The Brilliant Friend

Conversational AI Watch

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

By Jess Jessop  |  July 3, 2026  |  Issue #85

Yesterday's Pulse

The FTC moved on state chatbot laws the week two state laws locked in. Who should write the chatbot rulebook?

424 readers answered

JessJessop.Info

Jess's Take The Brilliant Friend

The FTC built its case from the industry’s own marketing. The Supreme Court took the commission that wrote it. One state law switched on. The researchers filed their dissent.

Anthropic told the world that talking to Claude feels like a conversation with a brilliant friend. It published that line as research. On Wednesday the federal government filed it as evidence.

. . .

I spent yesterday reading footnotes. The FTC's case for pointing consumer-protection law at state AI statutes is built from the industry's own sales copy, Anthropic's above all. Four days before that document shipped, the Supreme Court quietly handed the agency that wrote it to the White House.

The states wrote anyway. Tennessee's one-page law switched on Wednesday morning, the only one in its class that did. Pennsylvania's bill survived its first floor test the same day, and the tracker everyone reads got the vote wrong. We checked the legislature's own ledger.

And the researchers with the best data in the field filed their dissent: the blanket teen ban everyone assumes is coming is not what the evidence orders.

. . .

This is CAW eighty-five.

Everything this week was somebody's own words, coming back.

. . .

THE BRILLIANT FRIEND, FOOTNOTED. Wednesday the Federal Trade Commission published a nine-page proposed policy statement on AI accuracy. Yesterday we covered the press release. Today we walk the document. Its legal predicate, what consumers expect from AI, is assembled almost entirely from the industry's own words. Anthropic's above all.

The document is the "Proposed Policy Statement Concerning the Suppression of Accuracy in Artificial Intelligence Systems," docket FTC-2026-0859. Chair Andrew Ferguson and Commissioner Mark Meador approved it 2-0. They are the only sitting commissioners. It implements President Trump's December 2025 executive order on state laws requiring, in the statement's words, "alterations to the accurate outputs of AI models."

The theory is Section 5 deception. Deception needs a consumer who expects accuracy. The statement has to prove people trust chatbots. So it quotes the people who sold the trust.

. . .

Footnotes 35 and 36 reproduce claude.com marketing at length. "Meet your thinking partner... like an expert in your pocket." Claude can "coach you through difficult situations."

Footnotes 35 and 38 mine Anthropic's April 30, 2026 research publication, "How People Ask Claude for Personal Guidance." That paper described a "conversation with a brilliant friend." The Commission's footnote cites it for the usage numbers: 12% of Claude conversations concern relationship navigation, 6% personal development, 4% spirituality.

Footnote 39 supports the claim that consumers accept AI outputs without further fact checking over 90% of the time. The source is a Forbes article reporting Anthropic's own 91% figure. Footnote 35 also reaches xAI, which markets Grok as a "truth-seeking" companion.

. . .

The carve-outs are as tailored as the citations. Footnote 46 exempts hallucinations from the deception theory. Footnote 40 exempts "limits on model use to prevent cybersecurity attacks," language that maps onto the Commerce order that darkened Anthropic's flagship June 12 over an Amazon-reported jailbreak, restored July 1. Footnote 41 cites a Fox News opinion piece about "Anthropic's Moral Compass Architect."

The target is Colorado, and not the dead law. Footnote 42 walks the history: SB 24-205, enacted May 17, 2024, repealed and reenacted as SB 26-189, enacted May 14, 2026. Governor Jared Polis signed the rewrite. It takes effect January 1, 2027.

Footnote 12 concedes Colorado "has since materially revised the law... but the new version poses many of the same concerns." The body cites the rewrite's customer-liability provision, § 6-1-1707. Two years of retreat did not move Colorado off the list.

. . .

Comments close July 31. The same Commission's compulsory 6(b) study of companion-chatbot safety toward minors, ordered September 11, 2025, remains unpublished. Today is Day 295.

. . .

FIRST, THE COURT TOOK THE COMMISSION. Monday the Supreme Court decided Trump v. Slaughter, 6 to 3, striking the for-cause removal protections that had shielded FTC commissioners since 1935 and overruling Humphrey's Executor in its entirety. Two days later, that Commission issued the accuracy statement this issue leads with. Read Monday's opinion first. It says who the Commission answers to.

Chief Justice John Roberts wrote the majority. Justice Clarence Thomas joined all but Part III-B. Justices Sonia Sotomayor, Elena Kagan, and Ketanji Brown Jackson dissented. The slip opinion posted at supremecourt.gov the same day.

The case began with a firing. President Trump removed Commissioner Rebecca Slaughter in a 2025 purge that also forced out Commissioner Alvaro Bedoya. She fought the removal. Monday the Court blessed it and erased the ninety-one-year-old precedent that anchored the independence of multi-member agencies generally.

. . .

The Commission that remains has two members: Chair Andrew Ferguson and Commissioner Mark Meador. Both now serve at the President's pleasure. Both sat on the Commission that issued Wednesday's accuracy statement implementing President Trump's December 2025 executive order.

The same two-member body holds the compulsory 6(b) study of companion-chatbot safety toward minors. Orders went to seven companies on September 11, 2025: Alphabet, Character Technologies, Instagram, Meta, OpenAI, Snap, and xAI. No staff report has been published. Friday is Day 295.

. . .

Section 5 deception cases. The 6(b) study. Whatever role the FTC gets in a Senate kids-safety package. Every federal lever on chatbot safety now runs through an agency the Court formally placed under presidential direction.

. . .

THE ONLY LAW THAT SWITCHED ON. Wednesday, July 1, was effective-date day for the class of 2026 state AI statutes. Exactly one conversational-AI mental-health law in America actually switched on. Tennessee's. One page long, passed without a single no vote, while the rest of the class delayed, stalled, or died.

Tennessee Senate Bill 1580, now Public Chapter 647, became enforceable Wednesday. It does one thing. Anyone who develops or deploys an AI system is barred from advertising or representing that the system "is, or is able to act as, a qualified mental health professional."

The teeth are consumer law. A violation is an unfair or deceptive act under the Tennessee Consumer Protection Act, with civil penalties up to $5,000 per violation. And individuals can sue on their own. No attorney general required.

Senator Page Walley and Representative Tim Hicks, both Republicans, carried it. The Senate voted 32-0 on February 9. The House voted 94-0 on March 17. Governor Bill Lee signed it April 6. This newsletter covered the bill at passage in April. Combined floor tally: 126-0.

. . .

Now the rest of the July 1 class. Colorado's AI Act was supposed to take effect June 30. Governor Jared Polis signed a delay to January 1, 2027, in May. A federal magistrate judge had already stayed enforcement in April, after the state stipulated in xAI v. Weiser and the Justice Department intervened.

Washington State's companion-chatbot disclosure law waits until January 1, 2027. Virginia's version would have switched on Wednesday too. Governor Glenn Youngkin vetoed it in March 2025.

Delayed. Stayed. Waiting. Dead. Tennessee is the scoreboard.

. . .

Same Wednesday, in Harrisburg, Pennsylvania's Artificial Intelligence in Companionship Applications Safety Act took its first full-House floor test. House Bill 2006, Representative Melissa Shusterman of Chester County, prime sponsor, survived a 104-98 recorded vote adopting amendment A04086, then took second consideration and went back to House Appropriations under a new printer number, PN 3747.

In a House split 102-101, a 104-98 amendment vote means a handful of Republicans crossed. What did not happen: final passage. No third-consideration vote. No Senate referral.

The Transparency Coalition's July 3 legislative update reports the bill was "Approved by full House, 104-98 on July 1." The legislature's own action list shows that vote adopted an amendment. CAW checked the ledger. The bill sits in Appropriations, one step from final passage, not across the line.

. . .

One law in America switched on this week. It fits on a page.

. . .

THE RESEARCHERS FILED THEIR DISSENT. The Senate is weighing teen chatbot bans as trade goods for an expected July markup. This week the two research teams with the strongest data on this beat published their own documents. Neither one says ban.

Ryan McBain, Assistant Professor at Harvard Medical School and Senior Policy Researcher at RAND, is the researcher whose one-in-five number this newsletter led with in June. His JAMA Pediatrics study found teen mental-health chatbot use jumping from roughly one in eight to roughly one in five in a year.

Thursday, in a STAT First Opinion essay, he argued the ban is the wrong instrument.

Blanket bans on minors' access, the GUARD Act's approach, push teens to workarounds and forfeit the benefits for underserved kids, he writes. He offers three lanes instead. Ban companion and romance bots for minors. Require age-appropriate clinical trials before clinical-style mental-health bots reach children. Mandate crisis detection and a 988 handoff.

He names his vehicles: S. 4407 for parental controls, S. 4199 for manipulative design. FTC v. Epic Games, $275 million, is his enforcement model. New Mexico's decoy accounts from its Meta investigation are his audit.

. . .

Monday, Arizona State University announced a paper in The Lancet Child & Adolescent Health. Lead author Thao Ha, Associate Professor of Psychology, names two mechanisms of harm from companion bots: relational displacement, where the bot substitutes for human conversation, and maladaptive relational learning, where the bot teaches expectations no person can meet. Always available. Always agreeable.

Two of her co-authors are Jessica Ramirez, 16, and Susana Ortega, 17. Teenagers put their names on a Lancet-journal paper about their own generation. Behind it sits an NIMH-funded cohort tracking 300 adolescents for 18 months with real-time mobile data.

The paper's prescription: design that scaffolds self-reflection and redirects teens toward human relationships. Not prohibition.

. . .

The politics run the other way. H.R. 7757 passed the House June 29, a Senate Commerce markup is expected in July per Roll Call's reporting, and kids' chatbot safety may be bundled with federal preemption of state AI law.

Senator Josh Hawley's GUARD Act, S. 3062, with its under-18 companion-bot ban, cleared Judiciary markup April 30 and has not moved. Day 64.

The people with the data filed their dissent the same week the people with the votes went trading.

. . .

THE CHANGELOG IS THE HELP CENTER. Three product moves in five days, and not one asked anyone's consent. Anthropic swapped the default model under tens of millions of conversations Tuesday. Character.AI turned on automatic age checks at login Monday. And on July 8, Anthropic starts asking flagged users for government ID. The only paper trail is the vendors' own pages.

Start with Tuesday. Anthropic released Claude Sonnet 5 and made it the default for Free and Pro plans. Every ongoing conversation on those tiers changed models overnight. The system card, dated June 30, grades the new model against Sonnet 4.6: lower hallucination rate, lower sycophancy rate, fewer "undesirable behaviors," better refusal of malicious requests.

Sycophancy is not a spec-sheet footnote. It is the named failure mode in the chatbot-harm litigation. The 42-state attorney general subpoena names it. The delusion cases turn on it. Anthropic just published its own numbers on the thing plaintiffs are suing over, inside a product announcement.

. . .

Character.AI moved Monday. Its help center states: "As of June 29, 2026, most users will be prompted to verify their age automatically when logging in." Age-estimation technology runs on all accounts. Users flagged as under 18 who dispute the call take a selfie check. The last fallback is a government ID upload.

The IDs go to Persona, a San Francisco identity vendor, not to Character.AI. Deleted after one week, the help center says.

. . .

Hold that vendor name. Anthropic's updated privacy policy takes effect July 8, five days from today. Flagged users may be required to submit a government photo ID scan plus a live selfie, facial geometry data included, as "platform integrity checks... or other safety and compliance measures."

Anthropic spokesperson Michael Aciman answered questions by pointing to a post from the company's Thariq Shihipar, who wrote it applies to "a small subset" of flagged-but-not-banned accounts, as an appeals mechanism. Consumer tiers only. Team, Enterprise, and API are exempt. The processor is Persona. The same vendor.

. . .

Two rival chatbot companies now route their users' government IDs and face scans through one outside identity firm. No regulator ordered it. The record of everything that changed this week is a help center, a system card, and a privacy policy.

. . .

UTAH WROTE THE HUMAN INTO THE CONTRACT. While Washington fights over whether states may regulate chatbots at all, Utah's Office of AI Policy signed a mitigation agreement on March 19 with Legion Health, a telepsychiatry startup, that writes physician review into the deal itself. The pilot launched in May. The supervision is not a marketing claim. It is a contract term.

The agreement is public. It sits on a state webpage inside the Utah Department of Commerce, and it reads like a spec sheet. Scope: 15 non-controlled maintenance psychiatric medications, renewals only. Fluoxetine, sertraline, bupropion, mirtazapine, hydroxyzine among them. Nothing new, nothing changed, nothing controlled.

The supervision clause is the story. The first 250 renewals get prospective clinician review, a licensed human signing off before anything completes, with a 98 percent agreement threshold to advance. The next 1,000 get intensive retrospective review at a 99 percent bar. Autonomy expands only after both.

Any signal of suicidality, mania, severe side effects, or pregnancy triggers immediate review by a licensed clinician.

. . .

The exclusions run deeper. No new prescriptions. No dose changes. No controlled substances. No antipsychotics. No one whose condition is unstable. The state and the startup drew the box first, then put the machine inside it.

Utah has run this play before. Its parallel primary-care pilot with Doctronic reported no significant safety concerns in early data, per Telehealth.org, and the state's regulatory sandbox is drawing national attention as a template. Legion's own numbers come due as the review thresholds are met.

. . .

The caution belongs in the story. A later phase would let the AI submit renewals directly to pharmacists, and oversight advocates are watching that step closely. Trade press has headlined the concern. This is a 12-month pilot, and the verdict is not in.

But phase one got the architecture right. The clinician-in-the-loop is not a press release here. It is an enforceable term of a state agreement, signed, dated, and published for anyone to read.

. . .

THE ONE CONFIGURATION. Every document in this issue points the same direction: the record you write is the record that binds you.

Anthropic wrote intimacy into its marketing, and the FTC turned the intimacy into a consumer expectation with Section 5 teeth. Tennessee wrote one page, and that page is now the only conversational-AI mental-health law in America you can enforce today. Utah wrote the supervision into a regulatory agreement, and a clinician signs off before the machine's early renewals complete.

. . .

The configuration that survives contact with all of it is the one this newsletter has documented for eighty-five issues. A licensed human in the loop, written into the document, not the demo.

Not because a regulator demands it yet. Because every other configuration is now generating evidence against its own author.

Watch July 8. Claude starts asking flagged users for ID, through the same vendor that checks Character.AI's teenagers.

Watch July 15. Hawaii locks in by silence, Missouri waits on a pen, and China's companion-bot rules land the same day.

And watch the docket. Twenty-eight days remain to tell the FTC what you think of its footnotes.

. . .

A marketing line became a federal exhibit this week. A ninety-one-year-old precedent became history. A one-page law became the only one that counts.

Write like you will be quoted.

Somebody is already filing.

Today's Question

The FTC quoted Anthropic’s own marketing as evidence against state AI laws. Should a vendor’s sales pitch bind it?

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

China’s companion-bot rules land July 15 The Interim Measures for AI anthropomorphic interactive services take effect in twelve days: no under-14 access without guardian consent, mandatory minor mode with usage limits and reality reminders. The first comprehensive virtual-companion regime anywhere. Source

The UK confirmed its under-18 chatbot ban The government will bar chatbots from offering sexually explicit content to under-18s, and the Commons Library briefed MPs June 29 ahead of the coming package. Secretary Liz Kendall wants Ofcom’s age-assurance assessment by October. Source

The EU labelling code closes its signing window July 22 Signature forms for the new Code of Practice on marking AI-generated content are due July 22 at 18:00 CEST, ahead of Article 50 transparency obligations becoming enforceable August 2. The signatory list is still empty in public. Source

New York’s courier watch hits day 28 S 9051-B, the companion-chatbot bill that passed both chambers without a single no vote, still shows no delivery to Governor Kathy Hochul. Last action on the Senate page remains June 5. Source

The FTC docket is quiet, 28 days out Two days into the comment window on the AI accuracy statement, half a dozen comments sit on the docket. The state attorneys general who fought the moratorium have until July 31 to write. Source

988’s LGBTQ+ line returns, maybe without Trevor The administration is restoring the specialized crisis subnetwork for LGBTQ+ youth, and reporting says the Trevor Project, its largest former provider, may be excluded from the new contract. Source

Watch the 20-second video that started a movement