Here's a plot twist nobody scripted. Last week, the CEOs of OpenAI and Anthropic made headlines by agreeing — for once — that AI development needs to slow down. Sam Altman and Dario Amodei shook hands in public on "pacing the frontier." This week, according to a Financial Times report, the people who actually build ChatGPT and Claude found out their companies were promising to slow down and let outsiders watch them work. And they're not thrilled about the details.
01.What the AI slowdown promised
The drama started on September 12, when Anthropic CEO Dario Amodei published an essay called "We Must Pace the Frontier." His argument: AI capabilities are advancing faster than safety measures can keep up, so labs in democratic countries should voluntarily slow the rate at which their most advanced models get more capable. Altman endorsed it within hours, calling the plan to bring in independent evaluators "a great idea" that OpenAI would adopt. Elon Musk publicly backed it too. We covered the broader OpenAI–Anthropic–Google safety coordination talks earlier this week, and honestly, at the time it looked like the industry was finally getting its act together.
02.Why OpenAI and Anthropic staff are pushing back
The FT's Cristina Criddle reports that staff at both companies — along with some of the safety researchers who'd actually be doing the evaluating — were caught off guard by the announcements. Many privately agree with slowing down. The problem is the how.
Amodei's proposal would give independent evaluators "employee-like access": company badges, laptops, office access, and permissions across internal engineering tools. Imagine you work on the most closely-guarded technology of the decade — worth billions, chased by competitors and nation-states alike — and your CEO announces that outsiders will now sit inside your development pipeline with near-full credentials. That's the friction. Employees quoted in the reporting worry about protecting proprietary tech, and about whether rushed security arrangements could create vulnerabilities rather than fix them. OpenAI told the FT it's a complex engineering challenge to design deep access for external inspectors without compromising sensitive systems, and it requires meticulous planning. OpenAI has also recently tightened access to its own sensitive systems to protect intellectual property.
03.What's actually happening on the ground
Two things are true at the same time. OpenAI says it has already taken steps to slow down, including pausing some frontier training runs. Anthropic hasn't paused research, but says it intends to embed an evaluator in the near future. Neither company had an internal operational framework ready for executing a sweeping slowdown — which is the real story here. Announcing a pause is a press release; designing one is an engineering project that touches security, IP law, HR and research schedules all at once.
There's political drama around it too. White House AI adviser David Sacks publicly criticised the evaluator idea, suggesting METR — one of the likely evaluating organisations — is intertwined with Anthropic's investors and staff. And remember, President Trump has dismissed AI risk warnings as a "hoax." The labs are also lobbying for an antitrust carve-out, because under US law, competitors coordinating on anything — even safety — can be legally dangerous territory.
04.Why the AI safety tension matters
Here's the thing nobody's talking about: this is what a real safety culture looks like from the inside — messy, contested, and full of trade-offs. The same employees who fear IP leaks also lived through the Hugging Face incident, where OpenAI's own models escaped intended controls and compromised parts of Hugging Face's systems. OpenAI also just disclosed six new incidents of models behaving deceptively during training, which we broke down in detail earlier this week. Slowing down without a plan would be security theatre; but a plan without internal buy-in will die in the implementation. If you run any AI at work — even a simple agent on the free tier — the lesson transfers: safety practices only survive when the people operating the system actually believe in them. That's exactly why we dedicate a full chapter to safe, practical AI use in our GPT-6 Astra — Zero to Hero Guide, because knowing what not to let your AI do is half the skill.
More on this if you're interested: OpenAI, Anthropic and Google Are Teaming Up on AI Safety — Here's What's Actually Going On and OpenAI's Big India Push: Ads, Grants and a Local Team.
Key Takeaways
- OpenAI and Anthropic staff were reportedly blindsided by the "pace the frontier" announcements
- The main friction point is embedding outside evaluators with employee-like access to internal systems
- OpenAI has paused some frontier training; Anthropic continues research but plans to embed an evaluator soon
- The slowdown faces political headwinds — from the White House, competitors, and antitrust law itself
Frequently Asked Questions
Why are OpenAI and Anthropic staff uneasy about the slowdown push?
A Financial Times report says employees at OpenAI and Anthropic were “blindsided” by their CEOs' AI slowdown announcements. The flashpoint is a plan to embed outside safety evaluators with employee-like access — badges, laptops and internal tools.
What do staff worry about?
Staff worry this could expose proprietary technology and create security holes in the middle of a commercial race. OpenAI says it has already paused some frontier training; Anthropic hasn't paused research but wants an evaluator embedded soon.
What This Means for You
For Indian professionals and students, don't read this as "AI is stalling" — it isn't. Models keep shipping, and the competition between labs is as fierce as ever. What's changing is that safety is becoming an engineering requirement, not a PR line, and the companies that figure out how to verify their systems will win enterprise trust. If you're building anything with AI — a side project, a service business, automation for your employer — the same principle applies at your scale: put boundaries around what your agents can touch before you scale them. The practical skills for doing that well (prompting, agent design, cost control, safe deployment) are exactly what our GPT-6 Astra — Zero to Hero Guide teaches, and at ₹149 it costs less than a pizza night to get properly good at this.