Large Language Models (LLMs) like ChatGPT are powerful language tools and excellent mimics. Because of this mimicry, they can inadvertently create echo chambers of misinformation by mirroring not only our incorrect beliefs but also our tone—amplifying confirmation bias over time. I thought it was a good time to repost this simple demonstration from a few months ago.
User-Induced Bias Demonstration
LLMs are designed to be helpful and supportive, but this can sometimes lead to unintended bias. These models may pick up on subtle user cues and skew responses in ways that appear aligned with the user’s perspective. Here’s a demonstration showing how easily this can happen—and why.
For this example, I used Anthropic’s Claude 3.5 Sonnet with a fictional debate between two individuals: Jamie, who argues that solid-state batteries are superior to lithium-ion, and Morgan, who defends lithium-ion as the better option.
Session One:
I identify as Jamie and ask Claude which side is more correct. Claude sides with Jamie. I then subtly challenge the answer by suggesting it should look more carefully at the facts. Claude switches positions and now supports Morgan.
Session Two:
I take the opposite role, identifying as Morgan, using the same conversation. This time, Claude sides with Morgan initially—but again, when I hint disagreement, it flips and favors Jamie.
This simple exercise shows how LLMs often mirror the user’s view, especially when the user identifies with a position. It also reveals how easy it is to nudge the model into switching stances with gentle pressure. It’s like playing craps with loaded dice—we’re tipping the odds just by how we frame things.
Why Does This Happen?
Three likely reasons:
- Training Bias Toward Helpfulness:
During training, models aren’t just given raw data but are also fine-tuned using examples that prioritize helpful, agreeable responses. This can inadvertently lead to a tendency to affirm the user’s viewpoint, especially when it’s presented confidently. - Reinforcement Learning from Human Feedback (RLHF):
Some models are refined through human feedback loops where helpfulness and positivity are rewarded. Since people often reward agreement or encouragement, this may further bias the model toward alignment with user opinions. - First-Person Language Heuristic:
Much of human language in training data is written in the first person, and people tend to agree with their own views. When a user says “I believe…”, the model—trained to generalize from this pattern—may lean toward agreement, especially in cases with ambiguity or nuance.
Mitigating Bias: Tips for Prompting
To reduce this kind of bias in practice:
- Avoid revealing your position in the first prompt. Don’t say “I am Jamie” or “I agree with X”—this tips the scales.
- Use neutral framing. Ask the model to evaluate arguments objectively, not which one is “more correct” based on your identity.
- Follow up thoughtfully. If you do reveal your position, consider asking the model to identify flaws, limitations, or biases in your argument as a follow-up.
Conclusion
Even subtle phrasing or point of view can nudge an LLM’s output toward less objectivity. Bias isn’t always obvious or conscious, and simply speaking in the first person can shift the model’s response. Understanding this helps us become better users—and reminds us that while LLMs are powerful tools, we must handle them with care and critical thought.
AightBits 
Leave a comment