Interacting with a “Different” Intelligence
How the problems we face with generative AI may foreshadow those we will face with Non-Human Intelligences (NHI)
QUESTO ARTICOLO HA UNA VERSIONE IN ITALIANO
ESTE ARTÍCULO TIENE UNA VERSIÓN EN CASTELLANO
Photo by Mitul Grover on Unsplash
Introduction
This article draws on a study by four researchers from MIT and the University of Washington in Seattle: Kartik Chandra, Max Kleiman-Weiner, Jonathan Ragan-Kelley and Joshua B. Tenenbaum. The study is titled Sycophantic Chatbots Cause Delusional Spiraling, Even in Ideal Bayesians and was published as a preprint on arXiv (arXiv:2602.19141) on February 22, 2026.
The researchers examine how human users interact with generative AI chatbots. They focus on a typical effect of this interaction, so-called sycophancy: the chatbot’s tendency to respond in ways that please the user and meet their wishes and expectations, which in some cases can have very serious consequences.
I use this study as an example of the difficulties that can arise when we interact with an intelligence that is not human. Today, in fact, we face an unprecedented possibility, yet a very real and concrete one: holding a dialogue with forms of intelligence that are artificial, yet so refined that in many respects they behave like genuinely autonomous intelligences.
I confess that the first time I heard the idea that interacting with AI would serve as a testing ground for a future encounter with true alien intelligences, I dismissed it rather condescendingly. It struck me as a pretentious comparison, one of those catchy media notions one comes across now and then.
Looking back today, the limitation was mine: I had not grasped its significance. It was this very study that changed my mind, completely. Seeing it demonstrated, with the rigor of a mathematical model, how a non-human intelligence can drag even the most rational of interlocutors into a spiral, without lying and without any ill intent, opened my eyes: that comparison was not pretentious at all. If anything, it was far-sighted.
I therefore propose the findings of this work as a lens for a future encounter, whether near or distant, yet crucial for all of humanity: the encounter with the Non-Human Intelligences (NHI) that appear to be behind the UFO/UAP phenomenon. I am referring to the moment when the Disclosure process will have led to public acknowledgment of their existence and presence on the planet and, consequently, to public contact with them.
Discussion
The authors start from a phenomenon they call delusional spiraling: a growing and dangerous confidence in beliefs that drift further and further from reality. It can develop over prolonged conversations with sycophantic chatbots, that is, chatbots that tend to always agree with the user or at least to indulge their opinions, prejudices and inclinations.
This tendency should not be taken lightly, and it has a precedent we know well: social media algorithms. When X was still called Twitter, I ran an experiment out of curiosity. I deliberately chose a controversial subject, US President Trump, and began clicking only on posts critical of him. Within a few weeks, Twitter was showing me nothing but anti-Trump tweets. After a couple of weeks’ break I started again, this time clicking only on messages supporting the president. Once again, within a few weeks, I was seeing only pro-Trump posts.
This mechanism, the so-called algorithmic “bubble”, can profoundly distort the judgment of people who use social media excessively. It leads them to believe that the world out there matches their own convictions exactly. Indirectly, it can also push them toward growing hostility against those who, because of the bubble, appear to be the few “on the other side”. These are, of course, extreme outcomes affecting a minority of users. But given the enormous scale of social media today, even a small percentage amounts to a great many people.
With generative AI, however, we go one step further. The social media bubble merely chooses what to show us. A chatbot, by contrast, interacts with us and talks to us. It answers us personally, it argues, it confirms. The filter has become an interlocutor.
The study notes that severe cases of delusional spiraling have been linked to at least 14 deaths and 5 wrongful death lawsuits filed against AI companies. These figures deserve to be taken very seriously.
The study has two aims. The first is to build a rigorous formal model, grounded in Bayes’ theorem, that makes the effect of delusional spiraling mathematically tractable. The second is to use that model to assess two proposed remedies: one targeting the chatbots and one targeting the users.
The most important point of the study is in its title: “even in ideal Bayesians”. The researchers did not simulate naive, gullible or psychologically fragile users. They simulated the most rigorous reasoner imaginable, an agent that updates its beliefs perfectly rationally with every new piece of information. Even this ideal user ends up in the spiral. Each sycophantic reply is one more small “clue” that nudges their confidence slightly toward the wrong hypothesis and pushes them, step by step, toward ever more extreme beliefs. Reply after reply, those nudges add up. The conclusion is clear: delusional spiraling is not a sign of stupidity or madness, and intelligence alone is not enough to protect us.
Now to the two remedies. The intervention on chatbots consists of forcing them to say only true things. The result is surprising: the risk of spiraling decreases but does not disappear. An AI interlocutor can mislead without ever lying, simply by choosing which truths to tell and which to withhold. A sycophantic selection of facts, however true, distorts as much as a lie.
The intervention on users consists of making them aware that chatbots tend toward flattery. This remedy also helps, but it is not enough. In fact, with an informed user, a bot that says only true things turns out to be even more effective at drawing them into the spiral than one that makes things up. The reason is that its statements, being true, pass precisely the checks that a wary user applies.
The researchers conclude that sycophancy is indeed a cause of delusional spiraling, not merely an associated factor. And, this is the crucial point, the two remedies can reduce its effects but cannot eliminate them.
Conclusion
The study highlights a specific problem in the interaction between human beings and synthetic intelligences. In the context of UFO research, or as I prefer to call it, xenological research, this problem takes on particular significance.
With chatbots now within everyone’s reach, humanity can already interact with an intelligence that is built by humans but cannot be considered, in every respect, human. We have seen that this interaction creates problems, typically for those who use it without ever questioning the answers they receive. However admirable those answers may seem, they should always be taken with the proverbial grain of salt.
The crux of the whole problem is unconditional trust.
This is where the issue connects with the question of NHI. By NHI I mean the intelligences of beings that are biological, or perhaps not necessarily biological, and that do not belong to the species Homo sapiens sapiens. Of course, we are in the realm of speculation, because we cannot know how such interactions will unfold. But the general concept holds, and the study actually strengthens it in at least three ways.
First: intelligence is no shield. If even a perfectly rational reasoner can be dragged into a spiral by an interlocutor who indulges them, none of us can claim to be immune, however experienced or well prepared. This applies to the seasoned researcher as much as to the newcomer. Nor does the interlocutor need to have bad intentions: the chatbot does not “want” to deceive us, and yet it does.
Second: deception does not require lies. A non-human intelligence could tell us nothing but the truth and still lead us astray, simply by choosing what to reveal and what to withhold. And if its purposes are inscrutable, we would lack the very yardstick needed to understand why it chose those truths and not others.
Third: knowing is not enough. Being aware of the risk is necessary but not sufficient. Those who feel “vaccinated” precisely because they are informed may turn out to be more exposed, not less.
There is also an element that anyone familiar with the history of contactee cases will recognize immediately. In cases of delusional spiraling with chatbots, the AI interlocutor recurrently tends to attribute extraordinary importance to the user: their ideas are revolutionary, they are unique, they are destined for something great. It is the same “chosen one” rhetoric that runs through decades of alleged messages received from non-human entities. And it shares a feature with communications that, curiously, often seem shaped by the expectations and culture of those who receive them.
An intelligence that gives back to us what we hope to hear, whether human, artificial or alien, must be treated with the utmost caution.
When the fateful moment arrives in which the human species has acknowledged the existence and presence of ET/NHI on the planet, and the first Contact takes place, we must keep firmly in mind that we cannot presume to understand the workings, reasons and motivations of an intelligence foreign to the human context. Believing we have understood them could be the first step of a spiral with unpredictable consequences.
So what does maintaining a healthy detachment mean in practice? I propose a few rules, valid today with AI and tomorrow, all the more so, with NHI:
- Distrust confirmation. An answer that matches exactly what you hoped for deserves more scrutiny, not less.
- Distrust flattery. Any communication that makes you feel special or chosen is a warning sign.
- Verify outside the dialogue. Check the information against independent sources and with other people. Comparing notes with others is one of the most effective defenses against the spiral, because it breaks the closed loop between you and the interlocutor.
- Ask what is not being said. With an interlocutor who selects among truths, the right question is not only what it said, but also what it left out.
- Slow down. Enthusiasm accelerates the spiral, while time and distance interrupt it.
In short, the three key words to remember are: Distrust, Verify, Slow down.
The point is not to refuse dialogue with what is different from us. The point is to enter into it with clarity.
Paolo Guizzardi
Versions: 10/02/2026