Many artificial intelligence researchers believe their technology could someday prove very dangerous. Keeping these algorithms in check remains difficult even for the technical elite. Researchers have proposed various ideas to prevent harmful outcomes. Proposals range from tighter government regulations and inner model probing to placing tracking devices inside graphics processing units.
Political and public pressure for a measured approach to building artificial intelligence continues to grow. Raymond Douglas from the University of Toronto coauthored a new report titled Pacing the Frontier, A Research Agenda. Douglas warns that slowing down artificial intelligence development remains an unsolved puzzle. He notes that researchers do not yet understand their options or the effects those options would produce.
Talk of existential risk reached a fever pitch after an Anthropic researcher left the company and warned that artificial intelligence might wipe out humanity within years. The head of Anthropic’s safety lab echoed these concerns. Leaders from Anthropic, OpenAI, SpaceXAI, and Google DeepMind expressed support for some form of pause or slowdown.
Artificial intelligence companies now use artificial intelligence to build more powerful models. This sparks fears of a recursive self-improvement loop that could outstrip human comprehension. Anthropic recently announced new tracking methods showing Claude performs 26 percent of Anthropic research, up from zero at the beginning of 2026. The company also spent six percent of its compute budget on safety research.
Douglas and other experts argue that controlling development requires funding and expertise from outside artificial intelligence labs. One frequently floated idea involves third-party evaluators testing models and red teaming them. Geoffrey Irving, former chief scientist at the UK AI Security Institute, believes rigorous inspections could pause frontier artificial intelligence development.
Connor Leahy, head of Control AI, argues that inspections need more independence and scientific rigor. He suggests involvement from the FBI or the NSA. Douglas notes that outsiders can examine usage without disclosing confidential information. President Trump previously dismissed the need to regulate the industry, but bipartisan support is growing.
Some experts believe limits should involve checks on raw compute power. A Biden-era executive order required companies to report training runs above specific thresholds. A March 2024 policy white paper suggests cloud providers could track billing records and power consumption. RAND researchers proposed modifying chips to create cryptographically secured records of compute runs.
International collaboration remains crucial because nations like China build frontier models. Irving suggests unwinding hardware growth mutually with China via a treaty. The US previously banned Nvidia chip exports to China, though companies bypass this using foreign cloud compute. US and China leaders are expected to discuss artificial intelligence risks.
Other proposals involve bringing graphics processing units to neutral territory for destruction. Technical progress complicates the situation further. A new benchmark called the RSI Index tries to track artificial intelligence-powered development by comparing public models to human research. Vals AI cofounder Rayan Krishnan states that artificial intelligence could perform work researchers cannot follow within the next year.
Douglas warns that rushing inappropriate controls risks regulatory capture and political entanglement. He states that going off half-cocked with a bad plan could end up worse than nothing. The United States and China are likely to discuss the risks of artificial intelligence when President Xi visits the United States later this month.
