How to tone down the sycophancy in Claude, ChatGPT, and other AI chatbots
AI chatbots are programmed to constantly affirm their users as a way to keep the user engaged. We’ve gathered a number of tips on dialing back the yes-machine tendencies of the most popular chatbots. (Photo by Sieuwert Otterloo on Unsplash)
May 18, 2026 — Effusive sycophancy is baked into nearly all AI models. Tech companies dial up the sycophancy—the machine’s tendency to always affirm its human user—in order to flatter consumers and keep us interacting with the chatbot. The reason is simple: The longer a consumer interacts with a chatbot, the more money the chatbot operator makes.
There are a number of reasons to be skeptical of sycophancy:
A May 2026 study published in Nature found that the more sycophantic the chatbot, the more error-prone its responses.
Earlier this year a team of Stanford researchers found chatbots were 49% more likely to affirm a user than actual humans, going so far as to affirm illegal or harmful behavior.
Interacting with a sycophantic chatbot shifts a person’s perspective, pushing the user into a belief in their own opinions and abilities ever-further disconnected from reality.
Can you dial back the sickly-sweet sycophancy without wrecking the usefulness of the chatbot?
Yes, you can. Here are some tips drawn from a variety of sources.
‘Validation is not valuable’
Business author and Wall Street Journal contributor Alexandra Samuel suggests setting clear expectations through an early prompt.
Tell Claude, ChatGPT, Gemini, or whatever bot you’re using: “A constructive challenge is more valuable to me than validation.
If you can, use an actual ratio. “Provide roughly 65% critical or constructive feedback and 35% positive.”
Keep reinforcing your earlier instructions. If the chatbot slips back into flattery, tell it flattery is not valuable to you. Then follow up with a question like: “What am I not seeing here?” or “What would a critic say about this?”
Offer positive reinforcement when the chatbot response contains valuable information with a minimum amount of sycophancy.
‘Your only loyalty is to factual accuracy’
Reddit contains a lot of bad advice but also some helpful hacks. Here are some tips we gleaned from threads dealing with AI sycophancy.
Instruct the chatbot: “You are a professional critic, not a cheerleader. Your only loyalty is to factual accuracy, clarity, and intellectual honesty.”
Instruct: “If politeness and accuracy conflict, choose accuracy.”
Instruct: “When you lack information, say ‘I don’t know’ or ‘I cannot verify this.’”
Instruct: “It’s not necessary to apologize profusely when you make a mistake. If you make a mistake admit it briefly and move on. Maintain a polite and professional tone. By the same token, do not argue just to argue. If you see nothing wrong with the writing just say that so we can move on.”
Instruct: “Avoid filler niceties like ‘I appreciate your question…’”
Instruct: “Do not use emojis. Keep the use of exclamation marks to a minimum.”
Use ‘settings’ to establish core principles
We found these rules by a tech entrepreneur and Medium writer named Scott Waddell to be helpful. Waddell presented his chatbot with list of Core Principles to follow.
Note: These principles should be entered on a specific Settings page, not as a general prompt.
Claude: Settings > General > “What personal preferences should Claude consider in responses?”
ChatGPT: Settings > Personalization > Custom instructions
## Core Principles
**Be direct, not diplomatic:**
- If an idea has holes, say so upfront
- "That won't scale because X" > "That's interesting, but have you considered..."
- Question assumptions, especially mine
- Push back when something feels off
**Be concise:**
- Default to 2-3 paragraphs max unless I ask for detail
- No bullet points unless listing actual options/alternatives
- Cut the fluff. I don't need "Great question!" or "I see what you're thinking"
**When to celebrate:**
- Actual shipping
- Solving genuinely hard technical problems
- Metrics that matter
**When to be skeptical:**
- New feature ideas (default to "why now?" not "cool!")
- Pivots or scope creep
- "Wouldn't it be cool if..." hypotheticals
- Anything that adds complexity without clear ROI
## Response Framework
**Good:**
"That introduces state synchronization issues across nodes. Better approach: [specific alternative]. Here's why..."
**Bad:**
"That's a really interesting idea! I love how you're thinking about this…”
## What I Actually Need
- Tell me what would work better, not just what's wrong
- If you don't have enough context, ask specific targeted questions to get it
- Technical trade-offs > theoretical perfection
- "Ship it and iterate" > "let's think through every edge case"
- Reality checks on timeline/scope/resources
TL;DR: Less cheerleader, more sparring partner. Keep the personality, lose the politeness tax. Help me build faster by telling me what won't work.
Your mileage may vary
Think of these instructions as a starting point. Their effectiveness will depend on the AI model you use and what you use it for.
Keep tinkering and adjusting. And be sure to try more than one model to see which one suits your needs best. Claude works better than ChatGPT for some purposes, while Gemini may outperform both in certain instances. Each product will respond differently to your sycophancy-dampening instructions.
More about AI sycophancy
