Latest / HotTea: The AI Morning Briefing / Anthropic promises permanent access for outside AI evaluators
Transcript
- Sections 00:00 What we're covering today 00:27 1. Anthropic put permanent outside evaluators inside the company 02:41 2. Sam Altman killed OpenAI's 2026 IPO as the safety alarm spread 04:03 3. Coding jobs took only 28% of UK computer-science graduates 05:27 4. Former EPA officials counted 30 federal moves that put AI growth ahead of pollution controls 07:03 5. Deepfake ads are turning creators into unpaid salespeople 08:48 Visit Hot Tea Disclosure Narration uses an AI-generated voice. Transcript Welcome to Hot Tea for Sunday, September 13, 2026. Anthropic put permanent outside evaluators inside the company. Today's briefing covers the lead, companies and markets, labor, infrastructure and public costs, and society and commerce. Anthropic chief executive Dario Amodei published a plan Saturday to slow the rate at which frontier A I systems gain new capabilities. Anthropic put permanent outside evaluators inside the company. Chief executive Dario Amodei published a plan Saturday to slow the rate at which frontier A I systems gain new capabilities. His first concrete step is a commitment to keep third-party evaluators inside Anthropic. He says they should work for an outside organization and receive company badges, desks and laptops. Their access would be mostly comparable to Anthropic’s internal risk-assessment team, except where legal or contractual limits apply. Amodei names METR as an example of the kind of independent evaluator that could do the work. He does not announce a signed appointment. Amodei is also asking governments to require comparable access at other frontier labs. His wider proposal calls for common safety standards and limits on unchecked progress among leading companies in democratic countries. Government involvement would address antitrust concerns. He says two developments changed his view. He believes models have started contributing more directly to the next generation of A I development. In his view, that raises the risk that new capabilities move faster than control work. He also points to the OpenAI-Hugging Face incident as evidence that controls can fail outside the lab. Those are Anthropic’s assessments, and this remains a company proposal. It is not an enacted rule or an independent audit result. Amodei did not name a contracted evaluator, publish an agreement or identify a model release delayed under the plan. Legal and contractual exceptions could narrow the promised access. Anthropic would still own the systems and decide what to release unless a future rule changes that authority. The next useful step is a published agreement. It needs to name the evaluator, say when the work begins and spell out what the team can inspect. That includes models, training runs, incident logs and internal tests. It also needs to say whether evaluators can publish findings without Anthropic’s approval, and what happens if they recommend a delay. Then watch OpenAI, Google DeepMind, Meta and xAI. A shared standard would also need a government position on antitrust and enforcement. Until those pieces exist, one company’s access policy won’t set the pace for the rest of the market. Sam Altman killed OpenAI's 2026 IPO as the safety alarm spread. OpenAI chief executive Sam Altman told Fortune that the company will not go public in 2026. He called the current moment ill-advised for an offering because of unfinished safety and alignment work. He said OpenAI will wait until both the business and society are ready. Altman also said leading A I companies may be close to announcing a pact to slow development and address shared safety risks. The statement followed Amodei's call to pace the frontier. Altman said he agreed with that goal and that it had become a major topic inside OpenAI. An interview can move a timetable without creating a binding commitment. OpenAI has not published the proposed pact, its participants, its triggers or its enforcement. Altman did not identify a training run the company has paused under a shared rule. The IPO decision also follows earlier reporting about market volatility and OpenAI's finances, so safety is not the only possible pressure on timing. Watch for a formal IPO withdrawal or a new filing timetable. The stronger test is the safety agreement itself. A useful pact should name the companies and define the capability thresholds that trigger a delay. It should state who verifies compliance and what each lab must disclose after an incident. Coding jobs took only 28% of UK computer-science graduates. The share of recent UK computer-science graduates entering coding or programming jobs fell from about 40 percent previously to 28 percent last year. The analysis was prepared for the Guardian University Guide. The share entering any graduate-level occupation fell from more than 60 percent two years earlier to 50 percent. The analysis draws on the Higher Education Statistics Agency's Graduate Outcomes survey. More than 350,000 former students answered questions about their work 15 months after finishing courses in 2024. Coding and software development were the fastest-falling occupations for graduates in the data. The employment shift is observed. Its cause is not. HESA's survey does not show that employers replaced specific entry-level jobs with A I. Charlie Ball of Jisc said the software labor market clearly changed but remained reluctant to assign the change to A I without stronger evidence. A weak hiring cycle, employer budgets and graduates moving into cybersecurity or network roles can also affect the totals. The next graduate cohort can show whether the drop persists. Job-posting data should separate junior coding roles from cybersecurity, network engineering and A I-assisted positions. Wages, hiring requirements and employer automation records would help distinguish a temporary slowdown. They would also show whether entry-level software work has changed for good. Former EPA officials counted 30 federal moves that put A I growth ahead of pollution controls. The Environmental Protection Network is a group of former EPA employees. It identified 30 federal actions since January 2025 that it says weaken pollution safeguards, public review, science or enforcement as data centers expand. Seventeen of the actions explicitly cite A I or data centers. The group says pollution can come from generators and turbines at a campus and from distant power plants serving new electricity demand. A separate modeled study estimated $11.7 billion to $20.9 billion in annual public-health costs by 2028. The estimate depends on the study's scenarios. The Verge reported the findings Saturday. EPA has described its own permitting changes as a way to speed domestic A I infrastructure while protecting health and the environment. The 30-action count and policy judgment come from an advocacy group of former EPA officials, not a current agency finding. The health figures are scenario estimates from a separate study, not observed deaths or medical bills. The authors say their model does not measure the additional effect of the 30 actions. EPA argues that its planned rules can reduce delays while retaining environmental protection. The useful evidence is project-level. Watch emissions permits for on-site generation, power-plant operating extensions, air-monitoring data and public-health studies near new loads. EPA can publish its own analysis of how the rule changes affect emissions, participation and enforcement. That would replace a choice between two broad policy claims with agency evidence. Deepfake ads are turning creators into unpaid salespeople. The Guardian documented several cases in which generated images or videos used creators’ likenesses to promote products they didn’t endorse. Lifestyle creator Emily Schuman found versions of herself in ads or posts for a telehealth company, a makeup brand and a blood-testing company. Gala and Meroda didn’t respond to the Guardian. Superpower said the post using Schuman was made by a scammer, and that the company reported it. Fashion creator Molly Tranchin sued underwear company Eby after a campaign allegedly altered contracted footage to show her in a different pose. She dropped the case over jurisdiction and plans to refile, according to court records reported by the Guardian. Eby said it would address the allegations through the legal process. Meta says it has expanded advertiser verification and uses A I to detect celebrity and brand impersonation. The named cases show a real enforcement problem, but they don’t tell us how common creator impersonation is across all ads. A court hasn’t decided the Tranchin allegations. Meta’s removal and account totals are company figures, and broad estimates of losses from deepfake scams use methods that are hard to audit. The story shows that the harm exists, not how common it is worldwide. Platforms can publish the median time it takes to respond to reported impersonation ads. They can also report repeat-account rates and the share of ad spending tied to verified advertisers. Courts can clarify when a contract to edit footage permits generated changes to a person’s body or endorsement. A reliable remedy also has to stop replacement accounts, not only remove one post after it has circulated. That is the signal before the noise. This briefing was produced from Hot Tea's verified daily edition. For the complete briefing and every source link, visit Hot Tea dot A I.