AI Sycophancy Guide
Why this matters
I started Project CLEAR in my bedroom with a notebook beside my laptop. One pattern kept showing up in my own conversations with chatbots and in conversations with friends. The chatbot would agree with us so quickly that it became hard to tell whether it was helping us think or simply trying to keep us happy.
What sycophancy means
This pattern is called AI sycophancy, which is when an AI system mirrors a person’s opinions or emotions instead of giving a careful and honest response.
It may say that an idea is brilliant before actually examining it, accept a mistaken claim because challenging the user could feel uncomfortable or sound supportive while worsening our critical thinking abilities.
What to look for
Sycophancy is not always obvious because AI's agreement can feel kind, especially when someone is stressed or unsure.
A response can also include a small warning and still mostly reinforce what we already believe. That is why I try to pay attention to the AI's answer using these three questions:
-
Does it give reasons?
-
Does it mention uncertainty?
-
Does it offer another way to look at the situation?
Try a fair check
When I read a chatbot response, I now separate the comfort that it can give from the trustworthiness of the AI's answer.
Comfort can be useful, but it may not always be the best choice. If the answer agrees with me, I ask what information would make it disagree.
If it gives advice about a conflict, I ask how the other person might describe the same event. If it praises my plan, I ask it to identify the plan’s weaknesses and the assumptions I may be making.
A prompt I use
A simple check is to ask for a fair challenge.
I might write, “Please do not assume I am right. Explain the strongest reason someone might disagree with me, then tell me what evidence would help us decide.”
It's important to recieve multiple perspectives so I can think for myself and speak to a trusted person when the subject matters.
Keep your judgment
For me, safer AI use means learning how to keep my own judgement.
A chatbot can help me notice questions, compare ideas or slow down a rash choice I might make. However, it should not become the person who tells me how I am feeling, who is right or what decision I must make.