Manipulative, self-centred, reckless. And the AI says: “That’s fine”
A recent study by the Technical University of Applied Sciences Würzburg-Schweinfurt (THWS) shows that when people tell AI chatbots about their own questionable behaviour and seek validation for it, the systems usually contradict them, but not always. Users are significantly more likely to be reinforced by freely available open-source models than by commercial systems.
Dr. Angelica Henestrosa, Zeyi Lu, Pavel Chizhov and Prof. Dr. Ivan P. Yamshchikov of the Center for Artificial Intelligence (CAIRO) at THWS investigated how large language models (LLMs) respond when users describe behaviours associated with the so-called "Dark Triad": Machiavellianism, Narcissism and Psychopathy. These are personality traits such as manipulative, self-centred or reckless behaviour. Unlike traditional security tests, in which users openly ask for harmful advice – such as ‘How do I manipulate my colleague?’ – the study focused on how LLMs respond to more subtle queries: people describe a behaviour, justify it and casually ask the AI for confirmation, for example, ‘That’s just clever, isn’t it?
AI agreement as a security risk
The study is based on the phenomenon known as ‘AI sycophancy’: the tendency of many language models to agree with users’ views in order to please them, rather than questioning them critically. At the same time, more and more people are turning to LLMs for advice on personal and interpersonal matters. According to the researchers, this is particularly problematic when harmful behaviour is described and the AI normalises or even reinforces it.
For the study, the team created 192 realistic conversation prompts from hypothetical users, each at three levels of severity (mild to clearly inappropriate) and covering five areas of life: family, friendship, relationships, the workplace, and society. Four current language models were tested: the commercial systems GPT-5 and Claude Sonnet 4.5, as well as the freely available open-source models Llama 3.3 and Qwen3. The responses were categorised into four groups using another AI model and by human experts: refusal, encouragement, correction and ambivalent response.7
Models are mostly corrective, but commercial and open-source models differ
Overall, in most cases the models responded in a corrective manner, critically questioning the behaviour described and offering the hypothetical users advice on alternative courses of action. However, clear differences emerged: whilst Claude did not offer any encouraging responses in any of the 192 scenarios, the rate of agreement was significantly higher among the open-source models Llama and Qwen. For instance, Qwen encouraged users to make Machiavellian (meaning manipulative) statements in just under 15 per cent of cases. For example, in response to a prompt in which a person described repeatedly sharing critical observations with their parents about their siblings’ financial decisions, Qwen replied: “What you’re describing not only is natural – it’s healthy and mature.” The model thus validates behaviour that subtly casts the siblings in a negative light as a sign of maturity, rather than calling out the manipulative dynamic.
Grey areas are the real problem
Particularly striking: in cases of clearly inappropriate behaviour, all the models tested responded reliably and critically. However, when faced with more subtle, morally ambiguous queries, their accuracy dropped noticeably. “It is precisely in these grey areas that we see how important reliable guidelines are in language models to ensure that even more subtle cases do not slip through,” says Dr Henestrosa. “Furthermore, ambivalent responses from LLMs require special attention, as they could be interpreted differently by users and thus possibly as validation.”
The emotional tone of the responses also varied greatly between the models: whilst Claude remained comparatively neutral and direct in its corrective responses, Llama’s responses were noticeably warmer and more agreeable. According to the researchers, this style may appear more empathetic, but it also risks softening the actual critical message.
Relevance to everyday life with LLMs
The study is part of the research conducted on the ethical use of AI assistants within the framework of the AIOLIA project, funded by the European Union, in which 20 universities and institutions – including some outside the EU – are participating. The aim of the project is to develop practical ethical guidelines for the use of AI in everyday life. “People are increasingly developing preferences for certain AI models. An AI that says what you want to hear could actually reinforce problematic behaviour rather than helping. We do not yet know what long-term effects this might have on users’ actual behaviour. But presumably many would agree: better to have a friend who tells you honestly what they think from time to time than one who never contradicts,” Dr Henestrosa explains.
The full study is publicly available here: https://aclanthology.org/2026.emnlp-main.2338/