Twelve AI companies have now published safety policies for their frontier AI models. Nine of those define capability thresholds — lines in the sand where, if a model gets too capable, the company commits to slowing down, adding safeguards, or stopping entirely. Several of these policies followed the Seoul Frontier AI Safety Commitments, where leading AI companies voluntarily committed to “set out thresholds at which severe risks posed by a model or system, unless adequately mitigated, would be deemed intolerable.”
That’s the news that gets attention. The news that doesn’t: these thresholds don’t actually work as a shared safety system. The language across companies sounds similar but means different things. The way each company tests whether a threshold has been crossed is completely different. And there are few mechanisms for anyone outside these companies to verify or compare the results. The implication is that there is no concrete safety floor — “red line crossed at Lab A” does not mean the same thing as “red line crossed at Lab B”, even if they use the same language.
Shared words don’t guarantee shared outcomes
Let’s look at just one risk category from these frameworks: AI automating its own research.
Right now, one constraint on how fast AI improves is human researchers. But if AI systems start running the research process themselves — designing experiments, interpreting results, writing the next generation of algorithms — the bottleneck shifts. This is what Anthropic’s CEO Dario Amodei calls “a country of geniuses in a datacenter.” If it happens, every other AI-driven disruption speeds up with it: the pace of capability gains, the pressure on labor markets, the concentration of power among a smaller number of companies.
Among the labs that address AI research automation specifically, the language diverges immediately. Anthropic’s trigger is when a model can fully substitute for their entire set of Research Scientists and Research Engineers at competitive costs, or when it causes a doubling of the rate of AI progress. OpenAI’s critical threshold is “a superhuman research-scientist agent” or a system that causes “a generational model improvement in 1/5th the wall-clock time.” Google DeepMind triggers when AI progress is “substantially accelerating from historical rates.” Meta mentions AI research automation only as an “enabling capability” that evaluations should account for. xAI monitors it as a transparency metric but sets no formal trigger.
“We will consider this threshold to be met if we determine that either (1) our models would be able to fully substitute for our entire set of Research Scientists and Research Engineers, at competitive costs (i.e., within a factor of 5); or (2) there is ‘dramatic acceleration’ of the pace of AI progress for reasons that likely relate to the automation of AI R&D.” — Anthropic Responsible Scaling Policy v3.3
Thresholds can move, too
There’s a further problem that comparing snapshots hides: companies revise their own thresholds over time, and not always toward caution. Anthropic’s latest framework quietly dropped its formal cybersecurity threshold; others have reworded or recalibrated theirs between versions. When a commitment can be loosened by the same company that set it — potentially as its models start approaching the line — the “thresholds” stop functioning as commitments at all. A safety floor that each lab can move on its own isn’t a floor. This is hard to even track from the outside, since there’s no requirement to flag what changed between framework versions or why.
Measurement is where governance breaks down
Even if every company used identical language, we still couldn’t compare their assessments. Each lab tests for these thresholds using different benchmarks, different expert panels, different internal processes. For AI R&D evaluations specifically, none of the major labs have defined a numerical cutoff. The determinations rest heavily on qualitative expert judgment.
Here’s what that looks like in practice. Anthropic’s most recent determination that its newest model had not crossed the AI research automation threshold involved an internal survey in which 0 of 16 staff members believed the model could serve as a drop-in replacement for an entry-level researcher. That was one of the key inputs. Sixteen people at one company, answering one question. There is no way for a regulator, a competitor, or the public to run an equivalent assessment at another lab and get a comparable result.
A common standard is possible
It’s not all bad news. When you look past the surface differences in language, we can see elements of a shared structure.
In recent analysis done at CeSIA, when you formalize these definitions and translate them into comparable terms, the Anthropic and OpenAI thresholds land in roughly the same place — they’re asking for approximately the same amount of acceleration before triggering. This holds under plausible assumptions about recent rates of AI progress, though if progress has been speeding up recently (and there’s evidence it has), the gap between the two widens. Google DeepMind’s lower threshold — “substantially accelerating from historical rates” — is compatible with the same floor.
AI R&D harmonizes relatively cleanly because all three labs happened to frame their thresholds around the same type of underlying measure. Other risk domains are harder.
For biological and chemical risks, the labs do converge on a shared approach: they generally define thresholds in terms of how much a model helps a specific type of actor cause harm, relative to what that actor could do without AI. That’s a meaningful common structure, and a minimum floor is probably achievable. But the details diverge. OpenAI uses a 2021 baseline and “novice” actors. Anthropic uses 2023 and people with “undergraduate STEM degrees.” Google DeepMind uses “low to medium resourced actors.” Those differences matter when you’re asking whether a threshold crossed at one lab would register as crossed at another.
Cybersecurity is harder still. Anthropic’s most recent framework doesn’t include a formal cyber threshold. Google DeepMind and OpenAI kept theirs, but one defines the threshold in terms of the uplift a model provides to attackers and the other in terms of what the model can do autonomously. You can’t build a shared floor across labs when some don’t have a threshold and the ones that do are measuring fundamentally different things.
So the honest answer: the AI R&D case shows that harmonization is possible, but it doesn’t happen automatically. Getting there requires converting each company’s thresholds into a frame that can actually be compared.
Common language and common measurements still aren’t enough
Agreeing on what to measure is only the first step. The harder question is how to measure it, and who gets to check. Right now each lab runs its own evaluations with its own methods, and there are few mechanisms for an independent third party — like the EU Commission or the network of AI Safety Institutes — to verify or compare results.
And even beyond measuring, the risks need to be linked to concrete outcomes in the real world. Research from the Oxford Martin AI Governance Initiative (which CeSIA co-authored) found that safety frameworks introduce useful mechanisms, but at the same time they often “leave implicit fundamental considerations, including the definition of risk, the rationale for threshold selection, and the mapping of evaluation outcomes to risk severity and likelihood.” Just like an AI model passing a multiple-choice benchmark doesn’t make it fit to give legal advice in the real world, having a common language and common measurements still doesn’t tell you whether anyone is actually safer.
Where things stand
We have growing consensus on what capabilities to worry about. We have partial consensus on the language. We have almost no consensus on how to actually measure whether a dangerous threshold has been crossed, or how those measurements connect to real-world harms.
That last gap is the one that determines whether AI governance is real or increasingly just theater.
Bridging it is part of the technical governance work we’re doing at CeSIA. The goal is to identify evaluation methods that are technically sound, that institutions would actually adopt, and that a third party like a regulator could apply consistently across companies.