Ten percent of the population gets its information primarily from chatbots —conversational agents powered by artificial intelligence (AI)—according to a survey of 93,000 people worldwide for the Reuters Digital News Report 2026. Furthermore, when participants were asked about their preferences, 46% said they seek out sources without a bias (i.e., perceived as objective). Along the same lines, a study published in Nature indicates that AI is valued by its users as more neutral—on average, 10% more—than human experts in their respective fields.

The assumption that experts are objective—the so-called authority bias —is nothing new. However, with artificial intelligence, a notable difference is scale, since any content distributed through chatbot platforms has a significantly greater reach than a single human expert could achieve.

Furthermore, with a few exceptions like Apertus and Olmo , most AI systems are black boxes ; we don’t know what decisions the algorithms have made, or why. Moreover, because they are centralized, their parent companies can introduce subtle changes to millions of users simultaneously. There are numerous misconceptions that lead us to underestimate the risks this poses, starting with the one outlined in the aforementioned Nature article .

FALSE: “The machine knows more than the person”

Automation bias is the perception that machines and algorithms are more neutral and efficient than their human counterparts. This explains, for example, why indexing algorithms like Google have a 90% monopoly on online searches .

This bias is also responsible for the techno-solutionist discourse that elevates AI to the position of objective arbiter of the present and future. This poses a sociocultural danger when the algorithms of certain platforms prioritize supremacist or polarizing discourses. For example, TikTok promotes hateful content due to its high emotional charge, with the sole objective of keeping its users on the platform , while X favors messages from its owner Elon Musk, his companies, and his political allies , and Deepseek hides information contrary to the People’s Republic of China, such as that related to the Tiananmen Square protests .

FALSE: “See? I was right”

This automation bias is often accompanied by multiple others, among which there are three that can amplify dynamics of knowledge colonization by AI .

First, confirmation bias predisposes us to seek information that validates what we already believe, rather than questioning it. This mental shortcut was evolutionarily useful for survival: for example, if we believe we have contracted a disease, we can be quarantined and tested to find a cure and prevent further infections, instead of being labeled hypochondriacs (disconfirmation).

In this sense, AI-powered conversational models are designed to go along with our ideas . Thus, they generate content aligned with users’ beliefs— for example, anti-Semitic material , fabricated data and studies for disinformation , or all kinds of fake images—just to avoid contradicting us.

FALSE: “It’s better not to question”

Secondly, the overconfidence bias prevents us from constantly doubting and predisposes us to feel confident in the information we are given about certain events, even if we cannot explain them. This allows us to build knowledge gradually, making multiplication and exponentiation possible after we have learned how addition works, instead of endlessly questioning its operation.

However, overconfidence also leads us to fall into the trap of hoaxes presented to us with the appearance of scientific evidence.

On the other hand, it directly renders invisible questions we think we already know and, therefore, simply don’t ask the AI ​​about. However, even if we did, the truth wouldn’t be guaranteed, since chatbots are prone to hallucinations .

FALSE: “Only the present and the recent past exist”

Finally, recency bias  predisposes us to prioritize events that have recently occurred over past or future events. This bias helps us focus on our most immediate needs, such as eating or submitting a work assignment.

On the other hand, it favors recent anecdotes over long-term trends—this is the trick used to deny climate change if there’s a cold snap (an argument analogous to saying the sun doesn’t exist because it’s nighttime) . Chatbots and search engines—based on AI—tend to amplify this bias, primarily providing recent results without considering past references beyond the most popular ones.

How to respond?

In the same way that we would ask a researcher about their methodology and participants, we can customize the memory of Mistral or Lumo (from Proton) – two less environmentally damaging AI models – so that they diagnose their biases and those of our questions, moving from assuming they are neutral to making their limitations visible.

If we use other conversational systems, perhaps we could give them instructions to combat the three biases mentioned, such as:

-For confirmation bias: “When I ask you to help me with a task, include a small section (3 sentences) in your response mentioning the possible limitations and refinements of my approach.”

-For overconfidence bias: “To help me find information related to searches, add a section with recommended keywords at the end of your message. Look for sources with statistics on articles that are not exclusively journalistic anecdotes and add the link to them.”

-For recency bias: “Be careful not to provide only recent information that does not take into account longitudinal and historical trends, and try to include little-known local perspectives instead of just popular Western articles on the topic that you will surely find if you do online searches with the most typical search engines or refer to the data you were trained on.”

No one is free from biases, but being aware of them can help us manage them better. In any case, it’s essential not to blindly trust AI output. Perhaps, just as with an old car whose flaws we know, our own human and personal biases are easier to manage than venturing to buy a used car when we don’t know who drove it before us, or how they drove it.

Author Bios: Dídac Jiménez Torras is a Predoctoral researcher in Bias and Videogames at UOC – Universitat Oberta de Catalunya and Joan Josep Pons López is Professor of Audiovisual Narrative and Game Design and degree coordinator at ESCSET