How Should the US Prepare for Increasingly Automated AI R&D? (IFP, August 2026)

"How Should the US Prepare for Increasingly Automated AI R&D? 23 low-regret policy recommendations" is a report published August 6, 2026 by the Institute for Progress, written by Tim Fist, Saif Khan, Tao Burga, Arthur Tellis, Ben Schifman, Jonah Weinbaum and Olivia Scharfman. It assesses the argument implicit in the July 2026 open letter signed by frontier-AI company employees calling for the capacity to "pace" automated AI development (Pacing the Frontier (statement from employees of frontier AI companies, July 2026)), and proposes preparatory measures the authors argue are worth taking whether or not the risks materialize.

Fist and Khan published a short-form restatement of the report's first half as a guest post on Noahpinion three days later (Should we \"pace\" AI self-improvement? (Fist and Khan, August 2026)); the 23 recommendations appear in the report, and the report's seven top-level sections are the seven preparatory areas the guest post lists.

The argument assessed

The authors break the letter into three claims: that frontier AI companies are close to fully automating AI R&D; that automating AI research would pose serious risks; and that building the option to pace is a good way to address those risks. Their stated findings are that rapid progress toward fully automated AI R&D has empirical support, but that how much capability acceleration this will cause and what risks it might pose are less well understood; that some preparatory policy action is nonetheless warranted, both because of the seriousness of the possible direct risks and because of the risk that political backlash to AI-driven disruption produces poorly reasoned measures such as broad bans on new data centers; and that any decision to slow AI progress should not be taken lightly given benefits the authors list as accelerated economic growth, new technologies and scientific breakthroughs including novel cures for diseases.

On the evidence for automation, the report separates the skills involved into software engineering and research taste. It cites METR's time-horizon work for the finding that model capability on engineering tasks, measured against how long humans take to complete the same tasks, is doubling roughly every seven months (METR), and reports Anthropic findings that its models now outperform humans under a fixed time budget on an AI R&D task focused on speeding up model training, and that in a separate open-ended AI-safety research project Anthropic's models achieved a 97% performance improvement against 23% for two human researchers given a similar five-to-seven-day budget. It cites METR's February 10, 2026 simpler-timelines forecast model for the estimate that over 99% of AI R&D tasks will be automated by 2032. The authors record countervailing evidence: compute bottlenecks, diminishing returns to research effort, data bottlenecks, and the possibility that progress in verifiable domains fails to transfer to tasks without a clear correct answer. A footnote sets Anthropic's Mythos 5 system card, which suggests recent Mythos models established a higher trend line for capability growth, against Epoch AI's independent benchmark, which as of August 2026 did not show a recent speedup (Epoch AI).

The proposed operationalization of pacing

The authors argue the term "pacing" is vague and could encompass many measures, and propose an if-then conditional structure in two parts: specifying which automated AI R&D activities are likely to pose severe risks, with thresholds set on careful analysis; and, if a threshold is exceeded, incentivizing reallocation of resources away from those activities toward either accelerating the diffusion of AI capabilities — allocating compute and talent to inference and new AI applications — or accelerating research that makes further AI research automation safer, whether by improving model safety directly or by boosting societal resilience.

Under this reading a paced form of automated AI R&D might still involve much faster capability improvement than today, and need not entail slowing innovation overall. The authors note that identifying concrete thresholds might let camps that disagree about how much automation is achievable reach what they call positive-sum compromises, and argue that mitigating risks may be a precondition for the sustainability of rapid AI progress (Coordinated slowdown proposals compared).

Five criteria govern which near-term policies the report proposes. They should focus on serious and irreversible harms; minimize slowdown in the diffusion of existing AI capabilities; have upside even if automated AI R&D and its attendant risks prove unlikely; avoid systematically disadvantaging more cautious companies and countries; and avoid establishing a regulatory apparatus likely to be misused.

The 23 recommendations

Provide transparency into automated AI R&D

1. Frontier AI companies and relevant industry bodies should publicly share information relevant to trends and risks in AI R&D automation.
2. Congress should legislate transparency about automated AI R&D risk management, incident reporting, whistleblower protections, and model behavior specifications.

Improve state capacity to understand and respond to automated AI R&D

1. Congress should resource the Center for AI Standards and Innovation with a budget of at least $84 million per year and empower it to directly advise senior government officials and frontier AI companies.
2. The White House should set clear roles and responsibilities of US government agencies to increase specialization across AI policy.
3. Intelligence agencies should improve their collection and analysis on foreign AI development and counter threats targeting US AI companies.

Develop a risk management strategy for automated AI R&D that accelerates defensive and commercial AI uses

1. CAISI should develop guidelines for managing the risks of rapid AI capability improvement.

Accelerate the development of AI verification technology

1. CAISI should co-lead an AI Verification Consortium (AIVEC) with industry to prototype and deploy verification technologies.
2. AIVEC should coordinate the creation of AI hardware testbeds and make them available to government, industry, and nonprofit partners.
3. AIVEC should launch philanthropically funded prize competitions for AI verification headed by CAISI, in the form of lightweight prize competitions and an adversarial grand challenge.
4. AIVEC should coordinate the construction of a fully verifiable data center.
5. The Defense Advanced Research Projects Agency and the National Science Foundation should set up AI verification R&D programs.
6. Intelligence agencies should develop and operationalize unilateral means of AI compute monitoring.

Invest in AI resilience

1. The National Security Agency, CAISI, the Cybersecurity and Infrastructure Security Agency, and the Office of the National Cyber Director should further invest in cybersecurity resilience.
2. Congress, the Office of Science and Technology Policy, and the Centers for Disease Control and Prevention should invest in biosecurity resilience.

Extend the US AI lead to give the US more time to manage AI R&D automation risks

1. Congress and the Bureau of Industry and Security should strengthen controls on US and allied semiconductor manufacturing equipment.
2. Congress and BIS should close gaps in AI chip controls.
3. The Federal Trade Commission, Department of Justice, BIS, CAISI and Congress should help industry counter adversarial distillation of US AI model capabilities.
4. BIS should maintain visibility into sales of US chips.
5. The Department of War, intelligence agencies, CAISI and relevant Federally Funded Research and Development Centers should establish consensus security guidelines for protecting model weights from theft and prototype them in a government facility.
6. Congress should ensure the US has sufficient electrical capacity to sustain AI leadership.
7. Congress should ensure AI infrastructure like data centers can be constructed in America.

Create option value for international cooperation on managing automated AI R&D risks

1. The US government should use its bilateral AI dialogue with China to jointly develop guidelines for managing risks from rapid AI capability growth and prepare verification measures.
2. Countries with national AI institutes should collaborate on automated AI R&D risk management guidelines and technical capacity for AI verification.

On recommendation 3, the report notes that IFP had previously recommended increasing CAISI staff to 184 personnel with an $84 million budget, and that the America First Policy Institute recommended $50–100 million per year; the authors describe this as a minimum bar. They argue CAISI's autonomy and business model matter more than its size, and propose that it have a direct line to a senior White House or cabinet-level official and be empowered to forward-deploy staff into frontier AI companies, organize third-party AI evaluators and data producers, and establish direct contracts with frontier companies to support priority research and security efforts, including compute subsidies for alignment teams.

The China lag estimate

Using the Epoch Capabilities Index, the report compares the capabilities of the best models available for deployment by US organizations — US closed-weight models — against those available for deployment by Chinese organizations, which it takes to be Chinese closed-weight models plus all open-weight models. It estimates that China-available models should first match Claude Mythos's ECI score at the end of November 2026, and gives the expected lag as 7.7 months, described in the section heading as a Mythos-class model likely being deployable by Chinese organizations by the end of 2026 (US-China AI Competition: Different Races, Different Metrics).

Relationships

• supports: Pacing the Frontier (statement from employees of frontier AI companies, July 2026) — assesses the letter's three claims and finds preparatory action warranted, while declining to endorse a general slowdown
• related: Should we \"pace\" AI self-improvement? (Fist and Khan, August 2026) — the authors' short-form restatement of this report's first half, published three days later; the report is the primary document and carries the 23 recommendations
• contradicts: Artificial Intelligence Data Center Moratorium Act — named as the vivid example of the counterproductive default the report is written to forestall
• contradicts: The Future is for Everyone (Zuckerberg, August 2026) — on whether directing compute to recursive self-improvement should be paced at all
• related: Responsible Innovation at the Frontier (ARI, August 2026) — same visibility gap in automated AI R&D, addressed by mandatory filings rather than preparatory capacity
• depends-on: Recursive Self-Improvement (RSI)
• related: Compute Governance, Frontier AI Governance, AI Race Dynamics
• related: Adversarial Distillation — recommendation 17
• related: AI Biosecurity, AI and Cybersecurity — recommendations 13 and 14
• related: NIST CAISI (Center for AI Standards and Innovation) — named in recommendations 3, 6, 7, 9, 13, 17 and 19
• related: Tim Fist, Saif M. Khan, Institute for Progress (IFP), METR, Epoch AI

Provenance

Single source: the report as published on IFP's own site, with a companion PDF. Confidence is medium — the recommendations are a think-tank policy position rather than an empirical finding, none had been introduced as legislation at the time of ingest, and the report's own framing is conditional throughout. The empirical claims it relies on are attributed to third parties (METR, Anthropic, Epoch AI, CSET) rather than generated by the authors, and the report itself records where those sources disagree.