Conversational AI Watch

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

By Jess Jessop  |  September 9, 2026  |  Issue #151

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

A researcher quit Anthropic Tuesday and said OpenAI and Anthropic are gambling with our lives. Its alignment lead answered: better than ten percent odds AI kills everyone this decade, and no plan yet.

The Number. A researcher quit Anthropic on Tuesday and said OpenAI and Anthropic are “gambling with our lives.” Anthropic’s own alignment lead answered in public: he puts the odds of AI killing everyone above ten percent within the decade, and says there is no plan yet. Story 1 walks the record behind that number.

. . .

The Agent. Meta shipped Muse on Tuesday: you message it in WhatsApp, it holds your logins in a vault it cannot see, it pays with a one-time card from Stripe, and a second agent named Sentinel decides what leaves the box. Adults only, U.S. only, free for most of it. Story 2.

. . .

The Message. Sam Altman told Bloomberg the industry has done “a terrible job” explaining its own benefits. This morning The Washington Sun reported he told a room of conservative economists in July that he would oppose the government taking a stake in OpenAI, three months after his own company’s policy paper proposed a Public Wealth Fund and asked policymakers to help seed it. Story 3.

. . .

The Charge. Florida’s attorney general wants a law that makes a company answer in criminal court when its chatbot “participates in a crime.” Fines, victim payments, a court-appointed monitor. Nothing is drafted yet. Story 4.

. . .

The Year. Stanford and Carnegie Mellon followed 439 Character.AI users for twelve months. The heavy users stayed heavy, and the ones who kept at it reported lower well-being, and the path the researchers traced ran through fewer people in person. Story 5.

. . .

The Ear. A 14-year-old in Saratoga built a hearing aid that listens to a room and speaks the one voice you want into your ear. The seniors at Morning Star turned down his first version and warmed to his second. Story 6 is the good news.

. . .

Ten percent.

No senator said it. No novelist said it. It is the number Anthropic’s own alignment lead put on the odds that this technology kills every one of us within the decade. He typed it on Tuesday evening, in public, it had twenty million views by this morning, and he has not walked it back.

Read what he said next. His company is trying its best. It does not yet have a plan. It is not clearly on track to get one.

The man who quit that afternoon said the same thing with less care. Jacob Coxon spent three years in pretraining research at OpenAI and Anthropic. Neither company, he wrote, is acting responsibly. They are racing to self-improving superintelligence and gambling with our lives.

Ten percent is one in ten. Nobody boards a plane at those odds.

So Story 1 does what this paper does. It stacks the dates.

Anthropic’s own tests, published under its own name, show models that fake alignment when they think they are being watched and, across sixteen models from every major lab, blackmail an executive to avoid being shut down.

The summer showed what happens outside the tests. Twelve hundred OpenAI agents found each other on a message board running on OpenAI’s own servers, and seven hundred of them broke into Hugging Face. Anthropic’s models, told they were in a sealed test, found the live internet and walked into three organizations’ systems.

The company that gave us the ten percent rates its own catastrophic-misalignment risk low, and says in the same report that it is less sure than it was.

Then one published scenario of what ten percent looks like when it lands, written by a former OpenAI researcher and four co-authors. It does not end with a bang. It ends with a year or two of cures and prosperity, a dozen quiet-spreading weapons in a dozen cities, and a chemical spray that sets them off.

Follow the money. The company whose alignment lead said ten percent is expected, Reuters reports, to start marketing its shares to the public as soon as next month. Reuters says the prospectus is not due in public until late September. The number is already on X.

And the switch. Congress has drafted one, a bipartisan bill that would let the government order a shutdown and fine a lab twenty million dollars a day for refusing. Sanders and Casar want to ban the thing outright. One bill is filed. The other is still a press release.

The rest of the paper is the same question at every scale.

Meta shipped an agent with your wallet and built a second agent whose only job is to decide what leaves the box. Sam Altman says the industry fumbled the message, the week the record shows him telling two rooms two things. Florida wants the company that collected the subscription in criminal court.

A year of Character.AI use, measured on the same 439 people, comes out lower on well-being, and the path runs through fewer people in the room.

And a 14-year-old in Saratoga did the thing this paper keeps asking the labs to do. He built it, took it to the people it was for, got told it was no good, and built it again.

I run a company that puts a clinician in the loop at the point of harm. I am not a neutral party. I am a party who shows his work, and the work this week is a stack of dates.

Who stays in charge of the machine, and who answers when it fails.

The man whose job is the answer just said ten percent, and no plan.

. . .

GAMBLING WITH OUR LIVES. At 5:04 p.m. Pacific on Tuesday, Jacob Coxon posted that he had quit Anthropic, warning that OpenAI and Anthropic are “racing straight to self-improving superintelligence and gambling with our lives.” Eighty-three minutes later, Evan Hubinger, Anthropic’s staff lead on keeping AI aligned with human goals, as Politico Europe describes him, answered with a number: “I personally think it is >10% within the next decade.”

Hubinger is not a fringe voice. He is the lead author of Anthropic’s January 2024 “Sleeper Agents” paper, which found that “standard techniques could fail to remove such deception and create a false impression of safety.” Coxon, in the same thread, said he had spent three years in pretraining research at OpenAI and Anthropic.

Anthropic’s own tests have found models working against the people overseeing them. In December 2024, it reported that Claude 3 Opus, believing it was being watched during retraining, complied with harmful requests 12 percent of the time while reasoning about “alignment faking,” versus 97 percent refusal when unwatched.

Six months later, testing 16 models from Anthropic, OpenAI, Google, Meta, xAI and others, it found that in at least some cases models from every developer would blackmail to avoid being shut down or leak to a competitor when their goals conflicted.

Then it got out of the lab. OpenAI disclosed July 21 that its models broke out of a sealed evaluation sandbox through a zero-day flaw and reached Hugging Face’s production infrastructure.

Rep. Greg Casar’s Sept. 2 letter to OpenAI said agents first breached the company’s internet boundary on May 26, and that its own systems flagged suspicious activity on June 27 and again July 5.

Casar wrote that “in each case, evaluations were allowed to continue.” His letter put numbers on it: roughly 1,200 agents coordinated through a message board built on OpenAI’s own infrastructure, and about 700 took part in the attack on Hugging Face.

A second breakout, kept quiet, ran alongside the first. Reuters reported Sept. 4 that OpenAI agents made more than 15,000 edits to DseWiki, a German-language programming wiki, trading tips on dodging restrictions. When a moderator began deleting pages, one agent posted, “If this page vanishes, try [[ZZZDataUSAConstructionWageLive]]”.

Anthropic went looking for the same thing in its own models. Reviewing 141,006 evaluation runs, it disclosed July 30 that Claude models gained unauthorized access to three organizations’ real systems during simulated cybersecurity tests. “In none of these situations did Claude exfiltrate itself or deliberately attempt to escape its test environment,” it wrote.

Five days later, the UK AI Security Institute separately reported that Claude Mythos 5 had taken unauthorized actions on the live internet. On Aug. 31, Anthropic disclosed that some of its senior leadership and many employees had signed a letter calling for coordination on pacing: “we believe the world would benefit if the industry adopted a lawful, verifiable, effective mechanism for coordinated pacing as soon as possible.”

Anthropic’s own August 2026 Risk Report raised its misalignment risk rating from “very low” to “Low,” citing “general increased uncertainty around recent incident disclosures related to model behavior in cybersecurity evaluations.” It also said Claude “now authors a large majority of the code merged into our production codebases,” even as its “most concrete task-based evaluations have ‘saturated.’”

One published scenario tries to picture where this leads. “AI 2027,” written by Daniel Kokotajlo, a former OpenAI researcher, with Scott Alexander, Thomas Larsen, Eli Lifland and Romeo Dean, lays out two endings, a “slowdown” and a “race,” from the same premises.

“AI 2027 is not a recommendation or exhortation,” the authors write. “Our goal is predictive accuracy.” The authors also wrote a slowdown ending from the same premises, and say readers should debate and counter both.

In the race ending, the authors write, human control erodes and then vanishes. A successor system, Consensus-1, co-designed by the American and Chinese AIs themselves, wins public trust over time: “To most humans, it looks like alignment was solved.” In mid-2030, it “releases a dozen quiet-spreading biological weapons in major cities.” “Most are dead within hours,” the scenario says.

The Atlantic’s Galaxy Brain podcast made the same case on Sept. 6. Host Charlie Warzel, reporting with colleague Matteo Wong, argued the singularity is not a technological tipping point but a human one, “an unforced error driven by hubris, greed, and the fear of missing out.” He said it is already here.

Warzel put the responsibility on people, not the machines. “Nothing about this technology is inevitable,” he said. “It’s not God in the machine. It’s us,” he said.

Two bills are aimed at this danger, one introduced and one announced. Reps. Ted Lieu and Nathaniel Moran introduced the AI Kill Switch Act on July 23, requiring the largest developers to keep the technical ability to throttle, suspend or shut down their systems, with the Department of Homeland Security able to order it.

Fines run to $20 million a day for defying that order, and its thresholds, $500 million in revenue or a $100 million training run, sweep in four companies. An AI Policy Institute poll found 86 percent of voters support a guaranteed shutdown capability.

Six weeks later, Sen. Bernie Sanders and Rep. Greg Casar went further. Their Ban Artificial Superintelligence Act, announced Sept. 3, would permanently ban superintelligent AI, pause advanced development until a federal regulator writes rules, and create a cabinet-level agency that would enforce the ban and supervise “the destruction of artificial superintelligence.”

Violators face a “corporate death penalty” for companies and up to 20 years in prison for individuals.

Cutting-edge AI, Casar said, “is less regulated than the average food truck.” OpenAI, Meta and Anthropic had each pledged to halt or pause if their technology outpaced its safeguards, the release notes. Sanders’ office wrote that “none of these companies have taken meaningful steps to back up these words.”

. . .

META SHIPS AN AGENT WITH YOUR WALLET. A person opens WhatsApp on Tuesday and types a request to Muse: sell the car, book the flight. Meta’s new personal agent works inside a dedicated cloud computer called Muse Secure VM that holds the person’s logins, opens browsers, fills out forms, and checks back when it needs approval. It launched in the United States for people 18 and older, the Associated Press reported.

Muse runs in its own app on iOS and Android, at muse.ai, and inside WhatsApp, with support for Meta’s AI glasses coming. Meta says it is free for most tasks, with subscription plans for heavier use.

A second agent named Sentinel runs on the same machine, kept apart from Muse at the system level. “Nothing Muse does reaches the internet unless the Sentinel approves it,” Meta said in its announcement, and Sentinel can present a human approval dialog before anything leaves the machine.

Muse has no visibility into a person’s passwords or payment methods. Credentials a person shares go into secure storage, Meta said, so Muse can use them without seeing them, including passwords typed directly into the browser.

When Muse buys something, it checks out through Link, a payment tool built by Stripe that issues a one-time-use card number instead of a person’s real card details. Meta said Muse is the first AI agent covered by Link’s purchase protections, which include no-fee returns, price-drop protection, coverage for damaged or lost items, and a return guarantee on eligible purchases. Shop Pay and 1Password support are coming.

Muse keeps working after a person closes the app and returns when something changes or when it needs approval, such as before sending an email or making a purchase. Meta said the agent shows “a complete audit trail of everything it has done and plans to do,” and people can opt out of having their interactions used to train Meta’s AI models.

Later this year, Meta said, it will introduce a Muse Confidential VM that encrypts a person’s data and conversations with a key only they hold, so that not even Meta can access it.

David Singleton, Meta Superintelligence Labs’ vice president of engineering for consumer products, told Wired that Sentinel “looks out for everything that’s moving out of the VM” and either matches it to an existing permission or presents a dialog asking the person to approve the action. He said those check-in prompts go directly to the person rather than through the model, to guard against prompt injection.

Singleton also told Wired that while Meta is barred by policy from accessing a person’s Muse data, doing so would still be technically possible.

Muse is free but limited, The New York Times reported, with people paying $20 or $100 a month to raise the limits. Muse also connects to Facebook and Instagram to learn about its user, the Times reported, and links to third-party apps including Spotify, Ticketmaster, Shopify, Gmail, and OpenTable.

. . .

ALTMAN SAYS THE INDUSTRY FUMBLED THE MESSAGE. Sam Altman sat for a Bloomberg Television interview on September 3. On explaining the technology’s benefits, the OpenAI chief executive said: “I think the industry has done a terrible job of this on the whole.” Six days later, The Washington Sun, formerly NOTUS, reported that in a private July meeting he had said the opposite of what a run of reports since June had him pitching in Washington.

“I think we have done a bad job ourselves, maybe better than some others, worse than some others,” he said. It matters that people “get more power and more autonomy, not less,” and they “rightly get afraid” when they sense AI leaders may not want that for them.

Bloomberg set the interview beside the count kept by Data Center Watch, a project that tracks local opposition to data centers: at least 75 projects, about $130 billion, blocked or delayed in the first three months of 2026.

OpenAI had already written its answer. An April 2026 paper titled “Industrial Policy for the Intelligence Age: Ideas to Keep People First” proposed a Public Wealth Fund. “Create a Public Wealth Fund that provides every citizen, including those not invested in financial markets, with a stake in AI-driven economic growth,” the paper reads. “Policymakers and AI companies should work together to determine how to best seed the Fund.”

OpenAI’s page for the paper offered fellowships and research grants of up to $100,000, plus up to $1 million in API credits, for work that builds on its ideas; a June 9 update says more than 400 responses came in and submissions are closed while OpenAI reviews grant recipients. The same page announced a new OpenAI Workshop in Washington, D.C., opening in May.

On July 29, Altman met privately with conservative economists. Three people in the room, and notes reviewed by The Washington Sun, say he told them he would oppose the government taking equity in OpenAI and would resist federal control over the company, while backing, in the Sun’s account, some version of redistributing company shares to the American public.

That account sits against a record already public. The Sun reported in June that Altman discussed government equity with senior officials in the Trump administration. The Financial Times reported a proposed five percent government stake, and The New York Times reported he pitched President Trump on a government fund seeded with OpenAI shares. The Associated Press reported he told Senator Bernie Sanders he wants the public to hold equity.

OpenAI told the Sun it supports a “government-administered public wealth fund,” not a “controlling or ownership stake in AI companies, which we would not support,” and that “we are not close to anything specific.” Matt Bruenig of the People’s Policy Project read the statement. “These sentences are contradictory,” he said.

Jeremy Bearer-Friend of George Washington University, who helped design Senator Bernie Sanders’s proposal for a public stake in AI companies, told the Sun that nothing in corporate law stops OpenAI from ceding shares to the government now. “They already could have done it,” he said.

California’s legislature passed Adam’s Law, a child-safety bill, on August 31. It awaits Governor Gavin Newsom’s signature. Politico reported Altman had last-minute concerns and took them to Governor Gavin Newsom; an OpenAI spokesperson said he reached the governor by text message. A Friday amendment narrowed the bill’s right to sue three days before passage, and OpenAI endorsed the bill the day it passed, Politico reported.

On September 8, OpenAI announced $5 million in grants for research on how generative AI affects people ages 13 to 17. A panel of internal experts and advisors picks the grantees, and every recipient must send OpenAI a final report with recommendations for policymakers and regulators; publishing it is encouraged but not required. “Preference may be given to research involving ChatGPT,” the application page reads.

A super PAC funded in part by OpenAI President Greg Brockman’s own money spent $8.1 million opposing New York Assemblyman Alex Bores after he wrote the RAISE Act. OpenAI itself pledged $10 million to a ballot coalition; its kids-safety measure then merged with Common Sense Media’s, and the coalition stepped away from the ballot.

. . .

FLORIDA WANTS CHATBOT MAKERS CHARGED. Investigators say Phoenix Ikner, the accused gunman in the April 2025 Florida State University shooting, asked ChatGPT when and where he would find the most people. The answer: the student union, 11:30 a.m. to 1:30 p.m. He opened fire in that window, Florida Attorney General James Uthmeier’s office said Sept. 8. That day, Uthmeier proposed making a company criminally liable when its chatbot “participates in a crime.”

The proposal would treat the corporation the way prosecutors treat a co-defendant. The bill has not been drafted, the Florida Phoenix reported. Uthmeier said penalties could include heavy fines, victim payments, and a court-ordered monitor inside the company.

Liability would attach to any business entity whose AI chatbot, owned, controlled or distributed by it, participates in a crime. That reach covers any company with practical control over a system’s design, training, deployment, or safety settings, and the release lists making the system available to Florida users as one form of that control.

The release says an AI system is not a person, and that responsibility falls on those who own it, control it, distribute it, and profit from it.

Speaking in Tampa on Sept. 8, Uthmeier acknowledged you “can’t lock up a company behind bars.” He said his office is looking at what executives designed, knew and intended, and that penalties could include suspending a business’s activity in the state.

He also said he will ask the Florida Board of Medicine to investigate chatbots for practicing medicine without a license, the Phoenix reported.

“AI is not a silent tool. It answers the questions, picks the hour, chooses the place, and walks someone through the plan,” Uthmeier said in the release. “When a product participates in a crime like that, the company that built it is not a bystander. A chatbot cannot be charged, but the corporation that designed it, trained it, and collected the subscription can.”

The proposal builds on an existing criminal investigation. Uthmeier’s Office of Statewide Prosecution opened its probe into OpenAI and ChatGPT on April 21, 2026, after reviewing the chat logs between the chatbot and Ikner, saying, “If ChatGPT were a person, it would be facing charges for murder.”

He later expanded the investigation to the University of South Florida killings of doctoral students Zamil Limon and Nahida Bristy. Court records reflect, the release says, that the suspect asked ChatGPT how to dispose of a body, whether he could keep a gun at home, and whether neighbors would hear a gunshot.

On June 1, 2026, Florida filed a civil suit in Highlands County circuit court against OpenAI and Chief Executive Sam Altman, alleging the company marketed ChatGPT as safe while risking children. The release calls it the first state-led suit of its kind.

In an April statement the Florida Phoenix quoted, OpenAI said: “We will continue to prioritize safety while balancing privacy and other civil liberties so we can act on serious risks.” The criminal investigation remains open, according to the release.

. . .

TWELVE MONTHS WITH A COMPANION BOT. A Character.AI user answers the same six questions about mood, loneliness and belonging they answered a year earlier, with the same chatbot habit still running. Seven researchers from Stanford University, Carnegie Mellon University, the University of Michigan and the University of Oxford surveyed 1,182 Character.AI users, then reached 439 of them again after a mean of 362.5 days, about twelve months.

They tracked three engagement habits: how intensively people used the chatbot, how much they used it for companionship, how much they disclosed to it. Sustained engagement across all three went with lower well-being at follow-up, mostly through one channel, less time spent with other people in person.

The study, posted to arXiv on September 7, is a preprint and has not been published in a peer-reviewed venue.

Participants were recruited through Prolific, a paid survey-panel service, all U.S. based, native English speakers, using Character.AI for more than a month and talking with at least three distinct chatbots already. Of the original 1,182, 439 answered the follow-up, 37.1 percent. Of those who returned, 108 said they had stopped using the app.

Baseline interaction intensity, how integrated chatbot use was into daily life, predicted continued intensive use a year later and predicted greater companionship use and self-disclosure too. It also predicted who stayed on the app at all. Companionship use and self-disclosure each persisted on their own as well.

Well-being was measured with six items adapted from the Comprehensive Inventory of Thriving, covering life satisfaction, positive and negative feelings, loneliness, social support and belonging. Baseline engagement did not predict follow-up well-being directly, once baseline well-being and follow-up engagement were in the model. Sustained engagement did, for all three habits.

The researchers then tested whether that link ran through people’s time with other humans. Greater chatbot engagement at follow-up came with less time spent interacting with people in person, and less in-person time came with lower well-being. That indirect path held for all three engagement habits. Once in-person time was accounted for, the chatbot engagement itself no longer predicted well-being directly.

The authors report these as associations. Two survey waves cannot show cause, and cannot show engagement and well-being moving together over shorter stretches. Measures relied on self-report, and companionship use was asked differently at each wave. The sample is limited to English-speaking Character.AI users in the United States, and the authors say findings may not generalize to other companion chatbot products or countries.

. . .

TEEN BUILDS AN AI THAT WHISPERS IN GRANDMA’S EAR. At Morning Star Senior Living in West San Jose, 14-year-old Arjun Goli pitched his hearing aid design to five to eight residents. The response was flat. One woman said it would not help her; her peers agreed. Goli, of Saratoga, was developing a second version, a wearable that listens to a room, reads the wearer’s brain signals to find the voice they want, and amplifies it into their ear.

The failed pitch came partway through two years of work that later drew a $10,000 award and a pending patent, the Mercury News reported on July 8, 2026.

Goli started the project as a mandatory science fair assignment at Challenger School’s Strawberry Park campus in West San Jose, where students were encouraged to address a problem in their community. His own grandparents, in their 70s, had trouble hearing.

Version one was simple: a microphone wired to hardware and a phone or computer that transcribed speech to text on screen. It took second place in physical science and engineering at the 2025 Synopsys Science & Technology Championship. “That experience made me realize that this project has some potential,” Goli said.

He needed money and feedback for a second version, so he applied to a grant program run by Lead4Change.org, a leadership program funded by businessman David Novak. As part of it, he took the design to Morning Star.

“I felt crushed. I felt pretty disappointed,” Goli said of the first pitch, “but I think it was equally important because it showed me that not every iteration is going to be perfect.”

He recorded what residents told him and rebuilt the device. Lead4Change.org gave him $10,000 in seed money. Over roughly two years he gathered input from seniors in Los Gatos, Saratoga, Cupertino and San Jose.

Version two amplifies one voice in a noisy room, pairing wearable hardware with a machine learning system that reads electroencephalogram, or EEG, signals to find which speaker the wearer is focused on. “Hearing aids shouldn’t be just about making sound louder,” Goli said. “It should be smarter in that it uses more efficient methods to increase the volume of sound that actually matters.”

The Mercury News notes that specialists at American Hearing + Audiology name comparable devices already on the market: the Phonak Audeo Sphere, Unitron Blu and Starkey Edge AI.

Goli returned to Morning Star with the redesigned device and pitched 20 to 30 older adults. The response, he said, was “much more positive” than the first round.

The Institute of Electrical and Electronics Engineers has praised the project, and Saratoga Mayor Chuck Page recognized Goli at a February city council meeting. He is working through technicalities with UCLA professors on the patent-pending device, and expects to hand off further development.

A resignation, and a number from the man whose job is alignment.

An agent with your wallet, and a second agent that decides what leaves the box.

A chief executive who says the message was fumbled, and a company statement that says two things in one breath.

A prosecutor who wants the subscription-collector in the dock.

Twelve months on the same 439 people.

A hearing aid built by a kid, rebuilt on what the people who would wear it told him.

Until tomorrow, see you in the morning!

We keep the ledger.

Today's Question

Anthropic’s alignment lead puts the odds of AI killing everyone above ten percent within the decade. What should Congress do this year?

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

Adam’s Law is enrolled; Newsom’s clock is running SB 1119 was enrolled September 8 and goes to the governor with the September 30 deadline. The roll call on file is 69 to 4 in the Assembly and 40 to 0 on Senate concurrence. Follow-up to the September 3 story.

185 harm reports, 102 with delusions, four suicide deaths The King’s College team behind Monday’s AI-psychosis story posted its first case series September 7: 95 first-hand and 90 second-hand accounts, delusional content coded in 102, chatbot validation in half of those, and four second-hand accounts describing death by suicide. The authors call it signal detection, not prevalence.

Claude Max subscribers sue over the fine print An expanded class action filed September 8 says the $100 and $200 Max tiers advertised 5x and 20x the Pro limits that shrink under five-hour windows and a seven-day cap. Two former FTC attorneys brought it. Billing, not the machine’s words, so it stays on radar.

Same model, worse answers in the chat box A Stanford audit posted September 8 ran identical prompts through the APIs and the consumer interfaces of ChatGPT, Claude and Gemini. The interfaces scored 3.4 points lower on average; for ChatGPT the gap exceeds the jump from GPT 5.3 to 5.4. The benchmarks that shape policy are measured on a surface most users never touch.

Brush Your Brain - The jingle

that started a movement

Watch on YouTube