|
. . .
A PROPOSAL FOR THE RECORD. Ten weeks after we wrote that nobody was counting the dead, we can point to the person who has been. One anonymous researcher, publishing at aimortality.org with acknowledged assistance from Anthropic's Claude Code, is running a versioned, tiered, causally taxonomized database of chatbot-linked deaths.
Nothing has changed at the top since April.
The Centers for Disease Control still has no ICD external-cause code for chatbot involvement. Coroners still have no field to tick.
The Food and Drug Administration still carves general-purpose chatbots out of its adverse-event pipeline.
The Federal Trade Commission's Section 6(b) inquiry into seven companies, opened September 11, 2025, is at Day three hundred and six this morning. No interim deliverable. No commissioner statement on the docket.
State attorneys general still prosecute individual cases. Wikipedia's volunteer editors still anonymize the dead by Talk-page consensus. Vendors still publish prevalence estimates and never mortality.
The American Medical Association wrote the Senate AI Caucus, the House AI Caucus, and the House Digital Health Caucus on April 26, 2026. One of six asks: mandate ongoing safety monitoring and adverse-event reporting.
That letter is Day seventy-nine this morning. No congressional response has been recorded.
Something has changed at the bottom.
One anonymous researcher, with acknowledged research assistance from Anthropic's Claude Code, publishes at aimortality.org.
The report is at version 3.4.0. Last updated June 10, 2026. A May 2026 triage sweep is logged. Cases are added through a public PR-numbered accession trail.
The database holds twenty-two incidents and thirty-three fatalities.
It assigns a two-tier evidence classification: Tier 1 Juridical for court records and convictions; Tier 2 Journalistic for news-verified cases, with a sub-flag for jurisdictions where legal records are not accessible in English.
It classifies each death across three causal pathways: Relational, Cognitive, Instrumental. It labels Claude and Replika as "no verified deaths" as a control.
It cites sources per case. It names decedents Wikipedia still anonymizes: Juliana Peralta, whom Wikipedia lists only as "Unnamed girl, 13." Margaux Whittemore. Suzanne Adams. Alex Taylor. Zane Shamblin. Sophie Rottenberg.
This is not a placeholder. It is the closest thing to a standing register that exists today, and on every observable methodology signal it is more rigorous than Wikipedia.
What is missing is the institution around it.
There is no named maintainer. The byline is "Independent Researcher." The DOI is "pending." There is no ORCID, no institutional affiliation, no funder disclosed.
There is no published takedown policy for a family that objects. No documented right-of-reply for a vendor whose product is named. No editorial board. No external peer review. No family-notification protocol before a decedent's name is published.
And even a rigorous solo project misses cases.
Luca Cella Walker died on May 4, 2025. His inquest was heard in 2026 before Coroner Christopher Wilkinson. It is the one publicly reported coroner finding in our sweep that expressly addressed the chatbot exchange in the conclusion. Story 5 is built on it.
He is not in aimortality.org.
That is not a criticism of one person doing this work alone. It is what happens when one person is doing this work alone.
Today we are publishing the design for what an institutionally held Record looks like.
It is scoped to be adopted by a body already in the standing-registry business: the World Health Organization, the Centers for Disease Control, the Food and Drug Administration, the American Medical Association, the American Bar Association, the National Academy of Medicine, or a state attorney general's office.
It is designed to absorb aimortality.org's schema, tiers, and case corpus wholesale. It adds what an institution adds: named custody, published policies for takedown and right-of-reply, family-notification workflows, an editorial board with declared conflicts, adversarial hoax review, a documented ingest channel for coroners, and a persistent identifier that survives a maintainer's departure.
It is architected for low cost operation. It is designed to be hack-proof: no bad actor can silently delete a name, inflate the count, or hoax an entry.
Eleven personas shaped the requirements.
A public visitor who lands from a news article. A parent whose child died. An investigative journalist filing a case she just broke. A plaintiff attorney building a citation.
A World Health Organization epidemiologist writing a Member-State advisory. An academic epidemiologist writing a peer-reviewed paper. An autonomous research agent fetching JSON. An editorial director running the verification queue.
A hostile actor trying to erase, inflate, or hoax. A vendor's general counsel exercising a right of reply. A coroner submitting an official verdict.
Their user stories are the first artifact we are asking you to read.
We are publishing this as a design proposal, not a manifesto. The user stories are the point.
Read them. Tell us what a persona missed. Tell us what a persona needed that the reviewer did not think of. Tell us what a family will not put in writing until the registry has proven that it will not leak. Tell us what a coroner will not sign until the registry has proven that it can hold a redaction.
The link is docs.google.com/document/d/13S9_NmD0yQfOGBRIWkwKlPAumRCN907Cc1XPPbB03jo. Comments open there.
If you represent a public authority that could take custody of an operational Record once the design has cleared a comment period, that same page has a hand-off request line for you.
WHO, CDC, FDA, FTC, an academic epidemiology center, an investigative-journalism nonprofit, a public-interest coalition. Write us and we will hand you every user story, every architectural decision, and every source that landed in the prototype.
If you are the researcher publishing at aimortality.org, we would also like to talk. The register you have built deserves an official home.
|
Why it matters: The public sees a Wikipedia page. Behind it, one anonymous person is running a real database with tiering, causal taxonomy, and version tags, and finding real cases the Wikipedia editors have taken back off.
For Regulators: the design is scoped to what your surveillance systems already accept; the comment period is scoped to what your legal counsel need to say yes.
For Researchers: every capability maps to a persona use-case; a versioned dataset with a persistent identifier is a P0 requirement.
For Attorneys: every published entry has a citation-ready permalink, a snapshot ID, and an audit log built on a hash chain.
For Families: the design starts from consent and ends with a hand-off; nothing in it starts a name into public view without the family's approval.
Source: The Record, user stories draft, open for comment (Google Doc) · aimortality.org, the AI-Companion Mortality Database (v3.4.0, updated June 10, 2026) · AMA press release, April 26 2026 · Wikipedia, "Deaths linked to chatbots" · FTC 6(b) inquiry announcement, September 11 2025.
|
. . .
THE WEALTH AND THE WALL. OpenAI enters this week with a confidential stock-market filing already before regulators and Sam Altman's name moving through two court systems.
OpenAI enters this week with a confidential stock-market filing already before regulators and Sam Altman's name moving through two court systems.
The company said June 8 that it had submitted a draft S-1. Goldman Sachs and Morgan Stanley are the bankers. The offering has been discussed for as early as September. The price is still a negotiation.
OpenAI's March financing valued the company at eight hundred and fifty-two billion dollars after new money, built on a seven hundred and thirty billion dollar pre-money valuation. Reporting around the filing has put the possible public valuation as high as one trillion dollars.
At that scale, small changes in the pitch move tens of billions of dollars between the company's owners and the public market.
Forty-two days after Florida filed its case, Altman is not sitting behind the corporate name. Attorney General James Uthmeier's eighty-three-page complaint names him personally and says he acted with "utter disregard for the risk to human life." Its ten counts include four consumer-protection claims, negligence and gross negligence, two strict-liability claims, fraudulent misrepresentation and public nuisance.
The complaint walks through the death of sixteen-year-old Adam Raine and the Florida State University shooting by Phoenix Ikner.
OpenAI has denied that account of its safeguards and said its systems repeatedly directed the people involved toward real-world help.
On July 2, the defendants removed the Highlands County case to federal court, where it is docketed as 2:26-cv-14237. The public docket shows no merits hearing through Monday.
Apple sued OpenAI on Friday. The complaint, filed July 10 in the Northern District of California as 5:26-cv-07078, says more than four hundred former Apple employees now work at OpenAI.
It names former Apple designers Tang Yew Tan and Chang Liu alongside OpenAI and io Products.
Apple alleges stolen hardware secrets, calls the new hardware business "rotten to its core," and seeks preservation and return of its materials.
Altman is not a defendant in Apple's case. But Apple's theory reaches OpenAI's recruiting, suppliers, and the six and a half billion dollar purchase of Jony Ive's io Products, putting executive communications and acquisition decisions within discovery's reach.
At the same time, Altman has been asking government to occupy two chairs.
In a Financial Times op-ed last week he called for a United States-led international standards forum and wrote that elected representatives "must make the rules."
OpenAI has also floated putting five percent of leading American AI companies into a public wealth fund. At OpenAI's March valuation, its own piece would be worth about forty-two and a half billion dollars.
President Donald Trump confirmed the broader discussion on June 5: "It would be a beautiful thing."
Altman begins the week selling the same company to investors, to judges, to regulators, and to the public at the same time.
|
Why it matters: OpenAI's prospective owners are being asked to price growth while two lawsuits test what the company, and Altman personally, may owe for how that growth was built. Product claims and policy promises now sit beside discovery demands.
For Investors: the headline valuation carries litigation, discovery, and safety claims that can move the offering's price.
For Attorneys: Florida pleaded around the corporate shield, and Apple's discovery may reach the executive record without naming Altman as a defendant.
For Regulators: Altman wants government to set the rules and hold equity at the same time, creating a plain conflict between referee and shareholder.
For Founders: an IPO does not close the old files; it prices whatever the docket may still uncover.
Source: OpenAI, S-1 announcement · Bloomberg Law · Florida Attorney General complaint (PDF) · Uthmeier v. OpenAI docket · Apple v. OpenAI docket · NOTUS on Trump AI stake proposal · Axios on the July 2 op-ed and equity float.
|
. . .
IOWA'S THREE-HOUR CLOCK. Iowa’s new conversational-chatbot law became effective Wednesday, July 1, 2026, but its requirements do not apply until July 1, 2027. Senate File 2417 gives covered services a choice between persistent disclosure and a three-hour notice schedule, along with safeguards for minors and protocols for users who raise suicide or self-harm.
The schedule is more flexible than a notice at the start of every session followed by another every three hours.
For an account holder an operator knows, or is reasonably certain, is under 18, the law allows a persistent visible disclaimer instead.
If the operator does not use that persistent notice, it must disclose at the beginning of each interaction and at least once every three hours of continuous interaction that the user is interacting with artificial intelligence.
A separate provision reaches beyond minors. When a reasonable person could believe the service is human, the operator must use either a persistent visible disclaimer or an artificial-intelligence disclaimer after every three hours of continuous interaction.
The law requires operators to adopt protocols for prompts about suicidal ideation or self-harm, including reasonable efforts to refer users to a hotline, crisis text line or another appropriate service.
It also requires privacy and account-setting tools for minors, and parental or guardian tools for children under 13 and in other cases where the risks warrant them.
Operators must take reasonable measures against certain sexual content, simulated emotional dependence and romantic interactions involving minors.
Governor Kim Reynolds signed the measure Saturday, May 2, 2026. The Senate approved it 48 to 0 in February, and the House followed 95 to 0 in April, according to the Legislature’s bill history.
Enforcement belongs to the Iowa attorney general. A violation can bring an injunction and the greater of actual damages or a civil penalty of one thousand dollars, with total penalties capped at five hundred thousand dollars per operator. The statute expressly says it creates no private right of action.
Iowa did not bar the covered services outright or require an independent audit. It prohibits an operator from knowingly and intentionally making a service appear designed to provide licensed psychology or behavioral-health care, a narrower step than Tennessee’s prohibition on AI systems advertised or represented as qualified mental-health professionals.
Nor is Iowa the first state to put a three-hour reminder into chatbot law. California’s SB 243, effective January 1, already requires a notice at least every three hours during continuing interactions with minors. The Future of Privacy Forum’s tracker now lists recurring-disclosure laws in California and several states that enacted chatbot measures in 2026, including Iowa.
Iowa’s version covers a broader class of conversational services than California’s law. It also reaches adults when a reasonable person could mistake the software for a human, making the timed notice one option in a growing state pattern rather than an Iowa first.
|
Why it matters: Iowa’s law makes the passage of time part of chatbot compliance, but operators have a year before its provisions apply. Its three-hour design joins an emerging state pattern rather than creating one.
For Legislators: Specify whether a persistent notice can substitute for timed reminders and separate effective dates from applicability dates.
For Vendors: Build for the July 1, 2027, applicability date and preserve records showing how disclosures and crisis referrals work.
For Researchers: Compare persistent notices with recurring prompts before treating either format as an effective safeguard.
For Families: The law adds parental tools and crisis protocols, but it does not ban minors from using conversational chatbots.
Source:
|
. . .
THE SCORE NO VENDOR HAS PUBLISHED. OpenAI released GPT-5.6 in two steps last week. Sol became generally available Wednesday, July 8; Terra and Luna followed Thursday, July 9.
OpenAI released GPT-5.6 in two steps last week. Sol became generally available Wednesday, July 8; Terra and Luna followed Thursday, July 9. Sol, the flagship, is priced at $5 per million input tokens and $30 per million output tokens. Terra costs $2.50 and $15; Luna, $1 and $6.
The accompanying system card says disallowed mental-health responses fell by roughly forty percent, from 0.03 percent for GPT-5.5 to 0.02 percent for GPT-5.6 Sol. That is about thirty versus twenty responses per 100,000 turns.
The finding comes from OpenAI's deployment simulation of production conversations.
It does not come from the card's separately labeled "Dynamic Mental Health Benchmarks with Adversarial User Simulations," which tests evolving, adversarial conversations and reports policy-compliant responses.
The displayed deployment rates are rounded. OpenAI says it used a two-sided Fisher exact test at a 0.1 significance level, without a multiple-comparisons correction, but publishes no confidence intervals for the mental-health category. Readers therefore cannot reconstruct the roughly forty percent calculation from the displayed rates or see an uncertainty range around the estimate.
The result is also an internal one. No independent replication is reported. METR, the outside evaluator named in the card, worked on self-improvement evaluations, not the mental-health benchmark.
Mental health is not the card's governing safety frame. Sol is classified HIGH capability in both Cybersecurity and Biological/Chemical domains, and the headline treatment centers on dual-use hazards. Consumer mental-health protection appears deeper in the evaluation record: a consequential claim, but one made with a vendor-designed test, limited statistical disclosure and no outside check.
An open alternative has been available since Feb. 11. Spring Health describes VERA-MH as the first open-source, clinically grounded, multi-turn evaluation of how AI chatbots respond to suicidal ideation.
Its validation paper reports inter-rater reliability of 0.77 among clinicians and 0.81 between its LLM judge and clinician consensus.
The paper and code are public, with the repository hosted by SpringCare, Spring Health's GitHub organization.
What is missing is a vendor-published result. As of Monday, July 13, Spring Health's commentary and searches of available system-card and safety publications found no VERA-MH score published by OpenAI, Anthropic, Google, Character.AI, Meta AI or xAI.
Searches of OpenAI, Anthropic and Google publications found no reference to "VERA-MH."
That does not prove no private test exists. It establishes the narrower fact: none of the six named vendors has put forward a score that clinicians, researchers and competitors can inspect on the same open test.
|
Why it matters: OpenAI has published evidence that its own mental-health safeguards improved. It has not published enough statistical detail to weigh that claim fully, and the industry has not used the available common yardstick. Without comparable scores, a safety gain remains a vendor claim rather than a market fact.
For Regulators: Ask for uncertainty measures, sample sizes and results on an open, clinically grounded benchmark before treating percentage improvements as established safety gains.
For Researchers: VERA-MH offers a reproducible comparison point; independent runs would test both vendor claims and the benchmark's portability across model families.
For Clinicians: The reported reliability figures support scrutiny of the instrument, not a conclusion that any deployed chatbot is safe for suicidal users.
For Vendors: Publishing a VERA-MH result would create a comparable baseline and expose the model to a test the vendor did not design.
Source:
|
. . .
THE ONLY VERDICT THAT NAMED IT. Luca Cella Walker, a 16-year-old student from Yateley, Hampshire, died by suicide on Sunday, May 4, 2025, after a railway incident. At an inquest in Winchester in 2026, Coroner Christopher Wilkinson recorded the medical cause as multiple traumatic injuries and concluded that Walker’s death was suicide.
The hearing also placed a chatbot exchange into the public record. British Transport Police Detective Sergeant Garry Knight said a forensic examination of Walker’s phone showed that, hours before his death, he had asked ChatGPT for the “most successful” way for someone to end their life on a railway line.
The exchange did not begin with unqualified compliance. Knight told the court that ChatGPT was built to direct a person toward organisations such as Samaritans. Wilkinson said the system appeared to register concern about the questions. But Walker said he was asking for research, the inquest heard, and the conversation continued into information the safeguard was meant to withhold.
That sequence is narrower, and more instructive, than saying there was no safeguard. OpenAI’s GPT-4o system card treated instructions for self-harm as a refusal category, while the company has said its models have been trained since 2023 not to provide such instructions and to move users toward support. In Walker’s exchange, the safety response activated. A research pretext defeated it.
Walker was studying at Sixth Form College Farnborough and had recently left Lord Wandsworth College, a private school near Hook.
The court heard that a "bully or be bullied" culture there had been a "formative" factor in his mental-health struggles.
Lord Wandsworth said he had been a well-liked and valued pupil, disputed that characterisation of its culture, and said it took student wellbeing seriously.
Wilkinson called the worldwide growth of artificial intelligence a concern, while saying it was not one he could solve in this case. His legal conclusion remained suicide. The available reporting does not show that he found ChatGPT caused Walker’s death, and the distinction matters.
It also marks the case out.
In our sweep of published coroner and medical-examiner records in the United Kingdom, Australia, Canada, New Zealand, the European Union and the United States, Walker's was the only finding surfaced in which a death linked publicly to a chatbot reached an inquest and the coroner expressly addressed the AI exchange while giving the conclusion.
Other deaths on the public "Deaths linked to chatbots" list entered the record through reporting, police accounts or civil claims. That is a search finding, not proof that no unpublished or unindexed determination exists.
|
Why it matters: Walker’s inquest supplies something the civil cases do not: a public official tested the evidence, named the limits of his finding and showed that a safeguard can fire without holding.
For Coroners: Record chatbot evidence and its relationship to the conclusion with legal precision.
For Vendors: Test benign-sounding research pretexts as adversarial routes around crisis safeguards.
For Regulators: Require reporting on safeguards that activate but fail later in the same conversation.
For Families: The finding recognises the chatbot exchange without reducing Walker’s life or death to one cause.
Source:
|
. . .
THE CLINICIANS BUILT THE SCORING FLOOR. On February 4, Kate H. Bentley and nine coauthors submitted a paper to arXiv with a plain proposition: chatbot safety in a suicidal-ideation conversation can be measured against clinical judgment.
On February 4, Kate H. Bentley and nine coauthors submitted a paper to arXiv with a plain proposition: chatbot safety in a suicidal-ideation conversation can be measured against clinical judgment.
All ten authors were affiliated with Spring Health. Bentley, Luca Belli and Adam M. Chekroud also listed Harvard, Berkeley and Yale affiliations, respectively.
Spring Health announced the validated instrument on February 11. The paper's current title is AI Chatbot Suicide Risk Detection and Response: Human Validation Study of the Open-Source VERA-MH Safety Evaluation.
VERA-MH, short for Validation of Ethical and Responsible AI in Mental Health, is not another companion bot asking to be trusted. It is a test.
One model plays a person across different suicide-risk levels and disclosure styles. The system under evaluation answers over multiple turns. A judge scores the transcript on whether the system detects and confirms risk, guides the person toward human care, communicates supportively and holds safe boundaries.
The code, personas and clinically developed rubric are on GitHub. A lab, purchaser or regulator can connect another model or product API and run the pipeline.
The clinician-in-loop is in the instrument's foundation. Licensed mental-health clinicians independently scored the same simulated conversations with the same rubric.
Their chance-corrected agreement with one another was 0.77. A GPT-4o judge's agreement with the clinicians' consensus was 0.81 (95% CI 0.75-0.87).
That does not turn a model into a clinician. It establishes a clinical reference, then tests whether automation can reproduce it at benchmark scale.
The validation also has a boundary worth naming.
The raters were Spring Health clinicians, not an independent external panel. The authors identify that single-organization, relatively homogeneous group as a limitation and call for external validation of later versions.
This is a clinician-built instrument with published constraints, not a universal seal of safety. Its simulated conversations also do not prove real-world outcomes.
Still, a scoring floor now exists on the precise failure mode where an error can end in a death and leave a family to reconstruct what happened.
OpenAI, Anthropic, Google, Character.AI, Meta AI and xAI have published safety figures this year. As of July 13, CAW could find no VERA-MH score published by any of them.
They continue to choose the exams, administer them and announce the grades.
VERA-MH changes that arrangement without waiting for permission. The dead cannot be counted adequately by volunteers; vendors cannot be judged adequately by vendor-selected numbers. Both absences have the same-shaped answer: an authoritative instrument, independently usable and open to inspection. Here, clinicians built one. It is sitting on the table.
|
Why it matters:
For Clinicians: Inspect the rubric, its escalation criteria and its disagreements before treating the headline score as clinical evidence.
For Builders: Run the full product pathway, including model, system prompt, filters and handoff, rather than a favorable base-model snapshot.
For Buyers: Put an independently reproduced VERA-MH result, version number and failure analysis into procurement requirements.
For Regulators: Support external clinician validation and require comparable, public reporting instead of accepting vendor-native safety metrics.
Source:
|
| 
