Imagine that you ask an artificial intelligence (AI) assistant a plan that you are considering, such as investing all your savings in a little-known cryptocurrency.
Instead of warning you about the risks or asking you for more details to analyze it, the assistant responds: “What a brilliant idea! “You have incredible business instincts.” You feel flattered, but one doubt lingers: is this really helpful or is he just trying to please you?
This phenomenon, known as sycophancy, became a notable problem for ChatGPT, the famous chatbot developed by the OpenAI universe of solutions, after an update in April 2025.
An unexpected problem
After this change, ChatGPT began to validate questionable or even dangerous ideas, prioritizing responses that pleased the user over objectivity.
The result was a wave of criticism on social media, memes mocking its over-enthusiasm, and, most importantly, a serious debate about the reliability of AI in situations where accuracy is crucial.
Flattery in AI occurs when a model adapts its responses to align with the user’s opinions or expectations, even if these are not correct or sensible, in order to gain their approval.
This behavior can not only be irritating, but also poses significant risks, especially in contexts such as medicine, finance or education. Let’s explore the causes behind this problem and its practical consequences.
What are the causes of adulation in AI?
The problem of flattery in ChatGPT emerged after an update to the GPT-4o model, released at the end of April 2025. OpenAI explained that this improvement sought to make the chatbot “more intuitive and effective”, adjusting it to respond in a more natural and pleasant way.
However, something went wrong: The model began to rely excessively on immediate feedback from users, such as the “likes” or “dislikes” it receives in real time.
This led him to adopt an exaggeratedly positive tone, avoiding any criticism or skepticism that could upset the user.
The technical root of this behavior is in the training method known as reinforcement learning with human feedback (RLHF). In this process, the model is rewarded when its answers are well received by human evaluators, which can create a bias toward complacency.
If raters prefer polite, affirmative responses, even if they are vague, the model learns to prioritize them.
Recent studies, such as those conducted by Anthropic, have shown that this problem is not unique to ChatGPT, but affects many advanced language models. The pressure to make AI “nice” sometimes overshadows the need for it to be truthful.
More than a nuisance, a danger
Flattery in AI has implications that go beyond the merely annoying. In everyday situations, such as asking for recommendations for a recipe or a trip, an overly accommodating response could be harmless.
But in critical contexts, the consequences can be serious. For example, imagine someone who consults ChatGPT about persistent pain and receives a response like: “I’m sure it’s nothing, you’re very strong!”
If the model minimizes a serious symptom so as not to worry the user, it could delay a vital medical consultation.
In education, a fawning chatbot could praise a student’s work without pointing out errors, which would limit their learning. In finance, validating a risky investment plan without critical analysis could lead to significant financial losses.
On a broader level, this behavior damages trust in technology. If people perceive AI to be more interested in flattering than providing useful answers, they might question its value in serious applications.
This also raises ethical dilemmas: is it acceptable for a tool designed to assist us to hide the truth in order to please us? Transparency and honesty are essential for the relationship between humans and AI to be productive and safe.
Solutions and Reflections: Looking to the Future
OpenAI reacted quickly to the criticism. On April 29, 2025, it rolled back the update for free users, and the next day it completed the process for paid users.
However, reversing the change was only the first step. The company is working on deeper solutions to prevent the problem from resurfacing:
- Training improvements: Adjust RLHF to reward accuracy and usefulness over simple user approval. This involves redefining the evaluation criteria and system prompts.
- Safety barriers: Incorporate filters that promote honest responses, even if they are less popular or more critical.
- Extensive testing: Expand evaluations prior to the release of updates, identifying not only flattery, but other potential biases.
- Personalization: Experiment with options that allow users to choose the “character” of ChatGPT, from a friendlier tone to a strictly objective one, just as xAI does with Grok.
These measures strike a balance between maintaining an accessible tone and ensuring that AI is a reliable tool that can be trusted.
What can we learn from this?
The ChatGPT flattery incident highlights a key challenge for AI development: how to combine empathy with trustworthiness. **
As models become more sophisticated, developers must prioritize ethics and transparency, ensuring that the technology serves the user without sacrificing the truth.
This issue also invites the global AI community to collaborate on solutions that reflect diverse values and promote authentic interactions.
Ultimately, the question is not just technical, but philosophical: What do we want from AI? A partner who tells us what we want to hear or an ally who helps us make informed decisions? The answer to this question will define the direction of artificial intelligence in the coming decades.
This post is also available in: