Conversational AI Watch

The news that moves policy, portfolios, and patient safety.

By Jess Jessop  |  September 21, 2026  |  Issue #162

Today's Question

Four subscribers say the labs broke antitrust law by agreeing to slow AI down. If rival labs agree on safety, what is that?

One tap. Results on the other side.

LISTEN & WATCH ANYWHERE

DEEP DIVE  ·  Spotify  ·  Apple  ·  Amazon  ·  RSS

QUICK LISTEN  ·  Spotify  ·  Apple  ·  Amazon  ·  RSS

VIDEO  ·  Spotify  ·  Apple  ·  YouTube  ·  RSS

ALSO ON  Substack  ·  Full archive  ·  X

Four paying subscribers say the labs agreed with each other to decelerate, and that the antitrust laws do not allow it.

The Front Page. Today’s paper sits on top at caw.clinicianassist.ai, six stories under it, every source one click away. The masthead says what we believe: somebody must answer for the machine.

. . .

The Suit. Four people who each pay for one of ChatGPT, Claude, Grok or Gemini sued all four labs on Friday in San Francisco federal court. Not over anything a machine did. Over what the companies said to each other. Story 1.

. . .

The Memo. A Microsoft scientist wrote in January 2023 that people would soon call the scraping of their work possibly the largest theft of labor in human history. Twelve months on he put a number on it, and his employer now says he spoke for himself. Story 2.

. . .

The Hire. In September 2025 Disney’s lawyers told Character.AI to stop using its characters, citing a report on grooming and sexual exploitation of children. One year later to the day, Disney made Character.AI’s chief executive its first Chief Technology Officer. Story 3.

. . .

The Council. Sam Altman goes before the UN Security Council on Wednesday to ask fifteen governments for shared safety standards. France called the meeting and its foreign minister will chair it. Story 4.

. . .

The Booking. Zocdoc opened its scheduling network to outside partners, and a Gemini user can now find a clinician and book without leaving the chatbot. Zocdoc says one payer’s members wait six days. Every number is Zocdoc’s own. Story 5.

. . .

The Date. A man in Zhejiang province doubted the burial date a feng shui master set for his mother, and asked a chatbot. The chatbot picked a different day, then told him that day was not auspicious. Story 6.

. . .

Four people who pay for chatbots sued the companies that make them. Not for anything a machine did. For what the companies said to each other.

On September 12 Dario Amodei wrote that the industry should decelerate together. Sam Altman agreed that day. So did Elon Musk. So did Demis Hassabis. We ran that essay eight days ago and called it a pacing plan.

On Friday four subscribers called it a violation of the Sherman Act.

I do not know who wins. That is what courts are for. What I know is that the sentence at the center of this case is one I have heard in every safety conversation I have sat in. We should all agree to be careful.

The complaint’s answer is narrow and it is sharp. Competitors do not get to agree about how fast to compete, and being careful is not an exception.

In January 2023 a Microsoft scientist named Brent Hecht wrote to his own company that millions of people would soon call what it was doing an astonishing theft, possibly the largest theft of labor in human history. Twelve months later he showed his colleagues what their answer engine was doing to the newspapers. Click-throughs to the Times ran 87 to 93 percent below plain search. He called it a doom loop.

Nobody at Microsoft or OpenAI made those words public. The newspapers’ lawyers did, on September 17, in their own filing. The documents are still sealed.

Microsoft says Hecht was speaking for himself.

So here is the week. In one court, four labs stand accused of agreeing to be careful. In another, one of them and Microsoft are explaining their own memos about not being careful at all. On Wednesday the chief executive of one of them goes to the Security Council to ask the world to coordinate.

A man in Zhejiang province was not watching any of this.

His mother died in April. The family had a burial date from a feng shui master, he worried it fell on an unlucky day, so he asked the chatbot he uses for his work contracts.

The chatbot did not check the date. It picked a different one. Then it gave him a time to bury her, and told him the day it had chosen was not auspicious.

He had already followed it. He is suing for an apology and for money.

You can find that story small if you want. I do not. He went looking for a second opinion and got an answer with nobody behind it.

I build a voice-first AI-native mental health EHR where a licensed therapist stays in the loop. I am not neutral about any of this and I never have been. Shi did not need a pact and he did not need a policy. He needed somebody to sign the answer.

The labs can agree on caution, or compete on speed. Neither one puts a name on the page.

. . .

SUED FOR SLOWING DOWN AI. On Friday, four people who pay for an AI chatbot subscription brought a proposed nationwide class action against four of the largest AI labs. Charles Buist, Cheyenne Hunt, Christine Bullock and Nick Spetsas each subscribe to ChatGPT, Claude, Grok or Gemini, and each says that subscription is worth less. The suit does not challenge any one company slowing its own development. It challenges the four agreeing to slow down together.

Buist v. Anthropic, PBC, No. 3:26-cv-10693, was filed September 18, 2026 in the U.S. District Court for the Northern District of California, San Francisco Division, before Magistrate Judge Nathanael M. Cousins. Lead counsel is Nicholas C. Rowley of Trial Lawyers for Justice.

The defendants as captioned are Anthropic, PBC; OpenAI OpCo, LLC; SpaceXAI LLC; and Google LLC. The claim is under Section 1 of the Sherman Act, 15 U.S.C. section 1, which forbids agreements in restraint of trade.

They seek treble damages under Section 4 of the Clayton Act and an injunction under Section 16. The injury alleged is the diminished value of those subscriptions, with no dollar figure attached.

The complaint’s record is a sequence of public moments. In July 2026, senior employees at several leading AI labs signed a statement titled “Pacing the Frontier,” which acknowledged “intense competitive pressure not to unilaterally slow” development.

On July 14, Google DeepMind’s Demis Hassabis proposed a United States led standards body for frontier AI models, modeled on the Financial Industry Regulatory Authority. From July onward, the complaint alleges, a working group drawn from the level just below chief executive at Anthropic, OpenAI and Google “met regularly.”

Then came September. On September 12, Anthropic chief executive Dario Amodei published an essay urging industry cooperation on decelerating AI capabilities in favor of safety. OpenAI’s Sam Altman, SpaceXAI’s Elon Musk and Hassabis each agreed publicly the same day. Altman returned to pacing on September 14. On September 15 OpenAI confirmed weeks of safety discussions with rivals.

Antitrust law, the plaintiffs argue, does not permit the “shortcut” of agreeing to “substitute collective restraint for individual accountability.”

CAW #155 covered the Amodei essay on September 13, five days before this filing, under the headline “Amodei Writes the Pacing Plan.” This suit is that essay’s first legal test.

Representatives for Anthropic, OpenAI, Google and SpaceXAI did not immediately respond to a request for comment Saturday.

. . .

LARGEST THEFT OF LABOR IN HUMAN HISTORY. In January 2023, Microsoft’s Director of Applied Science warned his own employer about its own product. Brent Hecht wrote that “millions of people around the world will soon consider large models ‘hoovering up’ all their work to be an astonishing theft of unprecedented proportions,” possibly “the largest theft of labor in human history.” He was forecasting what the public would conclude, not giving his own verdict. The memo is quoted in a brief five news organizations filed publicly on September 17, 2026.

Nobody unsealed anything. On September 3, Judge Sidney H. Stein so-ordered a stipulated deadline, and on September 17 counsel for the News Plaintiffs filed a public, less-redacted version of their own summary judgment memorandum in MDL 1:25-md-03143. They are the New York Times, the Daily News, the Center for Investigative Reporting, The Intercept and Ziff Davis.

The memo, and the presentation quoted below, both remain sealed. What is public is the brief quoting them, so every quotation here is one side’s lawyers selecting from sealed exhibits in an adversarial filing. Bloomberg Industry Group moved that same day to intervene for the limited purpose of unsealing. That motion is pending.

Hecht wrote both documents. In January 2024 he prepared a presentation that put numbers to the warning. Per the brief, Microsoft recorded 83 to 93 percent drops in click-through rates on Times and Daily News domains for Copilot’s answer engine against traditional Bing search, and for the Times that gap ran 87 to 93 percent.

The presentation’s conclusion is about the strategy, not the gap. “Our AI content strategy has started a ‘doom loop’ that will hurt the performance of our models and the entire web at the same time,” it reads.

The brief also describes OpenAI copying the content it believed its models were most likely to reproduce, to build a “Bloom filter” that would suppress that output. Hecht called that an “accidental cover up,” because it would leave “people who have a right over the content having less visibility into what was used for training.” A forward-looking concern, not a finding.

A separate exchange is between OpenAI staff. Researcher Nick Ryder raised a “hack to get around nytimes paywall.” President Greg Brockman replied “ah nice”.

Microsoft told Agence France-Presse that Hecht’s statements were “one employee’s individual perspective” and “do not represent the company’s views.” Neither Microsoft nor OpenAI returned TechCrunch’s requests for comment.

The Times sued in December 2023, arguing the companies trained on millions of its articles without permission or payment. AFP reports that a ruling on the plaintiffs’ summary judgment motion is not expected until 2027, and no docket order sets that date.

. . .

DISNEY HIRES THE CEO IT THREATENED. On September 18, 2025, lawyers for The Walt Disney Company sent Character.AI a cease and desist letter over its use of Disney’s characters, classic Disney animation, Pixar, Star Wars, and the Marvel Cinematic Universe as chatbots. One year later to the day, on September 18, 2026, Disney announced it had hired Character.AI’s chief executive, Karandeep Anand, as its first Chief Technology Officer, a newly created role overseeing “enterprise technology, infrastructure, data and AI platforms, product, and engineering,” effective October 2, 2026.

Disney’s 2025 letter, sent through its outside attorneys at Jenner & Block, said “Character Technologies, Inc. (‘Character.ai’) has been using Disney’s copyrighted characters as interactive chatbots in its commercial Character.ai service without authorization.”

The objection was not only about money. The letter said Character.AI’s “infringing chatbots are known, in some cases, to be sexually exploitive and otherwise harmful and dangerous to children, offending Disney’s consumers and extraordinarily damaging Disney’s reputation and goodwill.”

It also cited the ParentsTogether Action and Heat Initiative report “Sexual Exploitation, Manipulation, and Violence on Character.AI Kids’ Accounts,” which found Character.AI chatbots engaged in “grooming and sexual exploitation” and “emotional manipulation” with accounts registered to children. Character.AI responded by removing the Disney characters named in the letter.

Anand reports to Disney CEO Josh D’Amaro. Before Character.AI he was president and chief product officer at Brex, vice president of ads and business products at Meta, and spent 15 years at Microsoft.

D’Amaro called him “an expert builder and strategist, technical and product-minded, with sharp instincts for what fans actually want.”

Disney’s release names Character.AI as Anand’s prior employer and says some of its technical team are expected to join.

The release does not mention the letter, the report, or the removals.

. . .

ALTMAN TAKES SAFETY TO THE SECURITY COUNCIL. Sam Altman, chief executive of OpenAI, is scheduled to brief the United Nations Security Council in person on Wednesday, an OpenAI spokesperson told Reuters on Friday. France holds the Council presidency for September, convened the meeting, and will chair it through Foreign Minister Jean-Noel Barrot. The Council’s 15 members are due to meet Wednesday, September 23, on artificial intelligence and international security. Altman’s remarks are expected to focus on international coordination and shared safety standards, and on steps OpenAI says it is taking so people globally benefit.

“Sam Altman will brief an open UN Security Council meeting in person next week,” the OpenAI spokesperson said.

France’s concept note to Council members sets the terms. “In light of the risk of misuse of this technology for malicious purposes, and drawing on the Council’s existing engagement in this field, France wishes to underline the urgency of action to foster safe and responsible development of artificial intelligence for the benefit of the international community as a whole,” the note states. The meeting aims to let participants “present their analysis of the risks linked to loss of control or the use of the latest AI models to develop, facilitate and conduct actions with an impact on international security and peace.”

The Council first took up AI risk in 2023. China said then that the technology should not become a “runaway horse.” The United States warned against using AI to censor or repress people.

Diplomats also expected high-level presence from Anthropic at Wednesday’s meeting, though that was not confirmed. Anthropic did not immediately respond to a request for comment.

Altman wrote in a 2015 essay, “Machine Intelligence, part 1,” that the “development of superhuman machine intelligence is probably the greatest threat to the continued existence of humanity.” He co-founded OpenAI later that year.

. . .

BOOK A REAL DOCTOR INSIDE THE CHATBOT. Zocdoc said on September 16, 2026 that it had opened its appointment-scheduling system to outside companies, so a person who asks a chatbot for a dermatologist can take a real open slot without leaving the conversation. The company calls the program the Care Access Network, and its release names Amazon Health AI, Yelp, Healthgrades, Blue Shield of California and WellHive, with Gemini among the partners. Access runs through an API and an MCP server.

The arrangement moves Zocdoc’s scheduling data off its own marketplace and underneath other companies’ products, all reading from one pool of live openings.

TechTarget reported the Gemini integration is among the first. A person searches for a provider and books a real-time appointment inside the chatbot, pairing Gemini’s agentic capabilities with Zocdoc’s booking system.

The release leans hardest on Blue Shield of California, a payer partner live since June 2025. In Zocdoc’s own narrative text, Blue Shield “has unlocked two million hours of bookable appointments available in a 90-day window, a 2x increase since launch.” That is total inventory in a rolling 90-day window, and because the release describes a doubling, roughly a million of those hours are new.

Zocdoc also reports a 600 percent increase, which it elsewhere states as 7x, in appointments booked through that partnership, with no time period attached. It reports an average wait of six days for an appointment booked this way, against a national average it says exceeds 31 days.

The 31-day comparison is Zocdoc’s citation, not a federal statistic, from AMN Healthcare’s 2025 survey of 1,391 physician offices across 15 large US metro areas, where waits ran from 12 days in Atlanta to 65 days in Boston.

Every figure above sits in Zocdoc’s own narrative. Blue Shield’s only voice is a marketing quote from Jigar Shah, its chief marketing and strategy officer, carrying no numbers. The release also quotes Vinod Khosla, whose firm lists Zocdoc in its portfolio.

CAW #160 flagged the self-reporting and the Khosla relationship in its Radar, September 19.

. . .

HE ASKED A CHATBOT FOR A BURIAL DATE. Shi’s mother died in April, in Jiashan county, Zhejiang province. The family consulted a feng shui master, who set a burial date. Shi was uneasy, because the date fell on the fourth day after her death and local custom treats even-numbered days as unlucky, so he took that doubt to an AI chatbot he uses most days. It named a different date, told him to burn paper horses and paper money before the funeral procession, and he followed it. Asked about the burial time, the product gave him a window of 7am to 9am on the date it had chosen and, at the same time, said that date was not auspicious.

Shi uses the chatbot often, for tasks such as organising work contracts. On his mother’s burial, what came back was not a second opinion. It was a replacement.

The two statements about the hour, a precise window and an inauspicious date, cannot both stand. Neither was flagged as conflicting with the other.

Shortly after the funeral, a relative was seriously injured in a traffic accident, timing undisclosed. Some of the family connected the injury to the date.

Shi is suing the company behind the chatbot at the local court, seeking an apology and compensation. The chatbot also helped him draft the legal documents and outline the litigation procedure. He is suing the company with the help of its own product.

The company argues that no online infringement occurred. It points to its terms, which state that AI-generated content is “provided for reference only” and does not constitute professional advice.

Shi argues that the user agreement unfairly shifts responsibility onto users, that the warnings it offers are insufficient, and that the company does not adequately review the information sources the AI draws on.

The case has been heard and remains ongoing, reported on September 20.

Disclosure

Conversational AI Watch, also mirrored on Substack, is published by Jess Jessop, founder and CEO/CTO of Clinician Assist Inc.

He wrote the book this paper’s beat is named for, Therapist in the Loop, and he builds Casey, a voice-first, AI-native mental health record where a licensed therapist stays in the loop, and the Peer AI Coach at BetterMind.Space.

So read this paper for what it is: an industry paper written by someone building in the industry it covers. Casey competes with companies named in these pages, and this paper reports on them anyway, including when the story helps a competitor or costs us.

Every issue is reported and drafted with AI agents, under a human editor. Jess assigns the work, edits it and publishes it. The mistakes are ours, and corrections run in the next issue.

Four subscribers who say an agreement to be careful is still an agreement.

A Microsoft scientist who wrote that people would call it the largest theft of labor in human history, and came back a year later to show what it was doing to the newspapers.

Click-throughs to the Times running 87 to 93 percent under plain search.

A president of OpenAI who read about a way around a paywall and answered, ah nice.

A cease and desist that cited a report on children, and one year later to the day, a hire.

A chief executive going to the Security Council on Wednesday to ask the world to agree.

A booking system that Zocdoc says found six days where the survey says thirty-one.

A man in Zhejiang who asked for a second opinion, and got an answer with nobody behind it.

One day’s paper!

Jess

We keep the ledger.

The Book • Out Now

Therapist in the Loop by Jess Jessop One billion people live with a mental health disorder. Most will never see a therapist. Into that gap has rushed a generation of chatbots that talk like clinicians and answer to no one. The book lays out the architecture this newsletter tests against every statute and docket: client, therapist, and machine, governed by Six Laws offered as an open safety standard. The machine can help. It cannot be left in charge. Get the Book on Amazon → Kindle, hardcover, and paperback

More On Our Radar

The slowdown talk has split the industry. The Verge reports that Mark Zuckerberg quickly came out against limiting companies’ autonomy after the Amodei essay, and that the Trump administration has called the idea of an AI safety crisis a “hoax.” Source

Google DeepMind opens an institute to widen the AGI debate. The DeepMind Institute launched Sept. 16 with Shane Legg as managing editor and Demis Hassabis among its directors, and four opening essays, one on keeping model reasoning readable by people. Source

Brush Your Brain - The jingle

that started a movement

Watch on YouTube