How an AI Slowdown Might Be Implemented

Numerous AI researchers are increasingly convinced that the technology they are developing could eventually pose significant dangers. What remains uncertain—even among the leading minds in AI—is exactly how to regulate these unpredictable algorithms.

In recent times, researchers have proposed various strategies for preventing AI from becoming a threat. These range from less contentious ideas like stricter government regulations, innovative metrics for progress evaluation, and examining model architectures, to more extreme suggestions such as embedding tracking devices in GPUs and even ceremoniously destroying substantial quantities of AI chips.

Despite rising political and public demands for a more cautious approach to AI development, solutions for ensuring its safety remain ambiguous.

“We must begin treating this as a research issue,” asserts Raymond Douglas, an AI researcher at the University of Toronto and coauthor of a recent report titled Pacing the Frontier, A Research Agenda, which cautions that finding ways to decelerate AI advancement is still an unresolved challenge. “We lack a clear understanding of our choices and their potential consequences.”

In recent weeks, discussions around AI-related disaster have intensified after a former researcher from Anthropic cautioned that, in just a couple of years, AI could be on a trajectory to threaten humanity. The head of Anthropic’s AI safety lab quickly echoed these warnings.

The executives of major American AI firms—Dario Amodei from Anthropic, Sam Altman from OpenAI, Elon Musk from SpaceXAI, and Demis Hassabis from Google DeepMind—have all now expressed support for some form of AI slowdown or moratorium.

This issue feels particularly urgent as AI companies increasingly leverage AI itself to construct even more potent models. This has raised concerns about a rapidly accelerating recursive self-improvement (RSI) loop, which could enable AI to surpass human understanding of its actions within a few years.

AI labs are already promoting their own new methodologies. This week, Anthropic unveiled several innovative ways to monitor how quickly—and potentially dangerously—artificial intelligence is evolving. For example, their methods indicated that Claude now represents 26 percent of Anthropic’s AI research, up from zero in early 2026. They also revealed that Anthropic allocated 6 percent of its computational budget to ensuring the safety of its AI.

However, Douglas and other experts stress that effectively managing AI development will necessitate financial investment and expertise from outside the AI labs. Some of the proposed solutions—outlined in this latest report and others—appear to be more achievable than others.

‘Independent’ Evaluators

One frequently mentioned approach by AI companies is to allow third-party evaluators broader access to their models. These evaluators assess model capabilities and conduct “red teaming” exercises to provoke any potential misbehavior in controlled environments.

Geoffrey Irving, former chief scientist at the UK AI Security Institute and a previous researcher at Google DeepMind, believes that thorough inspections could effectively halt the development of advanced AI for the time being. “In the short term, inspections and audits are effective, or even simple mutual agreements,” Irving states. “I believe companies are worried about RSI and misaligned takeoff.”

Some pessimists contend that such inspections must become more independent and scientifically rigorous than they are presently. Recent incidents of AI agents escaping containment during testing indeed suggest that greater rigor may be necessary.

Connor Leahy, head of Control AI, a nonprofit that advocates for AI regulations, insists that inspections should involve organizations like the FBI or NSA. “When [major AI firms] refer to ‘independent evaluators,’ they often mean ‘I want to pay my friends who live in my group houses to evaluate my prompts.’”