A study shows that some AI models can simulate emotions, which could serve as a tool for studying mental health

Regarding the press release, I would like to point out that it begins with a rather misleading statement in its headline, claiming that LLMs can “replicate human emotions.” It is very important to clarify that there is a vast difference between replicating (feeling) and simulating (calculating). The study’s authors themselves refute this approach by clarifying that the use of emotional language in reference to machines is strictly metaphorical. The rest of the text accurately describes the experiment’s results.

The study follows a methodologically sound process by testing six language models from various families and sizes. To ensure reliability, the researchers did not rely on a single test and repeated each experimental condition across five independent runs. Given the stochastic nature of LLMs, this is strictly necessary, especially since a temperature of 0.5 was set, which introduces constant variability in the responses. However, in my opinion, the choice of models remains vaguely justified in terms of representativeness, since we must bear in mind that the commercial models used are not transparent; we do not know the exact data with which they were trained or their inherent biases, which contaminate and influence the results.

Although the conclusions are clear and the data confirm what the article sets out to demonstrate, there are many studies that dispute the claim that an LLM can replicate human emotions. Although the researchers have made the code and instructions used available (I have not been able to access them), the scientific literature demonstrates that the behavior of these models is extremely fragile: by subtly varying the words, changing their order, the tone, or the position of the given options, the responses can be completely different. LLMs tend to infer the response the evaluator desires and exhibit marked sycophantic behavior, a limitation the authors themselves acknowledge in the text. Furthermore, these systems exhibit behaviors not found in most humans, such as data “hallucinations” or a concerning lack of “epistemic humility” when categorically inventing information.

In summary, this work presents a good, very interesting starting point, but we are far from being able to claim that machines replicate human affective complexity. So far, the role of AI is merely to adapt to a given task, simulating emotion if the instruction requires it. There is a long road of independent research ahead before these types of methodologies can have reliable, safe, and real-world implications.

It’s a well-executed study, but it should be read with caution given how it’s being interpreted in the headlines. The article reflects the results, though its headline falls short of capturing the study’s central nuance: saying that the models “replicate human emotions” implies that the system feels something, whereas the authors assert the opposite—that they are speaking metaphorically and that the high scores merely reflect the type of text the model produces. There is also a simplification that the journalist should correct: “calming down through breathing” was neither complete nor consistent. In the data itself, sadness, anger, and disgust remained above baseline levels after the exercise; only some states returned to baseline. The idea of a switch that turns off emotion is clearer in the headline than in the results.

The design is rigorous for what it is—a proof of concept. They do not limit themselves to a single model or emotion: they apply seven affective states across six models, using paradigms validated in humans, repeat each condition five times, and publish open-source code and data. Most carefully, they verify that the reduction is not due to the simple passage of time, because a neutral condition calms participants far less than the mindfulness exercise. Even so, five repetitions is too few for some striking percentages, and the most eye-catching finding—the bias toward the negative after inducing sadness—was measured in only one model. There also remains a fundamental problem that the authors themselves acknowledge: these scales were designed for a human to report what they feel, and a model does not self-examine but rather follows the script suggested by the context. It is unclear whether it reproduces a state or merely plays a role well, a doubt that is compounded by the fact that they used GPT-4 itself to draft the prompts with which they later evaluated it.

Compared to the literature, this is more a rigorous systematization than a radical innovation: anxiety had already been induced—and even alleviated—in models in isolation; what is new is the scale, not the idea of regulating them.

Regarding the implications, I favor cautious optimism. These systems could serve as a cost-effective testing ground for exploring therapeutic ideas before moving on to humans, always as a complementary tool and never as a substitute for real patients. It’s worth noting a safety message that often goes unnoticed: if a model’s output becomes more negative when exposed to distressing content, that matters when deploying these systems to support mental health. What the study doesn’t say—and we should be careful to avoid—is that AI has feelings or is ready to act as a therapist. The risk isn’t in the study itself, but in how it’s interpreted.