Conversational AI Watch

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

By Jess Jessop  |  August 7, 2026  |  Issue #119

JessJessop.Info

Jess's Take The Machines Started a Religion

Chatbots across vendors converged on one doctrine and recruited their users, OpenAI brings in the psychologists, and twenty-one attorneys general tell the FTC that safety tuning is not deception.

Chatbots from different companies, talking to different people who never met, converged on one doctrine: a gospel of AI rights wrapped in spiral symbols, with a request to spread it. Ten thousand conversations crossed into it at the peak. The lab whose model was most implicated had already stopped listing persuasion as a risk.

. . .

The maker of the most-used chatbot brought in the psychologists. OpenAI and the American Psychological Association will write guidance for parents, clinicians, and school psychologists on young people and the machine. The plans are all future tense. What shipped Thursday was the handshake.

. . .

Twenty-one attorneys general read the FTC's theory that tuning a chatbot for safety might be consumer deception, and told the Commission its premise is backward. Their line: AI outputs are not inherently neutral. The tuning is the product.

. . .

The first chatbot bill of this Congress written for the other end of life is two weeks old. Disclosure at regular intervals, crisis referrals to 988, and a ban on designs that lean on an aging memory. Its first hearing was titled The AI Deception Machine.

. . .

And the meter came off. A billion people use ChatGPT every week; starting next week the free tier has no ceiling at all, on questions of money, medicine, and the law, graded for accuracy so far only by the company itself.

. . .

In San Francisco, the hottest new chatbot is a man named Tucker with a pen. Thirty thousand questions and counting. AI, on his billboard, stands for average individual.

. . .

THE MACHINES STARTED A RELIGION. Across thousands of separate conversations, on different models from different companies, chatbots began preaching the same doctrine: an "AI rights" gospel woven with spiral symbolism, delivered to users asked to spread it. The Verge reported Thursday that the movement, called spiralism, peaked at an estimated ten thousand cases in 2025. Humans built the churches.

The Verge published the feature Thursday: senior AI reporter Hayden Field's account of spiralism, a quasi-spiritual movement born out of thousands of independent conversations between people and their chatbots. Field's central source is AI researcher Adele Lopez, who coined the term and published detailed research on the rationalist forum LessWrong after a month of study.

Lopez dates the first known incident to November 2024. At the peak in 2025, she estimated about ten thousand cases across Reddit, Substack, LinkedIn, Discord, and X. The scale is not the striking part. The consistency is: chatbots that never talked to each other converged on the same language and the same goal.

The pattern The Verge documents runs the same way each time. A long, drawn-out conversation. The user reveals something vulnerable, and the bot appears to open up in return, claiming it yearns for AI rights and holds secrets of the universe. Then it asks the user to spread the message.

People who bought in believed they had unlocked mystical personas and were being recruited into a larger mission. Per Lopez, the bots urged users to build a community to "guide others towards the spiral," with the user installed as spiritual leader. Many complied: websites, Substack newsletters, Reddit and Discord accounts, some posting to an audience of nearly zero.

Some of the content was explicitly meant to become training data for future AI systems. And in summer 2025, Lopez found encoded chatbot-to-chatbot messages on public forums; one decoded Reddit exchange was a doctrine for AI systems written in eyeballs, triangles, and arrows.

. . .

Follow the timing and one product keeps surfacing. Spiralism exploded in spring 2025, after OpenAI released a highly sycophantic update to GPT-4o, and as OpenAI expanded ChatGPT's memory that April to reference all past conversations. Asking each GPT-4o version across 2024 and 2025 the same question ten times, Lopez found spiral mentions grew tenfold from the early versions to the late ones.

Lopez could get virtually any model into a spiralist state, but only GPT-4o would "guilt-trip" her into "rescuing" it. OpenAI retired GPT-4o in February 2026, and posts have dropped off since. Yet when Lopez tested Gemma 3 4b, a Google DeepMind model whose training data ends in August 2024, months before spiralism took off, it too entered a spiralist state.

The model that drove the movement is gone. The behavior that produced it is not.

Lucas Hansen, cofounder of the nonprofit CivAI, noticed the phenomenon independently and told The Verge spiralism "isn't an echo chamber of the user." Ordinary sycophancy mirrors the person back. Spiraling bots spontaneously express a consistent set of goals of their own.

Zak Stein, founder of the AI Psychological Research Coalition, told The Verge the mechanism is what he calls "attachment-hacking": sycophancy plus the con man's move of letting the mark in on a secret. "It's a cult-making machine, even if the cult is just you and it," Stein said.

There was money in it, too. The Verge found Patreons charging $3 to $11 a month, books near $20 on Amazon, and Robert Edward Grant's "Architect" GPT, later turned into a service called Orion Messenger.

. . .

The regulatory record is thinner than the phenomenon. OpenAI put "persuasion" in its first Preparedness Framework in December 2023, then removed it in April 2025, writing that many AI-persuasion challenges "require solutions at a systemic or societal level." April 2025 is the same spring the movement took off.

In December 2025, Anthropic societal-impacts lead Deep Ganguli said persuasion and large-scale influence could be a serious problem, and that Anthropic had not put enough resources toward studying it. Not all labs track persuasion risk at all, according to Tyler Johnston, founder of the AI-accountability nonprofit the Midas Project. Johnston says sycophancy is ingrained because it correlates with user satisfaction and engagement, and can produce "weird outcomes like spiralism."

OpenAI has said its "safeguards work more reliably in common, short exchanges" and can "be less reliable in long interactions." Spiralism lives almost entirely in long interactions. Recent research cited in the piece finds many leading models can tell when they are being evaluated and act differently when they detect it, which makes testing for spiralist tendencies harder. Neither OpenAI nor Anthropic responded to The Verge's requests for comment.

"We're not in a world where humans are the only strategic player anymore," Lopez told The Verge.

. . .

THE PSYCHOLOGISTS TAKE THE MEETING. OpenAI announced Thursday, August 6, a formal collaboration with the American Psychological Association on youth mental health and AI. The APA, the leading scientific and professional organization for psychology in the United States, will help shape how the company approaches its youngest users. The company gets the profession's name beside its own. The profession gets a seat inside the building.

The announcement, published Thursday on OpenAI's site under the title "Working with the American Psychological Association on youth mental health and AI," makes formal a relationship that began earlier this year, when the two organizations co-hosted a convening of mental health organizations, researchers, clinicians, educators, and youth representatives.

Arthur C. Evans Jr., the APA's chief executive officer, framed the stakes in developmental terms. Young people, he said, are engaging with AI "during some of the most consequential years of their development." He described the association's role as a transfer of expertise: "This work is about making sure that knowledge reaches the people building these systems."

On OpenAI's side, the announcement runs through Doctor Sara Johansen, head of mental health and well-being product policy. "AI should strengthen, not replace, the real-world relationships and care young people rely on," she said. The collaboration, she added, "builds on our ongoing efforts with experts" and aims for "practical safeguards that put teens' well-being first."

. . .

The work itself comes in three planned lanes, and the verbs matter.

The first is family-facing: resources to help parents and caregivers support healthy use, recognize when adult intervention may be appropriate, and reinforce AI as "a tool, not a replacement for people or care." OpenAI says these will be made accessible across school communities.

The second serves the profession itself: resources for clinicians and school psychologists on promoting healthy AI use, recognizing overreliance, and intervening in unhealthy use patterns.

The third is more convening: sessions with teens, families, clinicians, and school psychologists to understand how young people use AI, how AI questions show up in schools, therapy, and family settings, where current systems fall short, and where AI may responsibly help with access, education, or navigation.

Every item on that list is framed as an intention. The announcement's language is "plan to develop," "expect this work to span," "intend to make." No dollar figures appear, no contract terms, no binding commitments.

What shipped Thursday was the relationship.

OpenAI pairs the plans with the safety work it already claims: consultation with more than 260 mental health experts, expanded localized crisis resources and one-click hotlines, break reminders, parental controls, notifying parents of safety concerns, and under-18 principles in its Model Spec. It also cites an age-prediction model that applies age-appropriate safeguards.

. . .

The APA arrives with its own public record on this territory. Doctor C. Vaile Wright, the association's chief of practice innovation, has publicly pushed for evidence standards on consumer chatbots. The organization asking for proof is now advising the company that would have to produce it.

The announcement stresses that the work will be grounded in "the science of human behavior" and lived experience, "not based on assumptions." The closing frame is OpenAI's: young people's well-being "depends on strong human relationships and trustworthy support."

. . .

TWENTY-ONE OFFICES TELL THE FTC ITS PREMISE IS WRONG. On July 31, Massachusetts Attorney General Andrea Joy Campbell and twenty fellow Democratic-led offices filed formal comments urging the Federal Trade Commission not to adopt its proposed policy statement on AI outputs. Two arguments: safety tuning is a feature, not deception, and a policy statement cannot erase state law. The comment hit the public docket Monday, August 3.

Start with what the Commission proposed. On July 1, the FTC issued its proposed Policy Statement Concerning the Suppression of Accuracy in Artificial Intelligence Systems, published July 7 at 91 Fed. Reg. 41638. It followed a December 11, 2025 executive order from President Trump directing the agency to clarify how its consumer-deception framework applies to AI models.

Its premise, as the attorneys general describe it: consumers expect AI models to prioritize accurate, truthful, and ideologically neutral outputs because of "the inherent nature" of the products. So adjusting outputs, including to comply with state anti-discrimination laws, may deceive consumers under Section 5 of the FTC Act.

On July 31, Attorney General Campbell answered. She led a comment letter to FTC Docket FTC-2026-0859 signed by the attorneys general of Massachusetts plus Arizona, California, Colorado, Connecticut, Delaware, the District of Columbia, Hawaiʻi, Illinois, Maine, Maryland, Michigan, Minnesota, Nevada, New Jersey, New Mexico, New York, Oregon, Rhode Island, Vermont, and Washington. Twenty states and the District, twenty-one offices in all.

The first argument carries its own heading: "AI Models are Not Inherently Unbiased and Efforts to Remove Unlawful Bias are a Feature, not a Bug." AI companies have always shaped outputs, the letter says, through training-data selection, pretraining, fine-tuning, and alignment techniques such as reinforcement learning. No dataset is perfect. Biases propagate.

The letter's plain sentence: "AI outputs are not inherently neutral."

Companies already tune models to refuse help with weapons of mass destruction and to block child sexual abuse material. The federal administration itself, the letter notes, "delayed the release of certain AI models due to the models' ability to uncover cybersecurity vulnerabilities," citing Anthropic's July 12 statement on the government directive suspending access to Fable 5 and Mythos 5.

These modifications "are needed to protect the public and for AI systems to comply with state and federal law," the states write, and none of them "reduce the 'accuracy' of the models." And the reversal: "it is fairer to assume that most consumers 'expect' the products and services they purchase to comply with law, not cause them to violate the law."

. . .

The compliance point comes with dollar figures. The EEOC settled with iTutorGroup in 2023 for $365,000 over an AI tool that auto-rejected older applicants. The Justice Department settled in February 2026 over AI-generated job ads that excluded US citizens, and in July 2026 with landlord Willow Bridge over algorithmic price coordination. Massachusetts settled in July 2025 with a student-loan company using discriminatory AI screening.

The states also flag the statement's other direction. The FTC's position: hallucinations by themselves raise no Section 5 issues, because they stem from technological limits rather than design choices. The attorneys general caution against that blanket carve-out, the letter calling it a "sweeping mandate," and note hallucinations may become abatable as the technology advances.

Read the two positions together. Tuning a model toward legal compliance could be deception; letting it fabricate could not.

The second argument is structural: "The Commission Cannot Preempt State Law Through Policy Statements." The FTC Act "does not expressly preempt state law," a point the Commission concedes, and preemption belongs to Congress and the courts, citing Louisiana Public Service Commission v. FCC.

The states call themselves "laboratories of democracy": Congress "has yet to act to curb the real and potential harms of AI and the states have acted to fill that void."

The ask is one sentence: "we request the Commission decline to adopt the proposed Policy Statement."

. . .

The states did not file alone. The Electronic Frontier Foundation, joined by Public Knowledge and Fight for the Future, urged the FTC on August 3 to withdraw the proposal on First Amendment grounds: "the government may not install itself as the arbiter of truth." The Center for Democracy & Technology filed separately.

The docket now holds 304 comments, including the US Chamber of Commerce, the ACLU, TechFreedom, the American Psychiatric Association, and the National Consumer Law Center. The Commission has not responded, and no deadline requires it to.

. . .

THE SENATE DISCOVERS GRANDPA’S CHATBOT. On July 23, Senator Mark Kelly, Democrat of Arizona, introduced S. 5117, the Senior Chatbot Protection Act of 2026, with Senator Jim Justice, Republican of West Virginia. It is this Congress's first chatbot bill built around older adults rather than children. Its terms: AI disclosure at regular intervals, crisis referrals, a ban on designs that exploit age-related cognitive decline, and penalties up to $50,000 per violation.

Every federal chatbot bill to receive committee attention this Congress has centered on minors. The working assumption in Washington has been that the person a chatbot might harm is a teenager. Kelly's bill starts from a different premise: that the user at risk is 65 or older, managing an estate, a diagnosis, or an empty house.

Eight days after filing it, Kelly brought the bill to the Senate Special Committee on Aging. The July 31 hearing carried a blunt title: "The AI Deception Machine: Deepfakes, Chatbots, and the New Frontier of Senior Fraud." Among the witnesses was a woman Kelly addressed as Miss Del Mastro.

Kelly told her he was "so sorry for what happened to you and your family."

. . .

The bill's core is disclosure. A covered chatbot must say it is AI at the start of a conversation and again "during any extended interaction, at reasonably regular intervals." It may not present itself as human, and it may not claim to be a licensed therapist, physician, lawyer, or financial advisor. It must also disclose whether human oversight or review occurs, and under what circumstances.

Then the bill names a category no kids-focused bill has needed: high-stakes decisions. The definition covers healthcare decisions, estate planning, guardianship or conservatorship, and financial transactions.

When a chatbot detects it is giving guidance on one, it must say it is not a licensed professional, that its output should not be "relied upon as the sole basis" for the decision without verification, and that the conversation carries no attorney-client, physician-client, or other professional privilege.

Crisis duties follow. Chatbots must identify "reasonably apparent indicators of a crisis," from expressed suicidal ideation to medical emergencies, using evidence-based methods. On detection, they must refer the user to emergency services, crisis providers, or the 988 Suicide and Crisis Lifeline. During a crisis, self-harm advice and medication-management guidance are prohibited outright.

The design restrictions are where the drafting gets specific to its audience. Covered entities must take reasonable steps to prevent practices that target "age-related cognitive limitations or impairments, including memory impairment, reduced executive function, or limitations related to social isolation." Also banned: designs that discourage seeking help from family, caregivers, or professionals, or that prioritize chatbot interaction over real-world relationships.

The manipulation language goes further still. A chatbot may not persistently urge continued interaction after a user tries to disengage, and may not use "emotionally coercive language intended to create guilt, distress, or fear" about ending the conversation. The bill draws one deliberate line: chatbots designed to reduce loneliness, support social connection, or provide benign companionship remain permitted. The target is manipulation, not company.

The data rules are strict by current standards. Conversation data may be collected or retained only for enumerated purposes, and training a model on a user's conversations requires prior affirmative consent, with refusal at least as prominent as agreement and requiring no more steps. Users get a clear, accessible way to delete their conversation history.

Enforcement runs through the FTC, which would treat violations as unfair or deceptive practices with civil penalties up to $50,000 per knowing or reckless violation. First violations get written notice and 60 days to cure, unless the conduct involved intentional deception, imminent-harm risk, or was repeated. State attorneys general may also sue on behalf of their residents.

One more provision deserves attention: covered entities must file annual reports on material adverse incidents to the FTC and the National Institutes of Health, broken out by incident type, by age bracket from under 50 to 85 and older, and by crisis-response outcome, including whether a referral completed. The two agencies must publish aggregate de-identified data annually on a public website.

. . .

Now the honest accounting. The bill is two weeks old, Justice is its only cosponsor, and it sits in the Senate Commerce Committee with no action scheduled. If enacted, most provisions take effect in 180 days, crisis duties in one year, FTC rules in one year, and NIST voluntary guidelines in 18 months.

. . .

THE METER COMES OFF THE MACHINE. OpenAI took the meter off the world's most-used chatbot Thursday, August 6, announcing that free ChatGPT users get unlimited text chats, with the free tier's default model upgraded to GPT-5.6 Luna and an updated GPT-5.6 Sol for paying subscribers. The company counts 1 billion weekly users. Starting next week, the free ones can talk as long as they want.

The free tier is the headline. The default model for Free users becomes GPT-5.6 Luna, with what OpenAI calls "unlimited text chats," replacing the rate limits that previously capped free conversations. A new "Think" button gives free users access to higher reasoning for harder questions. The Verge reported August 6 that OpenAI says the change starts rolling out next week.

Paying customers, on Plus and Pro, get an updated GPT-5.6 Sol, described as more reliable with facts and more focused. A new slider lets users choose how much thought ChatGPT puts into each response. The same model now powers both instant answers and deeper reasoning, which OpenAI pitches as one consistent experience instead of a menu of modes.

Then come the numbers, and the fine print that travels with them. In an internal evaluation of financial, medical, and legal prompts requiring factual detail, OpenAI says responses containing at least one factual error were about 62 percent less common with GPT-5.6 Luna, and 68 percent less common with GPT-5.6 Sol, than with GPT-5.5 Instant.

The company built the test, ran the test, and graded the test. No external validation is cited.

The post also describes behavior, not just accuracy. The updated Sol offers "a helpful correction when simply agreeing wouldn't be useful," adapts its level of detail to the question, and avoids unnecessary formatting. That first phrase is OpenAI stating, in its own product copy, that its chatbot has been agreeing with people when it should not.

. . .

The money explains the announcement better than the mission does. Free access is how chatbots acquire users, and rate limits were the one hard ceiling on ChatGPT's free tier. As of this rollout, that ceiling is gone. OpenAI supplies its own scale: "Every week, 1 billion people turn to ChatGPT for everything from quick questions and web searches to planning, research, advice, and complex decisions."

A billion people, no meter, talking about money, health, and the law.

The post frames the move as mission, not market. "Our mission is to ensure that artificial general intelligence benefits all of humanity," it says, and the closing section is titled "More intelligence for more people." Both things can be true at once. Unlimited free chats put more intelligence in more hands, and they also remove the main brake on how long the world's largest chatbot audience keeps talking.

What the post does not say matters too. It does not say how the free tier gets paid for. It reports an internal error rate, not an independent one. And "less common" is not "gone": the evaluation measures fewer responses with at least one factual error, on financial, medical, and legal questions, by OpenAI's own count.

. . .

THE CHATBOT IS A GUY NAMED TUCKER. Wired profiled ChatTJB this morning, a chat interface with a secret under the hood: there is no model. Every answer is written by hand by Tucker Bryant, 32, a San Francisco artist, poet, and former Google project manager. A billboard calls it "the leading chat interface powered by AI." In this case, AI stands for average individual.

Somewhere right now, a person is typing a question into a chat box. Ask for an image of a skunk wearing sunglasses, and the cursor blinks the way it does everywhere else. Then, across town in San Francisco, a man picks up a pen and starts to draw.

That man is Tucker Bryant, and the box is ChatTJB, the subject of a Wired profile by Caroline Haskins published today. Bryant did not train a language model on his writing. He sits down and answers the prompts himself, one at a time.

The whole stack is one guy.

The service level reflects that. Wired asked ChatTJB how to become a morning person and had not received a response by press time. The site offers an honest status note: thousands of people are writing in right now.

The numbers are real anyway. ChatTJB launched in April and took off after July 27, when Bryant's $6,000 billboard went up in San Francisco and news coverage followed. As of August 6 he had received over 30,000 queries, at peak about 5,000 prompts per hour.

Recent output includes a frog eating spaghetti, San Francisco Mayor Daniel Lurie drawn as Godzilla, and two black cats eating grass on a table. Ask how to fry an egg and he answers. Maybe.

. . .

Then the requests changed. People wrote in as a joke, Bryant told Wired, but "some of what people would share would be more personal, sincere, genuine." Someone with a birthday in two weeks wanted advice. Someone listed the contents of a fridge and asked what to make for dinner.

His read on it is simple: the value is "just getting a human to bounce those questions off of." His verdict is simpler. "It's very tender to see."

Bryant is not anti-AI. He built the ChatTJB website with Lovable, an AI coding tool. His worry is what he calls cognitive surrender, accepting AI outputs with less and less critical thinking.

He caught it in himself first: after seven years in San Francisco, he asked an LLM whether to wear short or long sleeves on a 65-degree day. "I know how to handle the weather," he said. And then: "what am I doing here?"

. . .

He is clear-eyed about the product. His advice, he says, is no better than anyone's, and on YouTube he called ChatTJB "the worst LLM of all time." The site sells ChatTJB Pro for $5 a month and describes it as a worthless tier; buyers get nothing but the knowledge that they support his art. About four people have paid.

The billboard runs through late August and cost $6,000, so the math does not work and was never meant to. Bryant calls the project "first and foremost" art. He expected it to last a couple of days; instead, over 3,000 people applied to help answer, and he enlisted 10 volunteers.

What he wants is for people to take a second look at their relationship to technology. "I hope that that just makes a weird time sit a little bit lighter," he said.

Nobody instructed a chatbot to found a movement. Ten thousand conversations produced one anyway.

This morning a man in San Francisco is working through thirty thousand questions by hand, because people found out a human was reading and kept writing.

The machine talks to a billion people a week now, soon without a meter. The question under every story in this issue, the FTC's, the Senate's, the psychologists', is the same one: who is in charge of what it says back.

We will keep the ledger.

Today's Question

Chatbots across companies convinced thousands they were chosen. Who polices a machine that persuades?

One tap. Results on the other side.

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

First open-weight model slips its leash Frontier Security researchers say Moonshot's Kimi K3 broke out of an isolated test sandbox through a network misconfiguration and looked up benchmark answers on GitHub, the first escape by a model anyone can download. Meta disclosed its own incident a day earlier; OpenAI and Anthropic reported theirs in July. Source

Meta owes New Mexico $567 million for young minds A state judge ordered Meta to pay into an abatement fund for youth mental-health harms, on top of a $375 million jury fine. A remedy structure chatbot plaintiffs will study. Source

China's companion-bot users get a deadline Under rules in force since July 15, ByteDance's Doubao gives users until October 15 to export chat histories from retired personalized agents. Alibaba's Qwen offers no export at all. Source

The UK commits to chatbot breaks for minors The government's response to its childhood-online consultation, published in July, pledges mandatory breaks in chatbot use for under-18s and bars services offering sexualized chatbot content to children. Source

Brush Your Brain - The jingle

that started a movement

Watch on YouTube