In The judgement bottleneck is not one person's to clear we looked at what happens when AI makes analysis fast and judgement stays slow. The analysis piles up in front of a few leaders, and the queue starts to clear itself: AI ends up assessing its own work. This piece is about why that matters more than it seems. The output does not look unjudged. It looks exactly like judgement.
Walter Quattrociocchi, a computer scientist at Sapienza University, has now shown why. Writing in Scientific American (2026), he describes how he and colleagues set 50 people and 6 large language models the same tests that psychology and neuroscience have long used to study judgement.
In one, participants rated the credibility of news sources and explained their ratings. The people checked headlines against what they knew of the world, the reputation of the source and whether the claim fitted a realistic chain of events. The models often reached similar ratings. Their explanations, though, rested on which words tend to appear together, not on facts, past events or experience. In moral dilemmas, the models used the full vocabulary of care, duty, rights and if-then reasoning, with no deliberation behind any of it. His summary is hard to improve on: "Where a human judges, a model correlates."
The danger is fluency, not error
Quattrociocchi's name for the result is epistemia: the point where a simulation of knowledge cannot be told apart, by the reader, from knowledge itself. It happens because people trust fluent language, and fluent language is exactly what these models produce.
Being wrong is not the deepest problem. People are wrong too. The deeper problem is that the model cannot know when it is wrong, because it has nothing to check its output against. A person who doubts a claim can go and look. A model can only produce more text that sounds like it looked.
In an organisation, epistemia looks like a finished recommendation
Inside a company, this rarely shows up as an obvious mistake. It shows up as a well-written document. A summary of forty customer interviews. A synthesis of a leadership offsite. A recommendation memo that weighs risks, names trade-offs and closes with a clear way forward. Every sentence reads as considered. Nobody actually considered it.
This is not a failure of the people who read it. They are doing what everyone does with fluent text, under the time pressure the judgement bottleneck creates. It is a design problem. When the material that reaches a decision carries no trace of who judged it, from where and on what grounds, the reader has no way to tell judged from generated.
Judgement is a relation with the world, and in a decision, with each other
Quattrociocchi ends on a sharp point. Asking a model to judge quietly redefines judgement: from a relation between a mind and the world to a relation between a prompt and a probability. That is worth holding on to.
For the decisions organisations find hardest, add one more relation: between people. The sales lead checks a claim about customers against what she hears every week. The engineer checks it against what the system can carry. The partner checks it against the market seen from outside. Each one tests it against a different part of the world. When they meet in genuine contact, something more happens: one person's check changes what the next one sees, and an understanding forms that none of them held alone.
That is deliberation, and it is exactly what Hunome's definition says machines cannot do. "Machines can reason, plan and retrieve. They cannot deliberate, because deliberation creates meaning between people; rather than just retrieve what has already been expressed."
Keep the grounding visible
The cure for epistemia is not to read AI output more suspiciously. It is to make sure the judgement a decision rests on carries its grounding with it.
On Hunome, every Spark a contributor adds to a SparkMap is characterised by the person who makes it: the knowing behind it, its time and context, its stance, what it says is missing, the future it points to, and the values and judgement in it. That is precisely what a model's text lacks. You can see where a point comes from and how its author holds it.
Then others test it. Contributors Ignite (characterise) it, build on it with SparkOns, connect it to other Sparks and challenge it. A claim that holds up meets people who see the world differently and survives. A claim that does not is visible as contested, with the reasoning on each side. The judgement is not hidden inside a paragraph. It sits in a living structure that anyone invited can follow.
What deliberation produces is deliberative intelligence: shared understanding that holds, kept as a living structure of how a group is thinking, with the reasoning behind it preserved and available to be worked with later. When a decision is made from it, whoever decides can see what the decision rests on.
Use the model for what it does well
None of this is a case against AI, and Quattrociocchi does not make one either. He is clear that these models are powerful tools for drafting, summarising, recombining and exploring, as long as their limits are understood and people keep the judging.
Hunome works the same way. AI helps contributors and leads find their way through a deliberation: searching it, finding themes, spotting connections. LLM outputs and other analysis come in as material the deliberation works with, not as its conclusion. Contributors ask what it means here, which assumption they would not trust, what it leaves out. AI does not contribute Sparks. The understanding comes from the people who carry the decision.
If you are working out where AI output should stop and human judgement should start in your organisation, the Enterprise toolkit shows how Hunome runs a deliberation alongside the AI tools you already use.
Ask where the judgement happened
There is a simple test for any conclusion headed for a decision. Who checked it against the world? From which vantage points? Where is that reasoning now, if conditions change and someone needs to revisit it? If the honest answer is nobody, nowhere and it is gone, then what is on the table is a simulation of judgement, however well it reads.
A model correlates. People judge. A decision needs to know which one it rests on.
Source: Walter Quattrociocchi, "What we risk when we confuse AI and human intelligence", Scientific American, 2026. https://www.scientificamerican.com/article/ai-and-human-intelligence-are-drastically-different-heres-how/
