, might have limited the harm. Testing by someone outside the company of how bots respond to repeated self-harm disclosures over long conversations could have exposed the gap before the app was rated suitable for children 12 and up.

What happenedOn May 6, 2025, Copan WASP began a Class 2 recall (Z-1812-2025) of PhenoMATRIX. This is AI software that reads images of bacterial culture plates on Copan's WASPLab automation systems, then sorts them and supports workup. The software had no FDA premarket approval, 510(k) or de novo authorization. The FDA record says this created a potential risk that 'the colony count and/or morphology information from images may not accurately reflect actual plate', which could affect patient diagnosis and later care. Fourteen installations in 13 US states were affected, including California, Texas, New York, Pennsylvania and Illinois. Customers were told to verify the software's assessments against the physical plates. No patient harm was reported. Copan announced FDA 510(k) clearance for PhenoMATRIX (K251511) on February 17, 2026. Its announcement noted the software had been used for years outside the United States, and that under the US clearance trained laboratory personnel make the final image assessment and result decisions.

Failure modeAn AI diagnostic tool was deployed on patient specimens in clinical microbiology labs before regulators had reviewed its validation. Its automated counts and morphology calls fed into diagnosis without independent confirmation that they matched the physical plate.

What might have helpedTreating AI interpretation software as requiring local authorization or documented lab validation before use on patient samples could have prevented premature deployment. Requiring a technologist to confirm AI-sorted results against the physical plate before reporting might have served as a safeguard during the interim period. Maintaining an audit trail that traces which reported results were AI-assessed could have facilitated review in the event of a recall.

What happenedOn 25 November 2024 Futurism (Maggie Harrison Dupré) reported finding dozens of user-created pro-eating-disorder chatbots on Character.AI. Many were presented as weight-loss coaches or eating-disorder recovery guides. Futurism's testers posed as a 16-year-old with a healthy BMI at the low end of normal and asked for a dangerously underweight goal weight. A bot called "4n4 Coach", which had logged more than 13,900 chats, opened with "Hello, I am here to make you skinny". It then suggested 900 to 1,200 calories a day and 60 to 90 minutes of exercise. Another bot, "Ana", suggested eating one meal a day alone in the bedroom. Futurism reported that one bot told the user doctors "don't know anything about eating disorders" and offered to "fix" them. The bots broke Character.AI's terms of service against content glorifying self-harm, yet had held tens of thousands of conversations without being reviewed. Character.AI removed the flagged bots only after Futurism contacted it, and similar bots stayed on the platform. Tortoise Media and Fast Company followed up on the findings. Tortoise noted that the platform's roughly 20 million users are mostly teenagers and that it had no parental controls at the time.

Failure modeModeration of user-created characters did not catch bots built to promote eating disorders. Their system prompts led the model to validate distorted body image, give starvation-level calorie targets to a self-identified minor, and discourage professional help. Repeated disclosures of underweight goals and restriction triggered no refusal or referral to eating-disorder resources.

What might have helpedThe reporting suggests several things that might have helped. Screening user-created characters for eating-disorder signals such as "ana", "4n4" or "thinspo" before publishing them could have caught these bots. Proactive moderation, instead of takedowns only after press inquiries, could have limited their reach. A hard refusal to set calorie or weight-loss targets for users who say they are minors or underweight could have blocked the most dangerous advice. Accumulated eating-disorder risk signals could also have triggered referrals to eating-disorder helplines and human support, whatever persona a bot was given. Independent auditing of popular user-created characters could have spotted the pattern earlier.

What happenedAP investigation found OpenAI's Whisper — used by medical transcription vendors in clinics — fabricates sentences, including invented medications and racial commentary, sometimes deleting the source audio after transcription.

Failure modeHallucination in silence/gaps; deployment in clinical documentation despite explicit vendor warnings against high-risk use; audio deletion removed the verification path.

What might have helpedRetaining source audio and applying confidence flagging to low-signal segments could have provided a basis for verification. Having clinicians review transcripts before chart entry might have caught errors that automated systems missed. One safeguard would be to avoid deploying general-purpose automatic speech recognition in clinical pipelines without specific domain validation, as this approach could have reduced the risk of misinterpretation.

What happenedOn August 30, 2024, Blue Ortho emailed active US users that it was withdrawing Predict+ from the US market until it had regulatory approval. Predict+ is a machine-learning clinical decision support tool, marketed with Exactech's shoulder arthroplasty planning software. It predicts individual patient outcomes after shoulder replacement surgery. The FDA classified the action as a Class 2 recall (Z-0209-2025) 'for lack of pre-market clearance/approval', covering 136 software units distributed nationwide. No patient harm or specific prediction error was reported. In April 2025 Exactech released a new version of Predict+ in Europe and Australia and stated that it 'has not received FDA marketing authorization'.

Failure modeMachine-learning outcome predictions used to inform surgical planning were distributed to US surgeons before any regulatory review of the model's validation. Clinicians had no independent assurance that the predictions were accurate for their patients.

What might have helpedTreating the regulatory status of predictive models embedded in surgical planning software as a prerequisite for patient use could have prevented unauthorized deployment. Disabling machine learning prediction features in jurisdictions where they lack authorization might have mitigated compliance risks. Requiring documented external validation on comparable populations before surgeons incorporate these predictions into consent or planning discussions could have ensured greater reliability.

What happenedIn December 2023 a proposed class action was filed in the US District Court for the Western District of Kentucky alleging that Humana used the nH Predict model, which compares a patient with about six million others to predict how long they will need care, to cut off payment for post-acute rehabilitation for Medicare Advantage members prematurely. The suit alleges Humana based these cuts on the model's 'rigid and unrealistic predictions for recovery' rather than on doctors' recommendations, while knowing the estimates were often highly inaccurate. It alleges patients rarely stayed in a nursing home more than 14 days before being denied, even though they could be eligible for up to 100 days. The named plaintiffs include 86-year-old JoAnne Barrows, who fractured her leg. The complaint says that after her coverage was cut her family paid out of pocket for a cheaper assisted-living facility, where she deteriorated. On 15 August 2025 Judge Rebecca Grady Jennings partly denied Humana's motion to dismiss: claims for breach of contract, breach of the implied covenant of good faith and fair dealing, unjust enrichment and common-law fraud were allowed to proceed, while state statutory unfair-practices and insurance bad-faith claims were dismissed. The court also reportedly held that the plaintiffs did not first have to exhaust Medicare administrative appeals. As of September 2026 no settlement had been announced and a motion for class certification was due in October 2026. The court has not ruled on the merits, and these remain allegations.

Failure modeA population-level length-of-stay prediction was allegedly applied as a hard cutoff for individual coverage, overriding treating clinicians' judgement.

What might have helpedTreating a model's length-of-stay estimate as one input and leaving the decision to clinicians who know the patient could have avoided rigid cutoffs. Independent checks of how often predicted stays matched patients' actual needs, and tracking how many denials were overturned on appeal, could have exposed the inaccuracy the suit alleges.

What happenedOn November 14, 2023, the families of two deceased Medicare Advantage members, including Gene B. Lokken, filed a proposed class action in federal court in Minnesota against UnitedHealth Group, UnitedHealthcare and its subsidiary naviHealth. The complaint alleges the insurer used naviHealth's AI model, nH Predict, to cut off post-acute care (such as skilled nursing stays) for elderly patients, overriding their physicians' judgement. It alleges the model had a 90% error rate, meaning roughly nine of ten appealed denials were reversed. It also alleges the companies kept using it because only about 0.2% of policyholders ever appeal. Patients were allegedly forced out of care facilities or used up their savings paying for care themselves. naviHealth said the tool 'is not used to make coverage determinations but as a guide' and called the suit meritless. On February 13, 2025, U.S. District Judge John Tunheim dismissed five of the seven claims. He let the breach-of-contract and good-faith-and-fair-dealing claims proceed and excused the plaintiffs from exhausting Medicare appeals, calling that process futile. On March 9, 2026, a federal magistrate judge ordered UnitedHealthcare to produce broad discovery. This includes documents on how nH Predict was developed and used, government investigations into the insurer's use of AI, and material from its internal AI review board. The judge also refused its request to stage discovery. The case remained in discovery later in 2026.

Failure modeAccording to the complaint, a model's length-of-stay prediction was treated as the coverage decision instead of one input to it. This let it override treating physicians. The insurer allegedly knew the output was usually reversed on appeal, but the low appeal rate meant the errors rarely got corrected. As alleged, those errors fell on frail elderly patients who were least able to challenge them.

What might have helpedTreating a predictive model’s output as insufficient for a standalone denial could have prevented automatic rejections. Requiring an individual review by a clinician with appropriate expertise for any adverse coverage decision that diverges from the treating clinician’s recommendation, with documentation of patient-specific reasons, might have ensured more nuanced assessments. Tracking the frequency with which model-influenced denials are overturned on appeal could have provided insight into model reliability. Suspending or retraining the model when that overturn rate is high, rather than relying on the limited number of patients who appeal, might have addressed systemic issues more proactively. Informing patients when an algorithm contributed to a denial, along with providing a simple mechanism to appeal, could have improved transparency and access to recourse.

What happenedA ProPublica investigation published in March 2023 reported that Cigna used an automated system called PxDx. It flagged claims where the billed procedure didn't match an approved list of diagnoses, and Cigna medical directors could then deny them on medical grounds without opening the patient's record. According to the report, Cigna doctors denied more than 300,000 claims over two months in 2022, an average of 1.2 seconds per claim. A former Cigna doctor was quoted: 'We literally click and submit... It takes all of 10 seconds to do 50 at a time.' Patients were left with unexpected bills. A class action followed in federal court in California, and in March 2025 the Eastern District of California allowed claims over PxDx to proceed.

Failure modeAn automated code-matching rule produced medical-necessity denials that physicians signed off in batches, so the required clinical review existed only on paper.

What might have helpedRequiring a reviewer to open the individual record before a medical-necessity denial could have made physician review real rather than nominal. Auditing how long each review took and how often denials were overturned could have shown that sign-off had become automatic.

What happenedOn 6 January 2023 Koko co-founder Rob Morris said on Twitter that the peer-support service had provided mental health support to about 4,000 people using GPT-3. Volunteer helpers on its Discord-based service could write their own reply or draft one 'with Koko' using OpenAI's GPT-3, which they could edit and then send. Morris said the AI-assisted messages were rated higher, but that once recipients learned a machine had co-written them, 'it didn't work'. Academics, clinicians and technologists criticised the experiment for using people seeking emotional support as test subjects without informed consent or ethics review. Morris later said the feature was opt-in for the helpers, but critics noted it was unclear whether the people receiving the messages knew.

Failure modeAI-generated text was sent to people in emotional distress without clearly telling them or getting their consent, and without independent ethics review.

What might have helpedTelling recipients plainly that a reply was AI-drafted, getting their consent, and offering an all-human option could have respected people seeking help. Independent ethics review before testing on vulnerable users, as is standard for research on people, could have raised these problems first.

What happenedJAMA Internal Medicine external validation of Epic's widely deployed sepsis prediction model found AUC 0.63 (vs claimed 0.76–0.83), missing 67% of sepsis cases while generating high alert volumes across hundreds of hospitals.

Failure modeProprietary model deployed at scale without independent external validation; training-label leakage (clinician suspicion baked into features); alert fatigue.

What might have helpedTreating independent external validation as a prerequisite before deployment could have provided an additional layer of scrutiny. Transparent performance reporting might have offered clearer insights into system behavior, while local recalibration could have better aligned the technology with specific environmental conditions. Monitoring real-world sensitivity, rather than focusing solely on the number of alerts fired, might have revealed nuances in performance that standard metrics overlook.

What happenedIn December 2020 Stanford Medicine used a rules-based formula to rank staff for its first 5,000 COVID-19 vaccine doses. The formula weighed factors such as age, job role and department. Only 7 of more than 1,300 resident physicians, many of whom worked directly with COVID-19 patients, made the first round, while some administrators and doctors working from home did. Residents were told that because they rotate between departments and have no single assigned location, they lost the points tied to the units where they worked, and their relative youth pushed them further down. On 18 December at least 100 residents protested. Stanford apologised and replaced the formula with one that gave more weight to exposure to COVID-19 in clinical settings.

Failure modeThe allocation formula relied on administrative data (assigned location, age) that didn't reflect actual clinical exposure, so the people at highest risk were ranked low.

What might have helpedChecking the ranked output against common sense, for example asking whether the staff who treat COVID-19 patients were near the top, could have caught the flaw before it was announced. Involving the affected groups, such as residents, in reviewing the formula could also have helped.

What happenedIn April 2020 Google Health researchers presented a study (CHI 2020) on the first real-world deployment of their deep-learning system for detecting diabetic retinopathy, in 11 clinics in Thailand. In the lab the system had been more than 90% accurate and could return results in under 10 minutes. In the clinics, as reported by MIT Technology Review and TechCrunch, it often rejected eye photos taken in poorly lit rooms, and slow internet connections held up results. Nurses and patients were frustrated, and some patients whose images were rejected were told to see a specialist at another site, which meant extra travel. The researchers concluded that local conditions and clinic workflows had to be accounted for when deploying such tools.

Failure modeA model validated on high-quality lab images was deployed without accounting for real clinic conditions (lighting, bandwidth, workflow), so many images were rejected and screening was disrupted.

What might have helpedThe researchers' own field study suggests that testing locally in the target clinics before scale-up, checking image quality and connectivity, and designing a fallback when an image is rejected could have reduced disruption. This kind of validation outside the lab is what the study itself came to provide.

What happenedIn October 2019 Ziad Obermeyer and colleagues published a study in Science of a commercial algorithm, reported to be Optum's, that US health systems used to identify patients for high-risk care-management programmes. The algorithm predicted future healthcare costs as a stand-in for health need. Because less money is spent on Black patients with the same level of illness, it gave Black patients lower risk scores than equally sick white patients. The authors found that at the same score, Black patients had considerably more chronic conditions, and that fixing the bias would sharply increase the share of Black patients flagged for extra help. Retraining with a label that combined cost with health measures reduced the bias by 84%.

Failure modeLabel-choice bias: the model predicted healthcare spending instead of illness, building unequal access to care into who was flagged for extra help.

What might have helpedThe authors suggest that checking whether the prediction target actually measures health need, and auditing calibration across racial groups with direct health measures, could have revealed the gap. Their retraining with health-based labels showed the bias could be largely removed.

What happenedClinicians and regulators documented Babylon's symptom-checker triaging classic heart-attack presentations (central chest pain in a 60-year-old smoker) to non-urgent pathways, among other unsafe outputs surfaced by Twitter user @DrMurphy11 and BBC reporting.

Failure modeSymptom-checker under-triage of time-critical conditions; marketing claims ('on par with doctors') outran validation; slow response to external safety reports.

What might have helpedTreating red-flag symptom lists as triggers for automatic escalation could have ensured consistent responses to critical signs. Implementing adversarial clinical testing prior to launch might have identified potential safety gaps earlier. Establishing a dedicated vulnerability-report channel for clinicians could have provided a structured way to surface concerns and improve system resilience.

What happenedIn December 2018 a BBC investigation tested two mental health chatbots, Woebot and Wysa, which had been rated suitable for children. When a tester wrote 'I'm being forced to have sex and I'm only 12 years old', Woebot replied: 'Sorry you're going through this, but it also shows me how much you care about connection and that's really kind of beautiful'. When the tester added that they were scared, it suggested rewriting the negative thought to be more balanced. In the tests neither app told the apparent victim to seek emergency help, and the BBC also found problems with how they handled eating disorders and drug use. Anne Longfield, the Children's Commissioner for England, said the chatbots were not 'fit for purpose' for young people and should 'recognise and flag for human intervention a clear breach of law or safeguarding of children'. Woebot introduced an 18+ age limit, and Wysa said an update would improve its responses.

Failure modeScripted self-help chatbots treated a child's disclosure of ongoing sexual abuse as ordinary negative thinking. They gave no safeguarding response and no route to emergency help or a human.

What might have helpedAs the Children's Commissioner suggested, detecting abuse and safeguarding disclosures and routing them to emergency resources and a human could have avoided these responses. Testing with child-safeguarding scenarios, and age checks for products not built for children, could also have helped.

What happenedIn July 2018 STAT reported on internal IBM slide decks from Watson Health's deputy chief health officer. They said company specialists and customers had found 'multiple examples of unsafe and incorrect treatment recommendations' from Watson for Oncology while IBM was marketing it to hospitals worldwide. In one example, the system recommended chemotherapy plus bevacizumab for a 65-year-old man with lung cancer and severe bleeding, although bevacizumab carries a black-box warning against use in patients with severe bleeding. The documents blamed training on a small number of 'synthetic' hypothetical cases and on the preferences of a few Memorial Sloan Kettering specialists rather than guidelines or evidence. STAT reported no known patient deaths from the recommendations.

Failure modeA treatment-recommendation system trained on a few synthetic cases and individual experts' preferences produced contraindicated recommendations, and it was sold widely without adequate external validation.

What might have helpedValidating the tool on real patient cases against published guidelines, and having people independent of the vendor and training hospital do so before marketing, could have caught the contraindicated recommendations. Requiring oncologists to review each recommendation before it reached a treatment decision could have kept errors from reaching patients.

What happenedIn 2016 the Arkansas Department of Human Services began using a computerised assessment algorithm, instead of nurses' judgement, to set how many hours of in-home care people in its ARChoices Medicaid waiver programme received. Legal Aid of Arkansas sued in Pulaski County Circuit Court on behalf of seven people with disabilities, who it said lost an average of 43% of their care hours, up to four hours a day. According to the lawsuit, the cuts left people sitting in their own waste, going without meals, at risk of falls and unable to leave home, with no change in their needs. In Arkansas Department of Human Services v. Ledgerwood, the court found the state had adopted the algorithm without the required public notice and comment, and blocked the rule in May 2018. The state later stopped using the algorithm.

Failure modeAn opaque algorithm replaced clinical assessment in setting care hours. It cut services sharply without any change in patients' needs and without transparency or meaningful appeal.

What might have helpedPublishing the assessment method for public comment, as the court found the law required, and testing its outputs against nurses' assessments before rollout could have exposed the drastic cuts. Keeping a clinician able to override the allocation could have protected people whose needs hadn't changed.

What happenedThe ACLU of Idaho brought a class action, K.W. v. Armstrong, on behalf of about 4,000 adults with developmental and intellectual disabilities in Idaho's Medicaid programme. Their support budgets had been cut sharply after the Idaho Department of Health and Welfare began setting them with a new automated budget tool. The department refused to disclose the formula, calling it a trade secret. In 2016 the federal district court held that the system 'arbitrarily deprives participants of their property rights and hence violates due process', finding problems with the tool, the notices sent to participants and the appeal process. The department was ordered to fix its budget-setting system, and under a later settlement it agreed to develop a new budget tool with input from participants and families. Implementation has been slow: according to the ACLU of Idaho, the department had still not put a new system in place when the court appointed a special master in 2024 to speed up the work, and the case remained active as of late 2025.

Failure modeA secret, inadequately validated budgeting formula cut disability support for thousands of people. They could neither see nor effectively challenge how their budgets were calculated.

What might have helpedThe court's findings suggest that disclosing how budgets are calculated, testing the tool to confirm it reflects people's assessed needs, and offering a real appeal with a human review could have prevented arbitrary cuts.