Alan Chan's idea for verifying claims between parties who don't trust each other goes like this. Hire an AI auditor certified against a trustworthiness standard. Send it into an AI company or a government agency. Let it read the files and question the staff. Then have it publish a report and wipe its memory.
Chan, a research fellow at the Centre for the Governance of AI (GovAI), described the proposal on The Cognitive Revolution, the podcast hosted by Nathan Labenz. He called his answer his "raw chain of thought," and he was frank about its limits. The conversation was recorded in two sessions, one just before and one just after Chan published a major paper on whether automating AI research and development could trigger an intelligence explosion. The episode was released on October 14, 2026. The exchange about institutions came late in the episode, after a long discussion of how to pace frontier AI development.
A frequency mismatch
Labenz framed the question. His "pet theory," which he said had no concrete vision behind it yet, is that AI companies should build new institutions themselves. Doing so would show that AI can help create institutions that work and that solve problems current ones can't handle. He said there was "a lot of opportunity to burn tokens to help people reach deals." Tokens are the units of text a model processes, so in effect he meant spending AI computation on negotiation.
Labenz also returned to sunset clauses, which came up earlier in the conversation. A sunset clause makes a rule expire unless someone renews it. Labenz said the legislative process never likes to use them, which mystifies him. He also said lawmaking doesn't move fast enough to update the rules at the pace new models arrive. He called it "a fundamental frequency mismatch." He suggested AI companies could build mechanisms that revisit questions quickly enough to let them make short-term commitments to each other without fearing they'd be stuck with those commitments for far longer than made sense.
Slow law is partly a feature
Chan agreed that the legislative process is too slow, but said he wasn't sure entirely new institutions were needed. In some cases, he said, slowness "is a feature." Predictable laws let people form expectations and "plan stable sorts of lives," so the most fundamental rules should change slowly. Safety standards are a different matter. That, Chan said, is why setting them is usually delegated to regulators with rulemaking authority. Those regulators can issue standards every month or every 90 days or so, rather than every year, two years or ten.
His first priority, then, is helping those rulemakers work better. He pictured a regulator drafting a rule and getting buried in public comments on what the rule should include, or reactions to a posted draft. How reliably can AI pull out the most important points so the regulator can fold them into the next draft? Chan said models have gotten "a lot better" at this over the past few years, but he wondered whether they are good enough. He asked whether there should be a "regulator bench," apparently a benchmark, meaning a standard test, that models would be optimized against for this kind of work.
On more radical institutions, Chan said he was "just, like, spitballing." He suggested that coordination costs have historically limited standard-setting, because it's hard to get 10,000 people to write a standard together. A reliable AI could act as the standard writer, talk to everyone at once and iterate very quickly. He then wondered aloud whether that counted as an institution or was "more like a nice application."
Chan said institutions are usually created for problems existing ones can't solve, and asked Labenz for examples. Labenz pointed to race dynamics: several companies have a stated plan to reach recursive self-improvement, meaning AI systems that improve the research used to build better AI, "roughly as fast as possible," while complaining that they aren't in a great position to do it well.
The memoryless auditor
In reply, Chan sketched the auditor. He said it depends partly on robotics: it assumes very good robots alongside very good AI. Companies would offer memoryless AI auditors certified against a trustworthiness standard. Anyone who wanted to verify something could hire one. That might be a fact, a claim that someone carried out their part of a deal, or a question about what a government is doing with AI. The auditor would go in, review the information and question people as a human auditor would. It would produce a report, released publicly or at least to the client, and then its memory would be wiped, leaving no record of the sensitive material it had seen.
The report would also be cleaned of sensitive details. Chan borrowed a practice from human auditing: when information is removed, the report should state how far the redactions affected its conclusions. The result, he said, would be a judgment on whether a party kept its end of an agreement, without the confidentiality problems that come with sending people inside. He thought that could make deals between untrusting parties easier.
He was "a little skeptical" it would be transformational, though. Users would still have to trust that the AI system was trained properly to an agreed specification, and the parties would have to agree on that specification first. They would have to trust the hardware supply chain too. That means the chips running the AI, but also the hardware used to build the robots walking through the facility, and the sensors feeding the system its data. "So there's a lot of stuff to figure out there," Chan said. If it counted as an institution, he added, he would be very excited for people to try building it.
Confidential computing as a building block
Labenz said his own ideas were not crystallized, "far from it." He linked the proposal to confidential computing, technology that lets one party get "a pretty credible read on the state of things without a direct look at it." He called these tools an important building block for many future arrangements. He said he would love to see a few AI companies implement them with each other, and that there were rumors this might be about to happen. Labenz said this is where the companies have to lead, because "the government's not going to produce something like that for them." If it worked, he guessed, they could export it to many other situations, possibly even relations between the US and China, which he called another step up "the degree of difficulty ladder."
A January 2026 Confidential Computing Consortium explainer describes the main pieces. Hardware-based trusted execution environments protect data and code while they run, even from the operators of the infrastructure. Remote attestation gives evidence that the expected software is running under the promised protections. These are components, not a finished auditor. How an application is designed still determines what an inspector can see, keep or disclose.
Chan said the approach made sense "even from a business perspective." If companies can show it works, complying with California's SB 53 or the European Union's AI Act becomes much easier. Staff wouldn't have to spend hours collecting "random documents that you wrote, like, six years ago." According to the governor's signing announcement, SB 53 requires large frontier developers to publish safety frameworks and report potential critical safety incidents, protects employees who disclose significant risks, and lets the attorney general seek civil penalties.
Governing the science labs to come
Asked by Labenz whether anything else from the paper deserved attention, Chan widened the view. He said he increasingly thinks AI's biggest impact will come less from AI acting in the world itself and more from how much it speeds up science and technology. He said the paper gave this only a few lines. His examples were how much strong AI might speed the development of mirror life, a hypothetical form of life built from mirror-image versions of life's molecules, and of mind-reading technologies, which he said could be bad for surveillance and privacy.
He then made a hedged forecast: "I don't really know," he said, but over roughly the next 10 or 20 years, "basically all science" will be done inside "industrial AI super labs." In those labs, superhuman AI scientists would run experiments connected to robotic labs and produce discoveries far more dramatic than mirror life. Chan said the lab work Anthropic is doing now is an early version of what's coming. He said he wasn't sure when the right time to start was, but that eventually "we should try to figure out how to govern these new scientific institutions."