Spotting bias in results

No source is neutral—not even this one

Bias runs through media, studies, search algorithms, and AI models—including ChatGPT. Knowing about bias isn’t about becoming reflexively skeptical; it’s about knowing which correction to apply based on the source type. A clear-eyed user gets better results than a credulous one, often by asking just one more question. This lesson covers the three families of bias you’ll encounter: source bias, tool bias, and your own.

What distorts web sources

Editorial bias is the most visible: every outlet selects and frames facts according to its stance, and the same event can be covered in vastly different ways. You expose it by asking for a comparison of how several outlets with different leanings cover a topic. That makes the gap legible instead of dissolving it into an averaged summary.

Selection bias is sneakier. Search engines, and ChatGPT after them, favor popular, well-indexed sources; high-quality but less visible work stays off your radar. The fix is to explicitly request other source types: “Find academic publications or NGO reports on this topic, not just news articles.” On technical or social issues, the difference in results is striking.

Commercial bias affects anything related to purchasing. A large share of web content is sponsored or tied to affiliate programs, and a product comparison can reflect partnerships rather than an independent evaluation. Before making any equipment decision, make this a reflex: “Do the cited sources have commercial ties to the recommended products?”

Recency bias, finally, privileges what was just published even when older analyses are stronger. In that case, specify that you want foundational work, not just the latest headlines: on structural topics, yesterday’s article is rarely the best.

What distorts ChatGPT

The tool adds its own quirks, which OpenAI’s help center acknowledges outright: presenting a single perspective as absolute truth, oversimplifying a nuanced issue, or misrepresenting the weight of a scientific consensus. Its synthesis bias pushes it to present a consensus where the topic is debated, because a coherent answer is easier to produce than a contradictory one. Its availability bias overrepresents English-language sources compared to others. Its people-pleasing bias inclines it to give the answer you seem to expect, which makes leading questions especially dangerous. And its update lag means information from the last few hours may not be indexed yet.

You can largely correct for these quirks with four targeted prompts. “Are there experts or sources who disagree with this analysis?” reopens the debate that synthesis closed. “Also find sources in French / German / Spanish on this topic” restores linguistic balance—useful whenever the topic is European or national. “What are the strongest arguments AGAINST this conclusion?” makes you play devil’s advocate and neutralizes people-pleasing. And “What are the limits of this analysis? What information is missing?” gives you an inventory of blind spots, often more instructive than the answer itself.

What distorts you

Your own biases weigh at least as much. Confirmation bias leads you to seek what reinforces what you already believe; it’s the most common and the most costly. The halo effect makes you treat a source as reliable on all topics when it’s only reliable on its own. Anchoring bias gives the first piece of information you find disproportionate weight over everything you read afterward. Authority bias, finally, makes you place excessive trust in anything from a prestigious institution, regardless of the document’s rigor.

The countermove lies in how you phrase your request. Instead of asking for proof of your hypothesis, ask for evidence for and against the topic. The difference may look cosmetic; it changes the nature of what comes back to you.

Your turn

An exercise that stirs things up. Pick a topic on which you hold a firm opinion. First ask for the strongest arguments in favor of that opinion, then the strongest arguments against it, and finish with: “What biases could affect the sources you cited in your two answers?” Honestly observe whether the opposing arguments surprise you or lead you to nuance your position. This is the only part of this course you can’t delegate: your own biases are the ones you’re least equipped to spot.

Official sources: Does ChatGPT tell the truth? · Searching the web with ChatGPT (OpenAI help center)

Key takeaways

• Bias exists in sources, in ChatGPT, and in you
• Always ask for opposing viewpoints and the limits of an analysis
• Vary your sources (languages, media types, perspectives)
• Phrase your searches neutrally to avoid confirmation bias
• Awareness of your own biases is the first step toward high-quality research