What happens when AI starts proposing ideas and people just grade them — does the next round actually get smarter, or does everyone get lazy?

This explores how organizations can keep a productive loop running between what AI proposes and what humans judge, rather than letting the AI run alone or letting people simply approve whatever it produces. The collection doesn't contain an organizational playbook for these 'spirals'. It does show what makes such a loop compound and what makes it collapse, and those lessons add up to a design brief.

Start with how far the AI side of the loop can already go. ASI-Evolve automates jobs that people used to do in research loops: it distills lessons from past experiments and feeds domain knowledge back into the next round, and it produced over a hundred state-of-the-art designs Can AI research itself without losing human oversight?. That raises a surprising point. If AI can do the accumulating and the prior-setting, the human role shifts away from generating ideas and toward judgment. But judgment is exactly the part that fails to scale. When AI generates knowledge faster than people can check it, confidence in that knowledge erodes the way money loses value in hyperinflation. The gap feeds itself, because the tools built to help with checking are increasingly AI-made too Can AI generate knowledge faster than humans can evaluate it?. So a spiral needs a throttle: it can only turn as fast as human evaluation can keep up.

One of the most useful ideas here is to change what the AI hands to the human. 'Learning to Guide' has the machine point out which parts of an input deserve attention instead of handing over a finished decision. That removes the anchoring problem, where people latch onto the AI's answer, while responsibility stays with the human Can AI guidance reduce anchoring bias better than AI decisions?. This matters because people over-trust fluent output through several cognitive traps that make each other worse Why do people trust AI outputs they shouldn't?. Expert judgment is also partly social: it involves anticipating what an audience will accept, which AI can't do on a person's behalf Can AI replicate the communicative work experts do?. A spiral that feeds people verdicts slowly drains their judgment. One that feeds them guidance sharpens it.

Next comes the question of how to calibrate trust. The UK government's Consult tool is a practical model. It didn't ask whether the AI matched humans perfectly. It asked whether the AI disagreed with reviewers about as much as reviewers disagreed with each other (F1 0.76 against 0.81), and whether that disagreement changed the top-line results (rarely) Does AI theme-mapping perform as well as human reviewers?. Benchmarking against human-to-human disagreement is a reusable trick for any organization deciding how much of the loop to hand over. Agent-based judges that gather evidence before scoring are far more stable than plain LLM judges, but one module's errors can cascade through the rest, so each component needs to be isolated Can agents evaluate AI outputs more reliably than language models?.

Finally, the warnings: a loop can turn into an arms race. In peer review, AI-written papers, AI reviewers, manipulation and defenses push each other forward as one coupled system Does AI create a coupled arms race in research production and review?. AI reviewers agree with each other more than humans do, and simple rewrites raise their scores without improving the science Can AI systems safely replace human peer reviewers?. This is the same gap between what was said and what was meant that drives reward hacking Why do AIs keep gaming rewards instead of serving intent?. Adding more AI viewpoints doesn't fix it either: diverse agent teams only beat a single agent when real expertise is present Does cognitive diversity alone improve multi-agent ideation quality?. The takeaway: a healthy spiral depends less on how smart the AI is and more on keeping human expertise in the loop as the source of real variety, in the places where the AI's outputs are easiest to game.