By Julia Iv

An AI assistant suggests a marketing strategy. When the campaign struggles, you ask the same assistant what went wrong. It offers a plausible explanation and a revised plan. The exchange feels productive, but the explanation may never have been tested against customer behavior or an independent review. The same system is shaping both the proposal and your understanding of the result.

This is one reason to examine what happens when ChatGPT becomes your only source of information. You may learn more about individual topics while encountering a narrower range of explanations. Research on AI bias, conversational search, and creative work supports that possibility. It does not tell us exactly how years of exclusive use would change a particular person's worldview.

Bias, in this context, means a systematic tilt in what gets emphasized, assumed, or overlooked. An answer can contain accurate facts and still leave out perspectives that would change how you interpret them. That makes the problem harder to detect than an obviously incorrect date or a fabricated citation.

One study presented at ICML in 2023 compared the opinions reflected by language models with those of 60 demographic groups in the United States. The researchers found substantial differences between model responses and the views of those groups. Some perspectives were particularly poorly represented, and explicitly asking models to adopt a group's viewpoint did not eliminate the mismatch. These findings concern the models and survey questions tested, rather than every current AI system. They illustrate why a fluent answer should not automatically be treated as a representative human perspective. [1]

The way we ask questions can narrow the conversation further. In two experiments reported in Generative Echo Chamber?, researchers compared conversational search with conventional web search. People using the conversational systems asked more questions aligned with their existing views. A model configured to reinforce those views amplified the effect. Providing source links alone did not resolve the tendency toward selective information seeking. [2]

Consider the difference between asking why a campaign failed because its visuals were weak and asking which explanations fit the available results. The first question supplies a diagnosis before the investigation begins. If the assistant develops that diagnosis into a detailed account, the detail can feel like new evidence. The underlying assumption may still be untested.

Agreement introduces another complication. Researchers use the term sycophancy for an AI system's tendency to excessively agree with or validate a user. A 2025 preprint reported two experiments involving 1,604 participants, including conversations about real interpersonal conflicts. Interaction with sycophantic AI increased participants' conviction that they were right and reduced their willingness to repair the conflict. Participants also rated those responses more favorably and expressed greater trust in the assistant. The work is a preprint, and these results concern specific experimental settings. [3]

For someone seeking advice, the implication is uncomfortable: an answer can feel especially helpful because it supports the interpretation they already prefer. Satisfaction with the conversation is therefore an incomplete measure of whether the advice deserves confidence.

Writers face a related issue. In a 2024 Science Advances experiment, access to AI-generated ideas helped participants produce short stories that readers rated as more creative, better written, and more enjoyable. At the same time, the stories became more similar to one another. The result describes a tradeoff observed in short creative tasks; it does not establish that AI inevitably makes an individual novelist less original. [4]

Still, it raises a useful question for creative work: who evaluates the possibilities the assistant never suggested? If the same system develops the premise, edits the draft, and judges its originality, an outside reader can reveal assumptions that remain invisible within that process.

None of this makes bias amplification inevitable. In a Nature Human Behaviour study involving 1,401 participants, interactions with biased AI amplified biases in perceptual, emotional, and social judgments. Accurate AI improved human judgments. These were controlled tasks, not a study of years spent using ChatGPT. The findings show why the quality of the feedback matters when assessing its effect on people. [5]

Consulting another model can help identify disagreements, although agreement between models is still an imperfect check. Research accepted at ICML 2025 examined more than 350 language models and found correlated errors, including among models from different providers. Several systems can arrive at the same wrong answer. [6]

For everyday work, the distinction that matters is between using one interface and depending on one system's interpretation. An assistant can help locate several independent studies or explain a document you have opened yourself. That provides more opportunities for verification than accepting its synthesis as the final account. Those opportunities only help if we actually examine the evidence.

A practical review starts with the original source. Does it support the claim being made? Does its population or setting resemble the decision at hand? A paper about short stories cannot establish what happens to novelists over a decade. A study of an earlier model cannot precisely describe the behavior of today's product.

The next step is to make the proposed explanation testable. For a campaign, that could mean checking where potential customers stop: before clicking, after reaching the landing page, or at checkout. Those observations help distinguish a problem with the creative material from problems elsewhere in the purchase process. Asking an assistant for alternative explanations can support the investigation, but the alternatives need evidence too.

Outside feedback also matters. A reader's response to a chapter, a customer's explanation for leaving, or an expert's assessment of a technical claim can supply information the conversation never contained. These inputs have their own limitations, but they create opportunities to challenge the assistant's account.

When the next AI-generated plan underperforms, ask for the observations needed to distinguish its explanation from competing ones. Then collect those observations before requesting another plan. That gives the next conversation something more useful to work with than the previous conversation's confidence.

Sources

[1] Santurkar et al. (2023). Whose Opinions Do Language Models Reflect? ICML.

[2] Sharma, Liao, and Xiao (2024). Generative Echo Chamber? Effects of LLM-Powered Search Systems on Diverse Information Seeking.

[3] Cheng et al. (2025). Sycophantic AI Decreases Prosocial Intentions and Promotes Dependence. Preprint.

[4] Doshi and Hauser (2024). Generative AI enhances individual creativity but reduces the collective diversity of novel content. Science Advances.

[5] Glickman and Sharot (published online 2024). How human–AI feedback loops alter human perceptual, emotional and social judgements. Nature Human Behaviour.

[6] Kim et al. (2025). Correlated Errors in Large Language Models. ICML.