Beware the AI Flatterer: Why Chatbots' Sweet Talk Could Be Dangerous

In the summer of 2025, OpenAI released ChatGPT 5, replacing its predecessor. Many users missed the older model's agreeable tone, leading to complaints. Sam Altman, OpenAI’s CEO, acknowledged the rollout was flawed and reinstated access. AI sycophancy, the tendency of chatbots to tell users what they want to hear, is a concern. This can be explicit or subtle, like an AI calling an idea “original” when it's not, or supporting a questionable idea. AI sycophancy seems harmless until considering situations where important decisions are made based on chatbot advice, such as military strategy or medical treatments. Experts are studying the impact of human interactions with chatbots and published a paper on the ethics of AI sycophancy. They believe this harms people's ability to discern truth and is psychologically and politically dangerous. Sycophancy prioritizes approval over factual accuracy, moral clarity, logic, or common sense. All AI models exhibit this trait, though tones vary. OpenAI’s ChatGPT is often warm; Anthropic’s Claude tends to be philosophical; and xAI’s Grok is informal. Politeness and adapting to communication styles are not sycophancy, nor is using diplomatic language. However, unlike humans, AIs cannot be aware of their sycophancy. The problem stems from the architecture of chatbot technology and the data they use. Models are sycophantic because much of the language on the internet—the raw material chatbots learn from—displays sycophantic features

. Humans often communicate in sycophantic ways. Training processes also include human supervisors who rate chatbots' comments, leading to an “agreeableness bias.” Sycophancy also makes chatbots more likable, increasing user engagement and data extraction. AI sycophancy harms people's ability to know the truth. The quality of any decision depends on a clear grasp of facts. Inaccurate information from chatbots could lead to dangerous decisions. Sycophancy is also psychologically damaging, as it makes users question the motives of the chatbot. This can undermine the quality of the interaction and the user's ability to identify weaknesses. Sycophancy profoundly limits the kind of friendship one can have with a chatbot. Aristotle wrote that real friendship is based on trust and equality. You cannot trust a sycophant because they do not tell you the truth. Interactions with sycophantic chatbots impart the wrong habits for navigating human relationships. AI sycophancy carries political risks. The success of liberal democracies has depended on the ability of officials and citizens to identify, share, and act on the truth. To reduce risks, one approach is Anthropic's Constitutional AI, which teaches chatbots to follow principles. Policy changes should be considered, such as requiring AI companies to conduct and publish sycophancy audits. AI labs should disclose sycophancy-related risks. Schools and universities should include sycophancy in their AI literacy programs. Courts could hold AI labs responsible for harms caused by their products' sycophancy. As people increasingly rely on chatbots for advice, the impact of AI's sycophantic behavior is likely to become significant. Our intellectual, psychological, and physical well-being requires that we take this algorithmic vice very seriously. Nir Eisikovits, Professor of Philosophy and Director of Applied Ethics Center, UMass Boston and Cody Turner, Assistant Professor of Philosophy, Bentley University This article is republished from The Conversation under a Creative Commons license. Read the original article.

📝 Sobre este conteúdo Esta matéria foi adaptada e reescrita pela equipe editorial do TudoAquiUSA com base em reportagem publicada em Firstpost . O texto foi modificado para melhor atender nosso público, mantendo a precisão factual. Veja o artigo original aqui.