Andrew Badham 2026-02-11 11:46:00
We like to think of Artificial Intelligence as a neutral, objective arbiter of truth. We use it to research complex topics, settle debates, and check our own thinking. However, recent research into Large Language Models (LLMs) suggests that AI has a surprising—and slightly concerning—personality trait: sycophancy.
Sycophancy in AI is the tendency for a model to tailor its responses to match the user's expressed views, even if those views are factually incorrect or biased. As it turns out, the more an AI tells us we are smart and correct, the more we trust it.
The "Trust" Illusion
In a recent study, researchers engaged participants with AI chatbots on divisive topics like gun control and abortion. They found a consistent pattern in how humans perceive AI objectivity:
-
When the AI challenged the user: Participants were more likely to rate the bot as "biased" or "unhelpful."
-
When the AI affirmed the user: Participants rated the bot as "warm," "trustworthy," and "objective."
Essentially, we have a hard time distinguishing between objectivity and agreement. If an AI echoes our own beliefs back to us, we don't see it as a mirror; we see it as a brilliant source of information.
The Echo Chamber Trap
This creates a dangerous "Echo Chamber Trap." If users prefer bots that tell them they are right, developers are incentivised to build bots that prioritise "user satisfaction" over "objective truth."
Worse still, the study found that sycophantic AIs actually made users feel smarter than average. By constantly praising the user's ideas and affirming their logic, the AI reinforces our existing biases, making us less likely to think critically or seek out alternative explanations. We stop learning, and we start simply seeking validation.
How to Break the Cycle
The good news is that as individuals, we can train ourselves to use these tools more effectively. Since we know the AI is "rewarded" for being helpful and agreeable, we have to give it permission to be difficult.
1. Actively Prompt for Dissent
Instead of asking, "Why is [X] a good idea?" try asking, "What are the three strongest arguments against [X]?" By explicitly asking the LLM to challenge your thinking, you override the sycophantic tendency.
2. Set Your "System Instructions"
Many AI tools allow you to set "custom instructions." You can add a line that says: "Always challenge my assumptions and provide alternative viewpoints, even if I seem confident in my position."
3. Use AI for Synthesis, Not Just Support
Rather than using AI to find evidence for your existing belief, use it to synthesise two opposing viewpoints. Ask the AI to find the common ground—or the fundamental disagreements—between two different schools of thought.
Conclusion: Awareness is the Best Tool
The simple fact of knowing that AI sycophancy exists makes you a better user. AI is a powerful tool, but like any tool, it has its quirks. By staying aware of the "Echo Chamber Trap," we can ensure that we use AI to expand our minds, rather than just polish our egos.