Keeping Pace with the Frontier: An Early Warning System for Accelerated AI R&D Automation

AI is playing a growing role in building the next generation of AI. Recently, Anthropic reported that, as of August 2026, Claude was leading 26% of the developer’s AI Research and Development (AI R&D) activities, up from around 1% in March 2026.

AI R&D automation has the potential to transform AI progress, ushering in systems far more capable than today’s at a greater pace. How large that impact will be remains uncertain. But recent reports suggest that automating AI R&D could significantly compress AI progress, outpacing society’s ability to adapt and straining frontier risk mitigations.

Drawing on its federal blueprint for frontier AI governance released earlier this year, Americans for Responsible Innovation (ARI) proposes a federal program to track the growing automation of AI R&D. Hosted within a new office in the Department of Commerce, this program would offer periodic, structured transparency about AI R&D automation to policymakers. Sustained federal visibility is the right first step for AI R&D policy and could help catalyze further interventions if the effects of automating AI progress prove to exacerbate frontier AI’s threat potential.

What Is AI R&D, and What Is Being Automated?

AI R&D is the scientific and engineering work that improves the state of the art of AI technology. It is composed of a range of tasks, from writing and debugging code, to running and analyzing experiments, to deciding which experiments are worth running and which experimental results should be integrated into future training runs. Today, AI can do most code production at frontier AI companies. Working under human direction, agentic systems can complete well-defined tasks over long horizons that would ordinarily take a human researcher several hours or even days. However, there are still core limitations to AI R&D automation, such as AI’s limited ability to independently judge valuable research directions (known as “research taste”) and its inability to autonomously direct training runs end to end without human input.

Frontier AI companies aim to overcome these limitations and fully automate AI R&D, with OpenAI aiming to create an automated “AI researcher” by March 2028. Even if some amount of human involvement in AI R&D remains necessary, just partial AI R&D automation could significantly boost the pace of AI development. Nevertheless, how much of a boost to AI progress we can expect from significant automation remains an open question. Some have argued that compute and parallelization bottlenecks could impede automation-enabled “takeoff” scenarios. While the physical world and research taste remain important limiting factors, sufficiently intelligent systems may conceivably overcome them. Uncertainty about the exact implications of AI R&D automation supports exploring governance interventions to more accurately estimate potential hazards and facilitate any further responses as may be warranted.

An Emerging Area of Risk

Only frontier AI developers know how fast their internal models are progressing, and they currently deploy models internally for long-horizon tasks that span weeks to months, including AI R&D activities. Under the status quo, the public mostly relies on developers choosing to share information about the pace of AI R&D progress and automation. And even when developers do share germane information, they use their own metrics which are not necessarily comparable across all developers’ AI R&D pipelines.

While this status quo persists, safety mitigations struggle to keep up. As was demonstrated in the OpenAI agents’ hack of Hugging Face, frontier AI progress is beset by alignment and control problems, and the technical breakthroughs needed to address them are lagging behind advancements in AI capabilities. Today’s AI models also pose severe misuse risks: they could hand bad actors new tools, such as advanced offensive cyber capabilities comparable to those of Anthropic’s Mythos. 

As more of AI R&D becomes automated, the urgency of mitigating misalignment, loss of control, and misuse increases. And if progress in AI capabilities accelerates without corresponding progress in technical safety science, there will be less time for the testing, evaluation, and research required to defuse frontier AI’s threat potential.

Require Developers to Report When AI Takes on More of Its Own Research & Development

Congress should establish a federal early-warning system for accelerated AI research automation. The goal of this proposal is to give policymakers timely evidence of whether AI development is accelerating beyond the ability of human oversight and safety testing to keep pace. Washington’s visibility into that shift should not depend on what companies choose to announce. Transparency is also a popular consensus step among AI researchers as a means of resolving the uncertainty around AI R&D automation’s exact implications.

To this end, ARI proposes a dual-track framework to provide structured, appropriate visibility into AI R&D automation. This program would run through a new office hosted within the Department of Commerce, which would receive confidential quarterly filings from major frontier AI developers describing the state of their AI R&D automation activities. 

The office would have authority to follow up on filings and make referrals to the Department of Justice if a developer knowingly submits a false or incomplete filing; a whistleblower program would add a backstop. 

Filings would include information pertaining to four statutorily fixed categories of AI R&D activity, with more granular disclosure fields and subcategories delegated to rulemaking:

For each category, filings would be anchored in specific, auditable, quantifiable measures, such as the share of computing power applied to autonomous activities and the share of AI-generated inputs and decisions in AI R&D that humans review. Filings would be exempt from disclosure under the Freedom of Information Act to help ensure information security and intellectual property protection. To inform the public, the office itself would periodically publish aggregate reports on AI R&D automation trends across the AI industry.

Between the quarterly filings, developers would also report material shifts in covered AI R&D activities within four business days of discovering the shift, or of when their own monitoring should have revealed it. Each such filing would report the same measures as the quarterly filing for the affected category, describe what changed, and note any responses the developer has taken. This secondary regime preserves federal visibility into automated AI R&D even if it suddenly speeds up. The triggering events for material-shift reporting would be set in statute, with additional indicators added through the rulemaking process. Initial triggering events could include the full automation of a research loop and the deployment of a new architecture that performs a covered activity autonomously. The office would review each such filing on receipt, could ask follow-up questions, and could move the affected category to a higher reporting cadence as may be warranted.

An Early-Warning System for Accelerating AI Progress

Leaders at frontier companies believe that end-to-end AI R&D automation is possible, and some, such as Anthropic co-founder Jack Clark, think it is more likely than not to occur by the end of 2028. Google DeepMind researchers have also argued that tracking AI R&D automation may have “disproportionate benefits” for forecasting AI progress. ARI’s proposed transparency program for AI R&D automation would equip policymakers and independent experts with visibility into frontier AI progress, while also protecting developers’ intellectual property and security. Congress should advance AI R&D automation transparency legislation to ensure clarity for the American people and promote responsible innovation at the frontier.