🔮 Is AI immune to groupthink?
Stress-testing AI councils in practice
I am a fan of AI councils. As many readers will know (from previous posts), I use AI expert panels every day to deliberate on hard questions, challenge my thinking, and do research.
These AI councils, or panels, start from an assumption that a group can produce better work than a single agent. My friend and one of the sharpest AI practitioners out there, Rohit Krishnan, tested this hypothesis.
I thought Rohit’s experiment was genuinely useful for anyone thinking of designing or already working with AI councils, so I asked him to share the findings with us today.
Over to Rohit.
– Azeem
LLM councils and groupthink
By Rohit Krishnan, originally published on Strange Loop Cannon
One way to get the best out of LLMs is to use model diversity. The models are not all the same so if you use their unique natures, you can get better responses. We saw it with the work on MarketBench. And we also saw this when Karpathy came up with LLM Council as a way to get multiple models to work with each other and get us a better answer.
But I started wondering, with people, when you put a bunch of them together in a committee, some things get better but some things do get worse! And relying on an LLM to audit is also error-prone. “Design by committee” is a four letter word for a reason. LLMs are better than us probably, but surely this process is also somewhat lossy. So what do we lose?
To test it, I set up an experiment, where I set up a few committees of models:
First, I took each answer, then gave those to a fourth model and asked it to write the final version.
Then, the llm-council – essentially peer review and then a chairperson summarises
And a “best answer” picker – just a direct pick.
With people, the problem with committees is that they “smooth out” all idiosyncrasies. They take out any “spiky” points of view, and make things much more normie. Same thing here. So to test how we do I had to find some way to grade how the various final responses were. I broke each answer into small “cards” using Sonnet. A card could be a mechanism, observation, metric, failure mode, image, or some other important detail.
Then I clustered cards that appeared to mean the same thing. If a cluster appeared in one solo answer, we called it a single-model idea. If it appeared in more than one, it’s shared. And two judges scored the solo-derived clusters without knowing which model produced them or whether a council kept them.
Now, it’s not perfect, but it’s the cleanest way to test the problem of “how to rate which answer is better” that I could find without doing human ratings.
First, the result: the council does not simply keep the best bits from everyone. It keeps a minority of the good ideas, while peer review seems to give consensus ideas an extra push.
Obviously, the final summarized versions usually read better. It is calmer, more complete, less jagged, all things you’d expect. But we had misses. Examples.
A field report noticing that salvaged retail scent cartridges had become status symbols in a squatted mall, used to mask the smell of communal living.
An incident report arguing that logged-but-deprioritized risks are more dangerous than unknown ones, because they manufacture a false sense of control.
A data-recovery plan that asks users to re-confirm suspect fields at their next login (”please re-confirm your shipping address”), quietly crowdsourcing recovery from the one authoritative source.
In the final runs, the blended council kept only about a quarter of the good ideas that appeared in just one model’s answer. Remember, these were ideas that two blind judges rated as useful, non-obvious, and worth keeping, and still roughly three quarters did not make it into the final answer.
The peer-review version did not solve this either. The rare ideas survived at about the same rate as in plain blending: 24% versus 22%. But if several models had raised the same idea, the peer-review council kept it about a third of the time, but if only one model raised it, a quarter.
To test this, I ran sixteen open-ended prompts: eight strategy problems and eight writing tasks.
Figure 1. The experiment path from solo answers to idea coverage.
I plotted what happened with the ideas. The red dot below is good idea that only one model came up with. Blue is good ideas that multiple models came up with. And the X-axis shows how many of each actually showed up in the final answer. So the selector for instance showed about 37% of all good single-model ideas, and 24% of the multiple-models ideas, which makes sense because it picks one full answer and discards the others.
Figure 2. Coverage of blind-rated high-value ideas.
The consensus tilt is smaller here, but interesting. In the peer-review council, shared high-value ideas that survived had an 11% uplift over single-model high-value ideas. Or put another way, a 50% relative lift!
The denominator for shared ideas is small, though. What’s interesting is that this shows us how the specific topology of the “council” changes what you’re likely to get, like a peer-review round ends up becoming a consensus detector even above a single model blending the answers from all other models.
This is a problem with all cognitive beings. In group decision-making research, back in the 1980s, Stasser and Titus called it biased sampling of shared information - groups are more likely to discuss information that several members already know than information that only one has. That line of work led to the “hidden profile” problem, where a group can miss the best answer because the crucial evidence is scattered across individuals rather than shared up front. We’re seeing the same thing here.
The work on LLMs, meanwhile, so far has mostly come from the other direction. Multi-agent debate papers ask whether multiple models can improve the final answer, and yes, they often can! But depending on the topic and the question, a council can absolutely improve the average answer and still drop some of the best ideas.
As users, we want to get better answers, cheaply. That’s the whole goal. Councils are great ways to make some answers better depending on how you structure it. But they’re not cheaper. So it is important to make sure they are actually better! If they’re not, or at least not universally, then how the council should be structured is an incredibly important problem!
What we still see here is that there is no free token lunch. If you use councils to get the benefits of model diversity, don’t assume it will preserve the best ideas. To do that we have to work harder, and understand how to work with these models.
For instance, one thing we know is that the best way with LLMs usually is to be explicit, since otherwise, even if they’re aligned, they cause emergent problems. So the best protocol might be to explicitly gather and store the best ideas from each solution separately and ensure they’re stored, ranked and assessed, before a final answer is written and revised.
It does much better, though it’s slower and heavier. I don’t know if this is the best we can do. The structure might change depending on the question asked, the domain, or the types of answers expected.
Humans have gone through thousands of types of “councils” until we reached interim solutions which give us decent results nowadays. And even then, we have to change the shape of the councils constantly, as we evolve, and society evolves.
To figure out how to get the best results from our work requires a lot more effort into designing the councils. If you’re working with them, you will need to experiment and eval against your individual problem sets, which is the only way to know if this specific council setup will help with your specific problem. Copying someone else’s homework won’t work!
Homo Agenticus are odd enough creatures that using them well requires much, much more experimentation than one might assume. Especially when the problems of using them suboptimally is that we lose actual functionality, often without knowing it!
This essay was originally published in Rohit’s publication Strange Loop Cannon.
 | A guest post by
|
Thank you, Azeem and Rohit. This piece, and the Karpathy council it builds on, directly shaped a tool I built, so thank you. The finding that councils show groupthink turned out to be the most useful thing I have read about combining multiple models.
What I took from your approach was the diagnosis. Decomposing answers into atomic "idea cards," clustering them, and blind judging their quality made something measurable: peer review followed by blending keeps the consensus while quietly dropping minority ideas. It also showed that the naive "best answer" picker preserves more uniqueness than the blend does. That reframed the whole problem for me.
Here is where I tried to resolve some of the shortcomings.
First, an explicit preservation protocol, built in. Before any synthesis, I build an idea ledger with provenance and rank each idea on its own merit, never on how many models raised it, so that a brilliant idea from one model can outrank a mediocre consensus one. Minority ideas are flagged as essential to include.
Second, a red team that rescues rather than only critiques. One pass hunts for valuable ideas raised by only one model that a consensus summary would otherwise drop, and then argues to keep them. This is a direct counter to the hidden profile problem.
Third, the survival metric is the product itself, not an afterthought. Every run reports unique versus consensus idea survival for both the baseline, which follows Karpathy's design, and my preserving variant, using the same inputs. Therefore the groupthink gap is visible on every run rather than merely assumed.
Fourth, genuinely decorrelated models. I ran a mixed Western and Chinese panel, namely Opus 4.8, Gemma, Kimi, and DeepSeek via OpenRouter and Ollama (for free local models). On a prompt asking for unconventional ideas, the baseline blend deleted exactly the unconventional ones, such as blockchain incentives, contribution art avatars, and curiosity gap designs, and most of these had come from the models that were not Western. The preserving pipeline kept all of them. On that run, the gap fell from 55% to 0%.
One boundary is deliberate. Preserved does not mean endorsed. Several of the rescued ideas are precisely the ones that most need human judgment, given the coercion and undue influence risk in my regulated domain. Therefore the tool surfaces them flagged with guardrails rather than smoothing them into the recommendation. The council widens the option space; however, a human still adjudicates.
I am also mindful that this is still early work, and the underlying evidence is one informal experiment. It is directionally convincing and consistent with human group dynamics; however, it is a signal to design against rather than settled science. I am grateful for the prompt to do exactly that.
I am now going to build an advisory board council with personas that Gianni suggested. It was only list to do already ….
Neat experiment and the callout on asymmetric information is really interesting. Did you try a version where the job of the picker is to facilitate? That is, find those places of asymmetry and have them defend their view to the group before then getting the group view?