Altman Backs Slower AI Training While OpenAI Talks Joint Rules With Anthropic and Google
OpenAI has begun running safety checks before major training runs and has spent months in talks with Anthropic and Google about shared self-regulation, even as Sam Altman says rapid progress will continue.
Sam Altman is asking the AI industry to pace itself, and OpenAI is putting internal checks behind the pitch. The company now runs safety evaluations before it launches major training runs, a step that adds friction to the same research pipeline Altman says will keep moving fast. His position, aired in a piece first reported by The Decoder, is that slowing down and speeding up are not opposites: guardrails go in, and capability work continues.
The more consequential signal is who OpenAI is talking to. According to reporting from The Information, the company has been in discussions with Anthropic and Google for months about joint self-regulation. That is a notable shift for three labs that compete for the same researchers, compute and enterprise customers. If the talks produce anything concrete, the terms would be set by the companies themselves rather than by a regulator, which is exactly the arrangement safety critics have questioned.
The pre-training checks matter because that is the stage where the largest models are made and where the biggest share of compute gets spent. A safety review inserted before a major run can delay a launch, alter a training recipe, or kill a run outright. OpenAI has not published the criteria its reviewers use, how often a run gets flagged, or who has authority to stop one. Without those details, outside researchers cannot tell whether the process is a real gate or a formality that runs alongside the schedule.
Altman's framing also lands at an awkward moment for the sector. Governments in the EU, UK and US are writing or enforcing rules that would bind frontier labs, and industry-led self-regulation is often offered as a lighter alternative. Voluntary commitments from the same companies can shape what lawmakers accept as sufficient. That gives the OpenAI-Anthropic-Google talks weight beyond the three firms, because whatever they agree to could become the baseline other labs are measured against.
Anthropic has run its own responsible scaling policy since 2023, tying model deployment to evaluated risk levels. Google DeepMind has published frontier safety frameworks with similar tiered thresholds. OpenAI's move toward pre-run checks puts it closer to that pattern, though each company's thresholds and review methods differ. A joint framework would need to reconcile those differences, and none of the three has said what a shared standard would actually require.
What to watch next is whether the talks yield a published document, and whether it includes third-party audits or an outside body with authority to enforce anything. A statement of principles costs little. A mechanism that can pause a training run at a lab that did not agree to the pause is a different kind of commitment. Until that exists, the industry's self-regulation remains a promise made by parties with a direct interest in keeping the schedule.