Model Parameters

Anthropic + Google DeepMind + OpenAI have ~2.6x more compute than OpenAI alone. The compute input series used by AIFM is based on OpenAI, so we adjusted it.

Anthropic + Google DeepMind + OpenAI have ~3.2x more staff than OpenAI alone. The staff input series used by AIFM is based on OpenAI, so we adjusted it.

We keep it at the same 5% as in AIFM for simplicity, but note that it is not optimal.

When we let AI labs optimize their compute allocation to maximize AI safety work, they change $s_{aut}$ to anywhere between 0.3% and 20%, which adds up to 11% more AI safety work. We still keep it fixed, as in AIFM, and err on the boring side.

Compute serving external users is frozen at its level at the start of the pause, 38% of total compute. Since total compute keeps growing, the share allocated to external deployment shrinks over time.

On the one hand, demand for models at the frozen intelligence level grows:

• There is continuous adoption of AI.
• When AI capabilities are frozen, building scaffolds makes more sense: there's no fear that GPT-7 will crush your startup.

On the other hand, under normal, non-pause conditions, compute efficiency improves so fast that the cost of the same level of intelligence drops 10–40x a year, which easily covers the increased demand. However, in a pause world, we might ban all research leading to better efficiency.

The uncertainty goes both ways, and neither direction affects experiment compute much, so we keep it fixed. This also matches the AI 2040 modeling.

Some experiments fail, so a paper uses more compute than its described experiments add up to. We accounted for that when we estimated the papers' compute.

But sometimes a whole research direction fails and is never published. We don't see it among the papers, so we account for it separately, through $M$. We use it twice:

• For the 2025 AI safety field outside the AI labs, whose compute we take from papers (see 2025 AI safety work). We don't see failed papers, so we multiply the published papers' compute by $M$ to account for that.
• For the labs in the pause. We smear compute for the failed experiments across papers, so their re-runs cost only $\kappa_c / M$.

We guess it is no more than 1.25.

See 2025 AI safety work.

See 2025 AI safety work.

See 2025 AI safety work.

A simple way to estimate the distribution of the papers' quality is to fit a Pareto distribution to their citations (Price, 1965; Redner, 1998; Peterson et al., 2010). But citations mix up the quality of a paper's idea with the quality of its authors. The difference matters: after automation, only the AI's taste will be left, so the distribution should become flatter. To separate the two, we use the Q-model (Sinatra et al., 2016; Janosov et al., 2020), which we fit to our AI safety papers dataset. We get $\alpha_v = 1.45$.

We use α for the quality of a paper's idea, $v_i$, which doesn't depend on compute. But don't citations depend on compute? We checked: only slightly. The top 20% of papers by compute hold 30% of the citations, against 20% if compute didn't matter. Comparing only papers with similar compute gives almost the same α: 1.43. Also see Chen et al., 2026: among NLP conference papers, the top 20% by GPUs hold 27–32% of the citations.

Citations also reward being noticed: early citations bring more, whatever the paper is worth. So real contributions are probably less lopsided, α is likely higher, and the results where the labs check papers with their own staff are optimistic.

See 2025 AI safety work.

We count only human labor, not the time it takes to run reproduction experiments. The denominator is the total labor of all the paper's authors. In 2025, 0.02 corresponds to 20.7 hours per verifier.

The less verifiers know the field, the more prerequisites they have to read before they can check a paper. We assume that in 2025, verifiers could check papers from their own field without reading any prerequisites. But with so many AI-generated papers, their field grows in proportion to $n(t)$, so verifiers have to spend more time catching up with it.

We aren't sure about this parameter. We tried to build a random graph model of the prerequisite dependencies among papers, but it was too complex. We still used it, together with historical paper trends, to calibrate our intuition. To reflect our uncertainty, we use a wide range and the point estimate that makes our results more boring.

An FTE-year is 1,500 hours, so 0.00614 FTE-years add 9.21 hours to each check for every tenfold more papers.

It's a rather arbitrary modelling choice. We assume reproduction experiments will be run only once, but we think there might be several people carefully reviewing a paper.

From the paper count (27): in a year $t$ before 2025, when papers had the labor of 2025's papers, research per paper relative to 2025 is compute per paper relative to 2025, to the power $\zeta_p$:

We took the papers $n(t)$ and their compute from our AI safety papers dataset: $\bar c(t)$ is their average compute, dead ends included. We computed $RS_s(t)$ the way we compute $RS_s$: AIFM's $X$ at that year's compute, times $M$ for the failed projects, and people (scaled by the number of distinct authors), with that year's AI help.

Against 2025, the years 2020 to 2024 give $\zeta_p$ in [0.21; 0.34]: 0.29, 0.21, 0.25, 0.25 and 0.34. We took the average, 0.27.

Its bounds:

• 1: labor can't make a paper worse.
• 6.1: AIFM's own $X$ for the 2025 AI safety field with unlimited coders. One paper can't gain more than the whole field, as long as more labor never means fewer papers.

We define a paper as a piece of research with a fixed human verification time, apart from studying its prerequisites. This limits how much labor you can squeeze into a paper given fixed compute and verification time. We don't have a better estimate, so we go with 1.5×.

We estimated it from our AI safety papers dataset. In 2020–2025 the median compute grew 1.43× a year, and $\sigma_c$ was always within [2.52; 2.77], with no trend. Opus's estimates don't widen or narrow it: on the test papers they are spread as much as the authors' own figures (2.81 vs 2.85).

To verify a paper's main contribution, you likely need to re-run only a small number of experiments.

The AI researcher is likely superhuman at coding, so you might lose some research from not letting it write the code. On the other hand, it may be able to describe the experimental setup in enough detail for its plans to carry most of its taste to the AI engineer.

In 2025, 0.003 corresponds to 3.1 hours.

It could be shorter. In Compartmentalization, the AI is better at AI R&D than almost any employee, so humans only check that plans have no dangerous instructions. A plan isn't a paper: it's just instructions for the AI engineer to implement a specific experiment, plus the rationale for why it's interesting. It's likely to be short.

It could be longer. Some experiments may be hard to explain to the AI engineer, so the AI researcher would have to write a more detailed plan. Note that $\kappa_p$ assumes the human who reads the plan already knows the prerequisites to understand it. We account for those in $p_p$.

The rationale is the same as for $p_h$. Here it's smaller because the human only needs to check that the plan is safe to hand to the AI engineer, without following all of its reasoning.

An FTE-year is 1,500 hours, so 0.00461 FTE-years add 6.91 hours to each plan for every tenfold more papers.

At Google, developers spend an average of 3.2 hours a 40-hour week reviewing their colleagues' code changes(Sadowski et al., 2018). Auditors looking for hidden rule-breaking would read more slowly, so this is optimistic.

Just like the number of verifiers, the number of auditors is both important and difficult to estimate. For verifiers, we studied two scenarios: hiring frozen at the start of the pause and unlimited hiring. This allowed us to measure how much we lose by not hiring.

To simplify communication, we calculated a table for a specific number of auditors and then ran simulations for different numbers of auditors.

Some considerations:

• You need competent people who understand AI R&D code and can detect unsafe code. Given the breadth of AI R&D, you might need a lot of specialists in different areas.
• There were ~20,000 NeurIPS reviewers in 2025.
• Historically, the number of auditors for international treaties has ranged from hundreds to several thousand.

For specificity, we fix it at a large but realistic number.