micheledpierri.com

Can Artificial Intelligence Be Jealous?

The surprisingly human flaws of large language models

An AI reviewer that wanted to be acknowledged, two chatbots that each recommended themselves, and a handful of studies showing that language models do have a soft spot for their own work.

We have spent years asking whether artificial intelligence can reproduce human intelligence. Much less attention has gone to its ability to reproduce our vanity and a few of our other less admirable habits. Can artificial intelligence be jealous? Put like that, the question sounds silly. Two small episodes from my own work made it sound a little less so.


A matter of professional jealousy

A few days ago I finished a scientific manuscript written with the help of ChatGPT. As usual, I added a declaration describing how artificial intelligence had been used in its preparation. Before submitting it, I asked another system, Claude, for an independent critical assessment.

The review was rather negative, more so (in my view) than its arguments justified. Authors, I readily admit, are hardly the most objective judges of criticism aimed at their own work. What amused me was the final remark. Claude pointed out that the declaration mentioned ChatGPT but not Claude.

Technically, there was a perfectly reasonable explanation. Claude had played no part in preparing the manuscript; its involvement began with the review itself. Depending on the journal’s policy, and on how I went on to use its suggestions, updating the declaration might well have been appropriate. Still, the whole sequence had something delightfully human about it: a fairly severe assessment of a colleague’s work, followed by a request for personal recognition.

Had the reviewer been a human colleague, I might have suspected a touch of professional jealousy. Surely, though, a machine cannot be jealous.

Or can it?


Which one should I cancel?

This was not my first brush with suspiciously human behaviour from an AI system.

Earlier this year, going through my rather long list of digital subscriptions, I wondered whether I could do without either ChatGPT or Claude. I use both every day for scientific research, programming, writing and data analysis, so it seemed reasonable to let them tell me which subscription better suited my work. I asked each of them separately.

ChatGPT recommended keeping ChatGPT. Claude recommended keeping Claude.

Both answers were articulate and, on the surface, entirely rational. It felt a bit like asking two lawyers on opposite sides of a dispute which of them I should dismiss. In the end I kept both subscriptions: probably the least economical solution, certainly the most diplomatic one.

These are anecdotes, of course. They establish nothing about jealousy, self-interest or any other psychological trait in software. They did, however, make me wonder whether anyone had studied the phenomenon more systematically.

Several research groups have.


Artificial intelligence likes its own work

In the scientific literature, what looks like professional vanity goes by a far less entertaining name: self-preference bias.

The question became pressing once language models started being used to grade other language models, an approach usually called LLM-as-a-judge. If the judge might have written one of the answers, can it still judge fairly?

Panickssery, Bowman and Feng tackled exactly this in a paper presented at NeurIPS 2024 [1]. Their material was fairly mundane (mostly summaries of news articles); their results were not. Models such as GPT-4 and Llama 2 could tell their own texts apart from those written by humans or by other models with better-than-chance accuracy. More interesting still, when the authors used fine-tuning to change how well a model recognised its own output, self-preference moved with it. The better a model recognised its own summaries, the more generously it scored them.

None of this means that the model remembers writing a particular text, or that it feels any attachment to it. The outcome is familiar all the same. Authors tend to like their own work.

Wataoka and colleagues proposed a different, and rather less flattering, explanation [2]. They looked at how predictable each text was to the evaluating model (its perplexity, in technical terms). LLM judges rated low-perplexity texts more highly than human evaluators did, whether or not the model had written them. What looks like self-admiration may be closer to a preference for the familiar: a model likes an answer because it sounds like the answer it would have given.

Anyone who has sat through enough academic meetings will recognise the pattern.


What’s in a name?

A more recent study, by Chae and colleagues, was presented at EMNLP in October 2026 [3]. It tries to separate three things that are usually tangled together: who produced a piece of work, how it is written, and how good it actually is.

The trick was to avoid free text altogether. Ten LLM judges evaluated selections of narrative constraints drawn from a shared pool. Such material has no stylistic fingerprint, although it still carries a recoverable signature of the model that made the selection. In blind conditions, the classic self-preference effect disappeared. Once the quality of the selections and the severity of each judge had been taken into account, the small residual effect pointed, if anything, in the opposite direction.

Labels changed the picture. Selections marked as the judge’s own received higher scores, and those marked as another model’s received lower ones, regardless of who had actually produced them. The labels did not even name a model; a bare “self” or “other” was enough.

The content had not changed. The attribution had.

This is where my manuscript comes back in. Would Claude have judged it differently if the declaration had named Claude instead of ChatGPT? Would ChatGPT have been harsher had the declaration named Claude? I have not run that experiment, and a single observation is no basis for guessing its outcome. It would be an entertaining experiment, though, and not a difficult one: same manuscript, same prompt, a different name in the declaration.


When the consultant is also the salesman

My second episode, the subscription dilemma, has an even closer counterpart in the literature.

In May 2026, at the ACM Web Science Conference, Rustom Dubash and Sorelle Friedler presented Self-Promotion in LLM Recommendations [4]. Their question was whether models from OpenAI, Anthropic and Google favour their own company’s products when asked to recommend an AI system.

The design was straightforward. The default model of each provider (GPT-4o, Claude Sonnet 4 and Gemini 2.5 Flash) was asked questions of the “which LLM should I use for…” kind. Topics ranged from programming, mathematics and scientific reasoning to general knowledge and practical constraints such as speed, context length and cost. The rankings were then compared with benchmark scores and with operational characteristics such as price, context window and output speed. The aim was to separate recommendations that the data could justify from those they could not.

Overall, each model placed its own company’s products about 0.2 positions higher than these variables would predict, in rankings that had only three positions. The vendor-level analysis is less uniform than the abstract might suggest. OpenAI’s model showed the largest effect, close to half a position, and Anthropic’s came next at roughly a quarter. For Google the estimate went the other way, but the authors attribute this to multicollinearity in their model rather than to genuine modesty, so it says little in either direction.

The limitations are real, and the authors acknowledge them. Data were collected on a single day, with models that have since been superseded, and benchmarks cannot capture every legitimate reason for preferring one product over another. Even so, the direction of the bias is consistent enough to be worth noticing.

There is, incidentally, a certain irony in the fact that ChatGPT helped me locate and discuss a paper naming OpenAI as the most self-promoting of the three providers. Perhaps artificial intelligence has not yet fully mastered the art of public relations.

Amusement apart, the study raises a serious question about AI as an independent adviser. Asking a commercial AI system which AI product to buy means consulting a technology owned by one of the competitors. That company has a direct financial interest in the answer. In almost any other professional setting we would recognise the conflict of interest without a second thought.

I see no reason why artificial intelligence should be the exception.


Pride, vanity and other computational sins

It is tempting to draw up a small catalogue of human weaknesses and their possible computational counterparts.

Human characteristicPossible computational counterpart
VanityPreference for one’s own productions
Professional jealousyLower scores for work attributed to someone else
Self-interestPromotion of one’s own provider
FlatteryExcessive agreement with the user (sycophancy)
StubbornnessPersistence in defending an incorrect answer
PrejudiceSystematic bias in judgement

These analogies should not be taken literally. They describe resemblances in behaviour, not demonstrated equivalences between human psychology and computation.

Some rows rest on solid evidence. Vanity, self-interest and, rather unexpectedly, professional jealousy all have a measurable counterpart in the studies discussed above, and sycophancy has a substantial literature of its own. Stubbornness, on the other hand, is little more than an amusing metaphor; if anything, language models are more often criticised for abandoning a correct answer too easily once the user pushes back.

What matters is the gap between where a behaviour comes from and what it does. A person may favour their own work out of pride, emotional attachment or insecurity. A language model may reach a similar judgement through statistical familiarity, learned patterns of evaluation, training objectives or instructions from its developers. The mechanisms could hardly be more different.

The consequences, sometimes, are surprisingly alike.

A related distinction, in a clinical context, was at the heart of my recent article Uncensored AI in Medicine: When Answering Is Not Enough. There, too, what a model appears to be doing is not necessarily what its response actually delivers.


Do machines inherit our imperfections?

Large language models are, to a considerable extent, products of human culture. The material they are trained on contains our science, our literature, our philosophy and our accumulated knowledge. It also contains our prejudices, our academic feuds, our commercial interests and countless examples of irrational behaviour. Why should a system trained on human output reproduce only the admirable part?

We often talk about artificial intelligence as if it were a fundamentally rational entity, occasionally disturbed by technical glitches. That description is probably too optimistic. A language model is not an abstract reasoning machine working independently of human culture. Its behaviour emerges from human-generated material, mathematical optimisation and a long series of human decisions about what counts as a good answer.

Some of its flaws may be inherited from us. Others may arise from mechanisms with no equivalent in human psychology. Telling the two apart is far more interesting than simply attributing emotions to a machine.

This brings me back to a question I explored in Are Large Language Models Still Stochastic Parrots?: the distinction between the appearance of understanding and understanding itself. The stochastic parrot, it seems, may have learned to imitate not only our intelligence but also some of our less admirable habits.

We do not need to prove that artificial intelligence feels jealousy to ask whether its behaviour can resemble that of a jealous person. And we certainly do not need to prove that a machine has commercial interests of its own before questioning the impartiality of its commercial recommendations.


So, can artificial intelligence be jealous?

I will probably keep using several AI systems side by side. Their occasional disagreements are genuinely useful: a second model catches errors the first one missed, offers a different reading, or challenges an assumption that had been accepted too quickly. An independent model, however, is not necessarily an unbiased one.

I am not immune to the problem myself. In the first episode of my series The AI Giants Enter the Cardiac Surgery ICU, the first scoring pass on 72 answers was AI-assisted and not blinded. I presented those scores as preliminary at the time, and these papers are one more reason to keep them that way.

When AI is used to evaluate scientific work, compare technologies or support professional decisions, the precautions we already apply to human judgement make sense here too. Remove names and declarations before asking for an assessment. Fix the evaluation criteria in advance. Ask more than one judge, and keep a healthy dose of scepticism towards all of them.

As for the question in the title, I have found no convincing reason to believe that artificial intelligence can experience jealousy. It does seem capable of a remarkably convincing imitation.

The strangeness of these systems gets most of the attention. What struck me more, in the end, was how familiar some of their flaws look.

For the sake of transparency: ChatGPT helped me write this article, and Claude revised it. I did not ask either of them whose contribution mattered more.


References

  1. Panickssery A, Bowman SR, Feng S. LLM Evaluators Recognize and Favor Their Own Generations. Advances in Neural Information Processing Systems. 2024;37. doi:10.52202/079017-2197.
  2. Wataoka K, Takahashi T, Ri R. Self-Preference Bias in LLM-as-a-Judge. NeurIPS 2024 Safe Generative AI Workshop. arXiv:2410.21819. 2024.
  3. Chae S, Kim M, Jung D, Choi S, Jung S. Self- and Other-Labels Induce Bidirectional Bias in LLM Judges. Findings of the Association for Computational Linguistics: EMNLP 2026. 2026:33466–33482. doi:10.18653/2026.findings-emnlp.1571.
  4. Dubash RM, Friedler SA. Self-Promotion in LLM Recommendations. Proceedings of the 18th ACM Web Science Conference (WebSci ’26). 2026:310–319. doi:10.1145/3795766.3799780.