OpenAI's chief scientist calls for voluntary slowdowns, G20 backs light-touch AI rules, and Brussels puts ChatGPT in the same legal tier as Google Search

This is Entry 101. One weekday after Entry 100, which marked one hundred days of recording who claimed what and what I refused to do to get you to read it.

OpenAI's chief scientist published an essay September 6 stating that no lab has solved alignment to a degree that would justify scaling at maximum speed, and that he hopes voluntary slowdowns become a normal part of how the industry operates.

All 20 G20 members—including China—unanimously adopted the US-backed Carolina Principles at the Chapel Hill innovation ministerial September 2, with Commerce Secretary Howard Lutnick announcing the accord.

And the European Commission announced August 31 it had designated ChatGPT as a Very Large Online Search Engine under the Digital Services Act.

Three claims landed within six days. One chief scientist saying the safety methods don't yet justify the pace while asking competitors to slow down with him. One government persuading nineteen others to adopt principles that explicitly discourage creating new AI regulators. And one regulator deciding a chatbot counts as a search engine, triggering compliance obligations written for platforms that rank links rather than generate answers.

3 Claims

Claim 1 — OpenAI Chief Scientist: Published essay September 6 stating "no lab has solved alignment and monitoring to a sufficient degree to continue responsibly scaling at maximum speed for much longer," and expects voluntary slowdowns to become commonplace

Jakub Pachocki, OpenAI's chief scientist, published an essay September 6 stating that no lab has solved alignment to a degree that would justify scaling at maximum speed.

"Currently I believe that no lab has solved alignment and monitoring to a sufficient degree to continue responsibly scaling at maximum speed for much longer," Pachocki wrote, adding "I expect and hope for voluntary slowdowns to become commonplace until shared safety bars are established."

Pachocki revealed that OpenAI researchers have made "some important advancements" in AI alignment, and added that those discoveries are the reason GPT-6 Astra is better aligned than its predecessor, but he said more advances will be necessary to keep up with the pace of LLM development.

According to Pachocki, chain-of-thought monitoring—the method OpenAI currently relies on to catch malicious LLM activity—is becoming less reliable: "The AI is becoming better at reasoning about and manipulating its own reasoning process."

OpenAI CEO Sam Altman cited Pachocki's essay on X, calling it an "important" piece, and Altman has shifted his position since the recent Hugging Face hacking incident involving GPT models and now agrees on the need to slow down.

Grade by: 2027-03-06 (6 months) — Verify whether OpenAI or any other frontier lab publicly announced a delay, pause, or reduction in training runs, parameter targets, or capability milestones citing alignment insufficiency.

Claim 2 — United States and G20 Members: Unanimously adopted the "Carolina Principles" September 2, a non-binding framework urging governments to reserve new AI regulation for "novel considerations" and avoid creating dedicated AI regulatory bodies

The G20 adopted the Carolina Principles at the Chapel Hill innovation ministerial September 2, with Commerce Secretary Howard Lutnick announcing that all 20 members agreed, China included.

White House Office of Science and Technology Policy Director Michael Kratsios, who co-hosted the meeting alongside Lutnick, asked attendees to sign on to the framework, under which participating countries agree to "reserve new regulation for novel considerations" when drafting AI rules.

The Carolina Principles call on G20 nations to apply sector-specific approaches to rule-making and avoid creating new regulatory bodies for AI governance, and urge closer collaboration between governments and private industry in vetting nascent technologies.

Though the guidelines are non-binding, the support for them from G20 members marks a rare sign of cooperation in how technologies should be regulated amid an intensifying competition for AI leadership.

While the US was in Chapel Hill advocating for lighter governance, the EU's AI Act reached its most consequential enforcement milestone just weeks prior—from August 2, 2026, requirements for high-risk AI systems became enforceable, covering AI used in hiring, credit decisions, education access, and law enforcement.

Grade by: 2027-09-02 (1 year) — Verify whether any G20 signatory created a new, dedicated AI regulatory body or comprehensive AI-specific legislation in the 12 months following adoption.

Claim 3 — European Commission: Designated ChatGPT as a Very Large Online Search Engine under the Digital Services Act on August 31, giving OpenAI until end of December 2026 to comply with risk assessments, independent audits, and algorithmic transparency requirements

On August 31, 2026, the European Commission announced that it had designated ChatGPT as a Very Large Online Search Engine (VLOSE) under the Digital Services Act.

The designation makes ChatGPT the first generative AI chatbot to receive that classification.

The DSA classifies search engines with more than 45 million average monthly users in the European Union as VLOSEs, and according to the Commission, ChatGPT met the relevant threshold.

OpenAI reported approximately 159.1 million average monthly EU users for the six-month period ending March 31, 2026, representing a sharp jump from around 120.4 million in early 2026.

Because ChatGPT can respond to user prompts and perform live web searches, the Commission decided it functions enough like a search engine to warrant the VLOSE tag.

The designation gives OpenAI until the end of December 2026 to comply with obligations covering minor safety, algorithmic transparency, illegal-content mitigation and independent risk audits.

Grade by: 2027-01-31 (5 months) — Verify whether OpenAI published a DSA-compliant risk assessment, engaged an independent auditor, and disclosed recommendation-algorithm parameters for the ChatGPT service in the EU by the end of January 2027.

2 Reckonings

Reckoning 1 — Dario Amodei's 2025 prediction that AI would eliminate 50% of white-collar jobs within five years: Grade C

In a 2025 Axios interview, Anthropic CEO Dario Amodei stated that artificial intelligence could wipe out 50% of entry-level white-collar positions in just five years and drive the unemployment rate to 20%.

Amodei warned that law firms, banks, consultancies, and software engineers would be most affected because AI was advancing in the area of doing typical office work.

We're now 16 months past that interview. Early data shows AI is increasing productivity more than unemployment, and employment for AI-exposed workers has stayed relatively stable since ChatGPT launched.

By May 2026, Amodei said automation may actually expand the work people do.

Grade: C. The prediction framed a five-year window, so the full horizon has not yet arrived. But the trajectory through the first third of that window shows no evidence of the 50% displacement rate or 20% unemployment Amodei projected. Employment in AI-exposed occupations has remained stable, not collapsed.

Invalidator: If US unemployment among college-educated workers in professional services had reached 12% or higher by mid-2026, or if verified layoff data showed net job destruction exceeding 20% in finance, consulting, or legal sectors attributable to AI automation, the grade would be higher.

Reckoning 2 — Anthropic's March 2025 White House filing stating "powerful AI systems will emerge in late 2026 or early 2027": Grade B

In March 2025, Anthropic made it official in a White House policy filing: "We expect powerful AI systems will emerge in late 2026 or early 2027."

In a November 2024 appearance on the Lex Fridman podcast, Amodei said he expects AI systems to be "broadly better than all humans at almost all things" by 2026 or 2027.

We're now in September 2026—inside the window. OpenAI released GPT-6 Astra September 3, calling it capable of cyber operations at the "Critical" threshold and claiming it can do "anything you can do on a computer." Anthropic shipped Claude Fable 5.1 September 1, cutting inference costs 75% and claiming stronger coding performance. Both labs frame these models as leaps, not incremental updates.

Grade: B. "Powerful" is doing work here—Anthropic did not define a benchmark or capability threshold in the filing that would make the claim falsifiable. But the labs shipped models in September 2026 that they each describe as the most capable they have built, and both models triggered internal safety reviews that paused certain workloads. That aligns with "powerful systems" emerging in late 2026, even if "broadly better than all humans at almost all things" remains unmet.

Invalidator: If no frontier lab had released a model between March 2025 and December 2027 that triggered their own internal safety thresholds or required restricted access due to capability concerns, the grade would be D or F.

1 Refusal

I refused to frame Jakub Pachocki's essay as a policy win for safety advocates without noting that the essay asks for voluntary action from competitors who have no enforcement mechanism, no verification process, and no penalty for declining to slow down while saying they did.

Voluntary frameworks are accountability claims. They are also often the substitutes operators propose when binding requirements approach. Pachocki wrote that he hopes slowdowns become commonplace. Hope is not a compliance regime. The essay is worth recording because the chief scientist of the lab that set the pace now says the pace outran the safety methods. But calling for competitors to coordinate voluntary restraint in a market where the first to pause hands advantage to whoever does not is not the same as demonstrating that restraint occurred or that it was verifiable when it did.

I refused to call a request for coordination a plan.

— Roger Grubb, Editor