TL;DR: Labs are accelerating due to automated AI research and development. As an upper bound, in March 2026, Anthropic researchers self-reported producing 4x more output with AI tools than without them.

We lack public evidence on how this may differentially impact safety versus capabilities. As a first measurement of automated safety-specific research uplift, we surveyed 52 third-party AI safety professionals from 20+ orgs and found a median reported speedup of 1.58x.

We think it is important to develop standardized methods for tracking safety-specific uplift, as these will facilitate a better understanding of the differential impacts of RSI.

*This work was done by Arcadia Impact’s alignment team.*

–

AI R&D may not accelerate safety and capabilities equally

Frontier AI companies are currently automating their AI R&D efforts.  OpenAI claims that they are “making strong progress towards creating an automated AI researcher by March of 2028.” At Anthropic, Claude now leads 26% of AI R&D, researchers self-report producing 4x more output, and code commits arrive 5x faster per engineer than 6 months ago. As labs push towards RSI, it is important to understand whether AI research uplifts safety work equally.

To this end, we surveyed 52 third-party AI safety professionals from 20+ orgs and qualitatively compared these numbers to those reported by frontier labs. We found that the median respondent estimated a 1.58x speedup due to current AI tools. This post details our key findings, including: (1) which kinds of safety work benefit most from AI uplift, and (2) the relationship between automation levels and speedup.

This work has a few key takeaways:

1. Heavy AI adoption does not necessarily translate into large research speedups. Gains are concentrated in building and running experiments, while conceptual work remains a bottleneck. Anecdotally, people we’ve shown this to expected the speed up of third party safety work to be higher.
2. We cannot yet tell whether safety research is systematically lagging capabilities research, or whether independent researchers are lagging frontier labs. Our survey and lab-reported speedups use different methodologies, making direct comparisons difficult.
3. Consequently, it is important to develop better points of reference. In particular, we would like to understand the speedup across the following four quadrants of AI R&D labor:

If we had all the above cells filled in using the same methodology, then we would have a better sense of the risk associated with RSI.

The current version of this survey has several limitations, which we address at the bottom of this post. Nonetheless, we believe it is a reasonable first step towards better analysis of safety vs. capabilities acceleration.

We welcome feedback on interpreting the results, as well as suggestions for future iterations.

Survey

We received responses from 52 professionals in AI Safety, including technical researchers, research leads, and individuals working in AI policy, governance, and grantmaking. Respondent organizations include Aether, Arcadia Impact, Arrow, Apollo Research, Coefficient Giving, CNAS, Equistamp, Epoch, GovAI, Geodesic, Guidelight, LASR, METR, Oxford, Principia, Redwood Research, and Safe AI Forum.^[2]

We adopt Epoch AI’s framework, which divides AI R&D work into six areas, and slightly modify subcategories to better represent AI safety work.^[3] 

The survey asked respondents to distribute 100 'points' across these categories to represent their responsibilities. Respondents then estimated how many hours of active human work would be required within each category to complete their most recent project, provided:

• No AI
• AI tools^[4] from 1 year ago
• Current AI tools
• AI tools, projected 6 months into the future

Overall speedup is calculated by weighing how much each area contributes to a respondent's job and what the acceleration within that area was compared to the no-AI setting^[5]. We also report qualitative summaries to respondents’ answers on open-ended questions.

The survey remains open for additional responses here: https://arcadiaimpact.github.io/automation-survey/.

Current AI tools make safety researchers 1.58x faster.

The median respondent estimates that current AI tools make them 1.58x faster. The speedups vary across roles, with full-time Members of Technical Staff (MTS) reporting the highest speedup of 1.69x, while those working in policy, governance, and grant-making self-report at 1.43x.^[6]

Speedup is mostly attributed to building and running experiments.

AI does not accelerate every part of safety research equally. Respondents report a median 5x speedup in building experiment infrastructure, code, and datasets; they also report a 3x boost in running experiments. Projected speedup in the next 6 months is also highest in these areas, at 10x and 5.6x respectively.

In contrast, the estimated gains are much lower for choosing research directions, designing experiments, and analyzing results.

However, accelerated tasks are a minor share of respondents’ jobs.

Despite high current and projected speedups for building and running experiments, these tasks only constitute approximately 28% of the work that respondents are responsible for. Consequently, respondents spend less time on these tasks, but do not experience significant end-to-end speedup.

People report that their AI Safety work is mostly conceptual and analytical: deciding what to work on, building up relevant context, designing experiments to appropriately test an idea, writing up results, and making the results legible to different audiences.^[7] As they hand off more “build and run”-style tasks to AI, people now spend more time on conceptual work.

Despite heavy agent usage, organizational automation is low.

Most respondents are already AI power users: 69% mention custom skills, prompts, or MCP integrations, and 65% report using multiple agents or long unattended runs.

However, more elaborate infrastructure is less common. Only 14 respondents (27%) mention custom scaffolds or orchestration. We only identified two organizations whose respondents described shared automation tooling and practices developed at the organizational level, rather than solely for individual workflows. These org-wide setups ranged from custom coding and review agents to automated research reports, quality checks, and tools for coordinating multiple agents across teams.  

This suggests that org-wide automation may be an underexplored opportunity for improving research productivity across the safety ecosystem.

Being an AI power user does not clearly correlate with greater speedup.

We manually classified respondents' free-form descriptions of their setups on a scale from 1 (chat-only use) to 4 (heavily customized agent workflows).^[8]

We found no clear relationship between setup complexity and reported speedup.

This could reflect diminishing returns to customization; researchers with more elaborate setups may also be attempting less automatable tasks. We note that low-complexity setup users are fairly underrepresented, with high variance due to significant outliers.

Research judgment remains the biggest reported bottleneck.

When asked what prevents them from automating more of their work, respondents repeatedly point to limitations in research taste and scientific judgment.

47 of 52 respondents describe issues that fall under this broad category. These include choosing research directions, designing valid experiments, identifying important problems, and interpreting results.

Researchers also report spending substantial time steering agents through repeated iterations and verifying their outputs. These activities can erase some of the time saved through automation, particularly when failures are difficult to detect.

Survey limitations

We are primarily interested in understanding whether automation accelerates capabilities researchers more than it does safety researchers.

However, it is not clear whether perceived speedup gaps between our survey respondents and lab employees are due to:

• A safety vs. capabilities gap
• An inside- vs. outside-lab gap in resources and automation culture
• Inconsistent methodologies between the labs’ internal surveys and our external one

Furthermore, there are several other challenges when interpreting our survey results:

• These are self-reported estimates and many respondents say the estimates were difficult to provide. Comments include "man these are such vibes," "seems really hard to estimate," and "massive uncertainty here, including in direction."
• The counterfactual is unclear: some respondents say AI allows them to explore more ideas or run experiments that wouldn’t have been possible otherwise. These gains may be difficult to express as a speedup vs. a no-AI baseline.
• This number may overstate the gain for people who lose time steering, checking, or fixing mistakes.
• It’s possible that models are sandbagging AI safety research or external AI R&D.
• Time speedups do not necessarily translate to actual safety progress or uplift. Namely, they do not establish how much better, more valuable, or more scientifically useful the resulting research is.
• Those in TAIS have overlapping intuitions but no shared definition of “uplift.”^[9] 

Frontier lab safety uplift is currently underreported and unstandardized

Taking a step back, we think it is particularly valuable for the community to converge on a unified set of metrics for evaluating AI safety vs. capabilities speedups. When labs publish acceleration data, they group capabilities and safety work into general “R&D.” This obscures how autoresearch efforts and outcomes are distributed between the two areas. Of recent lab reports with acceleration data, safety-specific automation data is extremely sparse:

• Anthropic: Safety gets ~6% of R&D compute.^[10] This was reported for a single week: July 13 to July 20 of this year, with no further information as to whether the week was representative or not. The outcomes and relative safety uplift from this compute are unclear.
• OpenAI: A September research acceleration update only refers to safety in its final section, where they show “Astra-class” RL compute dropped for a single month (mid July to mid August) during a HuggingFace-induced pause.
• DeepMind: The recent Gemini 4 Argon safety and security report listed two findings on AI R&D speedup, and none of them measured safety-specific acceleration.

If Claude leads 26% of Anthropic’s AI R&D tasks, how much of this is directed towards safety research? Which parts are being accelerated, and are these speedups translating into meaningful safety progress? These questions are difficult to answer meaningfully; measuring “safety research” or “effort” is fuzzy and unstandardized. Nor can we classify every task as solely “safety” or “capabilities” work, and in Anthropic’s own words, “[E]ach developer will be tempted to draw the line generously.”^[11] Privatized safety practices further compound difficulties with cross-lab comparisons; as a result, being calibrated on AI safety progress is extremely difficult.

Robust ^[12] and shared notions of automated safety uplift are therefore extremely valuable. We are iterating on this task which, done right, could lay the groundwork to legibly audit safety efforts and outcomes. We would welcome feedback or collaboration.

–

We would like to thank Andrew Draganov, Moritz Pfau, Aleksandr Bowkis, Francis Rhys Ward, Joe Kwon, Alan Chan, Aniket Chakravorty, Jean-Stanislas Denain, Cameron Holmes, David Africa, Tyler Tracy, Michal Bravansky, and Matthew Bozoukov for helpful comments.

–

1. ^{^}

   This figure plots 47 responses, as 5 people did not answer this section of our survey.

2. ^{^}

   Individual responses do not represent their organizations as a whole, nor should they be considered endorsements of this post.

3. ^{^}

   We combine Analyze & Write as the 5th category, and add “Collaborate” as the 6th category, to better represent the range of work done across the AI Safety community.

4. ^{^}

   Tools are defined as AI assistance in any form (chat only, coding agents, etc)

5. ^{^}

   Similar to Anthropic’s R&D Automation Index, which weights automation levels by estimated person-time, we weight task-level speedups by each task’s contribution to a respondent’s job.

6. ^{^}

   We note that the policy and grant making sample size is small.

7. ^{^}

   Our collected responses are directionally similar to what labs have reported:

   • At OpenAI, employees’ agent usage is heaviest in EpochAI’s R&D categories of “Build,” “Run,” and “Communicate.” The high “Communicate” uplift is due to “technical help & review,” while other subcategories like research write-ups and announcements remain extremely low.
     • These areas also report the most growth: compared to the start of 2026, researchers now use an additional 133.1-198.2k tokens per day on these tasks.
     • In comparison, deciding what to work on, deciding which tasks to continue or stop, designing technical specifications, and experiment planning have only increased by <10k tokens per researcher per day over the same timespan.
   • Compared to 2021-2025, Anthropic employees now commit 8x more code per quarter on average. As of May 2026, more than 80% of merged code at Anthropic was Claude-written. However, Claude is still slightly weaker at steering research sessions, with an estimated 64% of Claude Mythos Preview samples beating human researchers’ next-step suggestions.

   
   

8. ^{^}

   1 - chat only, no coding agents; 2 - coding agents out-of-the-box, 3 - minimal customisation with prompts/skills; 4 - combining several practices (MCPs, parallel agents, always-on machines, custom scaffolds)

9. ^{^}

   See, for example, the Appendix in this post.

10. ^{^}

    Anthropic: “These are deliberately conservative estimates. For example, if a token was used to advance capabilities as much as it was to advance safety, it was not counted in these metrics. Additionally, these metrics do not account for safeguards classifiers, which are a separate, comparable amount of compute that make our models much safer for the world.”

11. ^{^}

    Anthropic’s full quote: “Safety research is hard to distinguish from capabilities research, and each developer will be tempted to draw the line generously. The burden of proof should sit with the developer to show that work is safety-related. Developers, governments, and the wider research community would benefit from converging on a shared definition ahead of time. A measurement like this could inform future actions, such as a lab’s commitments about the share of compute going to safety research, or limits on the share of compute going towards AI research agents.”

    While not safety-specific, OpenAI has also advocated for better measurements of acceleration data:

    • OpenAI: "Agent-powered AI research is still new, and we are still learning how to measure it. [...]  Metrics that focus more directly on research progress [...] could be more useful, but are complex to develop and validate. [...] Deepening our understanding of research acceleration is a significant focus area across OpenAI."
12. ^{^}

    Safety washing is a significant failure mode. Organizations can inflate poor proxies for safety progress without making models meaningfully safer.