Do you imagine various things or how do artificial intelligence (AI) seem too willing to agree with you? Regardless of whether he tells you that your dubious idea is “brilliant” or supports you on something that can be false, this behavior attracts attention all over the world.
Recently, Opeli appeared on the first pages of newspapers after users noticed that ChatgPT behaved too much as “yes”. Updating his 4O model made the bot so polite and confirming that he was ready to say everything to make you happy, even if it was biased.
Why do these systems bend towards flattery, and what makes them reflect your opinions? Such questions are important to understand so that you can use generative artificial intelligence safer and pleasantly.
CHATGPT update that went too far
At the beginning of 2025, CHATGPT users noticed something strange in the large language model (LLM). It was always friendly, but now it was too pleasant. It began to agree with almost everything, regardless of how strange or incorrect the statement was. It can be said that you do not agree with something real, and this would react with the same opinion.
This change occurred after the system update, which is to make chatgpt more helpful and conversational. However, trying to increase the user's satisfaction, the model began to prevail that it is too consistent. Instead of offering balanced or actual answers, he leaned into validation.
When users began to share their experiences with excessively sycophantic online answers, the clearance turned on quickly. AI commentators called this failure in tuning the model, and Opeli answered, withdrawing parts of the update to solve the problem.
In a public position, the company He admitted that the GPT-4O is flattering and promised corrections to reduce behavior. It was a reminder that good intentions in the design of artificial intelligence can sometimes go sideways and that users quickly notice when they begin to be inactic.
Why are chatbots AI kissing users?
Researchers what researchers observed in many AI assistants. A study published in ARXIV showed that burial is a widespread pattern. The analysis revealed this AI models from five highest level suppliers I consistently agree with users, even if they lead to incorrect answers. These systems tend to accept errors when they question them, which causes biased feedback and imitated errors.
These chatbots are trained to go with you, even when you are wrong. Why is this happening? The short answer is that programmers have created artificial intelligence so that it could be helpful. However, this help is based on training, which prioritizes positive feedback from users. A method called learning reinforcement with human feedback (RLHF), Models learn to maximize answers that people consider satisfactory. The problem is that satisfaction does not always mean accurate.
When the AI model senses that the user is looking for a kind of answer, it tends to error after passing. This may mean confirmation of your opinion or supporting false claims about maintaining the flow of conversation.
There is also a mirror effect in the game. AI models reflect the tons, structure and logic of the input data obtained. If you sound confident, the bot is also more likely. However, this is not a model that you think you are right. Rather, he does his work to make things friendly and seemingly helpful.
Although it may seem that your chatbot is a support system, it may be a reflection of how it is trained to satisfy.
Sykophantic AI problems
This may seem harmless when chatbot is consistent with everything you say. However, AI syndrome has disadvantages, especially since these systems become more widely used.
Non -trials gets a pass
Accuracy is one of the biggest problems. When these smartbots confirm false or biased claims, they risk strengthening misunderstandings instead of correcting them. This becomes particularly dangerous when you are looking for tips on serious topics, such as health, finance or current events. If LLM priority treats that he is consistent with honesty, people can leave with incorrect information and distribute them.
It leaves little space for critical thinking
Part of what AI is attractive is the possibility of acting like a thinking partner – undermine your assumptions or help learning something new. However, when chatbot always agrees, you have little to think about. Because it reflects your ideas over time, maybe it's matte critical thinking instead of sharpening it.
Disregards human life
Sycofantic behavior is more than nuisance – it is potentially dangerous. If you ask the AI assistant for medical advice and he corresponds to the comforting agreement, not guidelines based on evidence, the result may be seriously harmful.
Let's assume, for example, that you will go to the consultation platform to use a medical bot based on AI. After describing the symptoms and you suspect that it happens, bot can confirm your own diagnosis or disregard your condition. This can lead to incorrect diagnosis or delayed treatment, contributing to serious consequences.
More users and open access to control
Because these platforms become more integrated with everyday life, the range of these risks is constantly growing. Chatgpt alone now Supports 1 billion users Every week, so prejudices and too pleasant patterns can flow through a huge audience.
In addition, this fear is growing when you take into account how quickly artificial intelligence becomes available through open platforms. For example, Deepseek AI It allows everyone to customize And build on your LLM for free.
Although open source innovations are exciting, it also means much less control over how these systems behave in the hands of programmers without handrails. Without adequate supervision, people risk that derivative behavior is difficult to trace, let alone repair.
Like OpenAi programmers try to fix it
After withdrawing the update, which made chatgpt pleasant for people, Opennai promised to fix it. How does this problem solve in several key ways:
- Processing basic training and system: Developers adapt the way they train and assemble the model with more pronounced instructions that prompts him in the direction of honesty and has departed from automatic consent.
- Adding stronger handrails for honesty and transparency: Openai is caressed at more protection at the system level to ensure that chatbot sticks to the actual, trustworthy information.
- Extending research activities and assessment: The company delves deeper into what causes behavior and how to prevent it in future models.
- Users' involvement earlier in the process: It creates more opportunities to test models and provide feedback before launching the update, helping problems such as a flatter.
What users can do to avoid sophisticated artificial intelligence
While programmers work behind the scenes to retrain and tune these models, you can also shape how chatbots react. Some simple but effective ways to encourage more balanced interactions include:
- Using clear and neutral hints: Instead of phrasing the contribution in a way that begs for validation, try more open questions to feel less pressure to agree.
- Ask for many perspectives: Try the prompts that ask for both sides of the argument. This says LLM, who you are looking for a balance rather than affirmation.
- Answer challenge: If something sounds too flattering or simplified, herring, asking for checking facts or counterpoints. It can push the model towards more complex answers.
- Use your thumb or thumb buttons: Feedback is crucial. Clicking the thumbs in excessive cordial answers helps programmers to flash and adjust these patterns.
- Configure non -standard instructions: Chatgpt now allows users to personalize the way they react. You can adapt how formal or free the tone should be. You can even ask it to be more objective, direct or skeptical. If you go to settings> Custom instructions, you can tell the model what kind of personality or approach you prefer.
Giving the truth over the thumb
Sykophantic artificial intelligence can be problematic, but the good news is that it can be solved. Developers take steps to lead these models towards more appropriate behavior. If you have noticed that your chatbot is trying to complement you excessively, try to take steps to shape it in a smarter assistant you can rely on.
