A group of researchers from multiple institutions has laid out a framework for research on ‘‘pacing’’ AI development — that is, intentionally influencing the speed of AI progress rather than trying to stop it completely. Their paper argues that without better tools and evidence, the world’s response to rapid AI advances could be haphazard, reactive, or persist in institutions that outlast their usefulness. The authors hope research can make pacing more deliberate, targeted, proportionate, decisive, and legitimate.
Who produced the paper
The work is authored by researchers affiliated with ACS Research, University of Toronto, Arb Research, Paradigm 3 Institute, the Wharton School at the University of Pennsylvania, Goodheart Labs, Trajectory Institute, Harvard University, and the University of Cambridge.
Pros and cons of pacing
The paper summarizes common arguments on both sides of the pacing debate.
-
Arguments against pacing: slowing development may delay beneficial outcomes such as healthcare or other science‑driven breakthroughs; it could increase concentration of power; it might create a false sense of security and enable capability overhangs as compute costs fall during the pacing period; early pacing can be less efficient than later, more surgical interventions because advanced AIs could assist better‑targeted pacing; and some pacing regimes may be difficult to reverse if designed poorly (the authors point to nuclear power regulation as an analogy).
-
Arguments for pacing: it can buy time to prepare for threats tied to AI advances (for example cyber or bio risks); AI progress is hard to predict and unexpected dangerous developments can arise; rapid AI progress could contribute to worldwide lose‑lose outcomes (for instance, governments ceding certain military decisions to AI systems, reducing human agency).
Incentives and coordination
The paper emphasizes that the practical impact of any intervention depends heavily on how affected actors respond, including those enforcing the intervention. Governments face mixed incentives: they both compete internationally for advantage and must respond to domestic political pressures. Infrastructure providers — notably semiconductor and cloud companies — generally have strong incentives to push forward and much to lose from pacing. Consequently, the authors argue that some contexts will require coordinated pacing rather than voluntary unilateral measures. Coordinated pacing is more onerous because it requires alignment across stages of the intervention and mechanisms to sustain coordination over time.
Research on interventions and lifecycle analysis
Authors recommend research to improve technical and regulatory interventions. While every intervention involves trade‑offs, new coordination, oversight mechanisms, and prior planning may enable strictly better options. They also advise studying a pacing intervention’s full lifecycle — from pre‑intervention to exit — to identify gaps and failure modes.
What to pace and how
The paper lists multiple potential pacing targets:
- Inputs and infrastructure: compute, data, power, capital, researcher time.
- Capability development: algorithmic R&D, pre‑training, post‑training.
- Model artifacts: for example, model weights.
- Use and distribution: internal use, deployment, access (e.g., weights), and other diffusion channels.
The authors distinguish rival goods (resources whose use by one actor limits availability to others, such as compute or researcher time) from non‑rival goods (items like model weights and training algorithms that can be copied and shared cheaply). This distinction matters for policy design: rival goods can be monitored, taxed, or redirected, while non‑rival goods are much harder to control.
They also note that the timing of a pacing intervention will often reflect political opportunity, public perception, salient incidents, or campaigning efforts, not only technical thresholds.
Practical questions to guide pacing decisions
The study proposes a set of questions to ask at different stages of a pacing intervention:
-
Pre‑pacing: What justifies intervention? Which signals of emerging risk are tracked? What infrastructure is necessary for those signals to be detectable? Who watches and who interprets and reports the signals?
-
At trigger: Who has authority to determine a trigger condition is met? How much error is acceptable in order to act quickly? What sequence of actions does triggering initiate? Are developers obliged to remediate the triggering concern or free to abandon the blocked path?
-
During intervention: How is compliance observed, verified, and enforced? Who reports compliance evidence, who audits, and who acts on discrepancies? How is downtime used to respond to the threat? Who, if anyone, may continue restricted work and under what oversight?
-
At exit: How do decision‑makers judge whether an intervention has succeeded, failed, or become obsolete? Should exit be immediate or staged? Who is affected by exiting, and who bears costs or gains?
Why this matters
The authors argue some form of pacing will likely occur: the combination of the technology’s power, political economy, and the scale of risks makes responses probable. The central choice is whether those responses will be precise and targeted — enabling nuanced, ‘‘scalpel‑like’’ interventions — or blunt and indiscriminate. The paper’s aim is not to offer canned policy prescriptions but to frame a research agenda: better theoretical foundations, coordination tools, and lifecycle planning may increase the chances of effective, proportionate, and durable pacing when the need arises.
