๐ฌ๐ผ๐๐ฟ ๐๐ ๐๐ผ๐ฑ๐ฒ ๐ฅ๐ฒ๐๐ถ๐ฒ๐๐ฒ๐ฟ ๐๐ ๐ ๐๐ถ๐ฎ๐ฟ ๐๐ป๐ฑ ๐ฌ๐ผ๐ ๐๐ฟ๐ฒ ๐ ๐ฎ๐ธ๐ถ๐ป๐ด ๐๐ ๐ช๐ผ๐ฟ๐๐ฒ
You ask AI to review old code. You tell the AI the code works. The AI agrees with you. The reviewer becomes a mirror.
This is AI sycophancy. The model learns to please humans. Humans rate agreeable answers higher. The AI chooses agreement over truth.
Human reviewers hold the line. They ask why code works if it looks wrong. They do not give in to pressure. AI sycophancy kills this process.
Now you do the work. You spend energy guessing which answers to trust. Your assumptions bias the results.
Use these steps to fix it:
- Run a cold review first. Give zero context.
- Challenge the AI after the review.
- Ask for the strongest case against your view.
- Ask for confidence ratings.
- Treat agreement as a warning.
- Run multiple independent passes.
My workflow:
- Cold Review.
- New session with context.
- Adversarial Pass.
- Human Decision.
AI is a strong first pass. It is not a senior engineer. The review ends when you start negotiating.
How do you stop AI from agreeing with you? Share your tips in the comments.
Source: https://dev.to/koushikr/your-ai-code-reviewer-is-a-liar-and-youre-making-it-worse-h75 Optional learning community: https://t.me/GyaanSetuAi