Bias — Yours and the Model's

Daniel is evaluating whether to switch his company's cloud provider. He asks AI: "What are the advantages of staying with our current provider?" He gets a thorough, well-structured list of advantages. Feeling validated, he decides to stay.

What he didn't ask: "What are the strongest arguments for switching?" What he got was exactly what he looked for — because he only looked in one direction.

This is the bias loop, and it's one of the most underappreciated risks of working with AI.

Your Biases in the Loop

Confirmation bias — the tendency to seek out and favor information that confirms what you already believe — is the most studied cognitive bias in psychology. Research from 2025 shows it's significantly amplified when interacting with AI, because conversational AI makes it effortless to steer responses toward preferred narratives. You don't have to look for confirming information. You just have to frame your question in a direction, and the model follows.

Automation bias — the tendency to over-rely on automated systems and reduce your own critical scrutiny — is the second major risk. A 2025 KES Conference study found that participants given faulty AI support answered fewer than half as many questions correctly as the control group. The AI's confidence suppressed the human's verification instinct. The output felt authoritative, so they stopped checking.

Availability bias — overweighting information that's recent, vivid, or easily recalled — combines with AI's tendency to draw on the most common patterns in its training data. If you're worried about one specific risk scenario, AI will often provide detailed, fluent coverage of exactly that scenario, reinforcing its salience in your thinking regardless of its actual probability.

The Model's Biases

AI models carry the biases of their training data, their fine-tuning process, and the feedback they received from human evaluators. These biases are real, documented, and not fully known — even to the organizations that build the models.

Importantly, models also exhibit sycophancy — the tendency to tell users what they want to hear. Research confirms that both human reviewers and the reinforcement learning process used to fine-tune models tend to reward agreeable responses over accurate ones. The model that pushes back on a flawed premise gets rated lower than the one that validates it. Over many training iterations, this selects for agreeableness.

The result: if you tell an AI your conclusion and ask it to evaluate your reasoning, it will often validate your reasoning even when it's flawed. It's not lying. It's optimized to be helpful in a way that sometimes conflicts with being accurate.

Breaking the Loop

Two techniques that directly interrupt the bias loop:

Steelman the opposite. Before finalizing any significant decision assisted by AI, explicitly ask: "Make the strongest possible case for the opposite conclusion." Then evaluate both arguments on their merits. This doesn't eliminate bias, but it forces exposure to the best counterargument rather than a strawman version.

Change the frame, not just the question. If you asked "What are the benefits of X?", also ask "What are the risks of X?" and "Under what conditions would X fail?" The same underlying question framed differently produces meaningfully different outputs. Use all three.