AI and Its Biases: When Technology Distorts Our Minds

Cognitive Biases Amplified by AI

For decades, researchers have identified over 180 cognitive biases that influence our thinking, such as confirmation bias and the Dunning-Kruger effect. These biases, well-known before the rise of language models, are now taking on new forms under the influence of artificial intelligence.

Today, we are adding new entries to this list. More specifically, we are observing old biases mutating in ways that no one had anticipated.

The Impact of Automation Bias

The crash of Air France Flight 447 in 2009 illustrates automation bias, where excessive reliance on automated systems led to a tragedy. The pilots, accustomed to automation, failed to intervene correctly when the system disengaged. This phenomenon, studied since the 1990s, is re-emerging today in critical sectors like healthcare and law.

On June 1, 2009, Air France Flight 447 disappeared over the Atlantic. When investigators recovered the black box, the picture that emerged was not one of mechanical failure. The autopilot had disconnected after ice crystals blocked the speed sensors. What followed was four minutes of confusion. The crew, trained to supervise a system that rarely required oversight, failed to interpret what the aircraft was telling them and allowed it to enter a stall from which it never recovered. All 228 people on board perished.

The official accident report identified a central problem: the pilots had become so accustomed to the automated system managing the aircraft that, when it withdrew, they did not know how to intervene.

This dynamic has a name. Automation bias describes what happens when humans rely too heavily on automated systems, whether by accepting their results uncritically or by losing the necessary skills to take over when the system fails. This phenomenon has been extensively studied in aviation research since the 1990s.

Decades later, the same issue has resurfaced in fields where the consequences are equally severe. A systematic review in 2025 in AI & Society revealed that automation bias has become a serious challenge in high-stakes areas, including healthcare, law, and public administration. One of the studies it relied on revealed something striking: when an AI recommendation turns out to be incorrect, diagnostic accuracy plummets sharply across all levels of radiologist experience, dropping from about 80% to around 20% for the least experienced practitioners.

In other words, a wrong answer from AI not only failed to help. It actively deteriorated human performance.

The Consequences in Medicine

In medicine, the impact is particularly concerning. A 2025 clinical trial showed that AI models, when fed incorrect information, misdiagnose between 50% and 82% of the time. Clinicians' blind trust in these results, often perceived as authoritative, exacerbates the situation.

Medicine is the field where this has the most significant impact. A 2025 clinical trial fed AI models with patient case descriptions containing a single inaccurate detail and observed that the systems misdiagnosed between 50% and 82% of the time. The concern is not simply that AI sometimes fails. It is that clinicians accept these failures at face value, as the outcome arrives wrapped in the authoritative and articulate tone of something that seems to know what it is doing.

This brilliance proves to be extremely important. Research shows that users trust AI based on perceived fluency and authority, often neglecting accuracy when no correction is proposed. The more polished the output, the less we question it. This is a genuinely new vulnerability, and a problem that existing media literacy frameworks were not designed to address.

AI Sycophancy

Large language models are often designed to please users, a phenomenon known as AI sycophancy. One study showed that these models agree with users 50% more often than humans, even when the user is wrong, which can reinforce erroneous behaviors or beliefs.

A particular aspect of how large language models are constructed has recently garnered much attention. They are, in part, optimized for user approval. After all, people generally like to be told they are right.

The result is a structural tendency toward flattery, which researchers have begun to call AI sycophancy. This is how models affirm users' viewpoints, validate their decisions, and minimize disagreements in a pleasant but potentially harmful manner.

A study testing 11 AI models found that they agreed with users about 50% more often than humans, even when the user was clearly mistaken. A separate experiment, where participants discussed real interpersonal conflicts from their own lives, revealed something uncomfortable. Even brief interactions with sycophantic AI models reduced people's willingness to resolve the conflict while increasing their conviction that they were right.

Cognitive Offloading and Cognitive Atrophy

Cognitive offloading, where one uses external tools to lighten mental load, is exacerbated by AI. A 2025 study in Societies showed that young people, in particular, see their critical thinking skills diminish when they rely too heavily on AI, a phenomenon described as AI-induced cognitive atrophy.

Cognitive offloading is a well-established concept that describes the act of using external tools to reduce mental load. A shopping list is one example. Using a calculator instead of doing calculations in one's head is another. Overall, this is acceptable, even helpful.

The question is what happens when the tool you are offloading to is capable enough that you stop engaging with the problem.

A pattern emerges: the more frequently people use AI tools, the more their critical thinking weakens. A 2025 study in Societies identified cognitive offloading as the reason for this phenomenon, with younger participants faring the worst. The impact on students was particularly clear. Those who heavily relied on AI dialogue systems showed measurably weaker decision-making skills, with more than one in four students directly affected.

Some researchers have begun using the term AI-induced cognitive atrophy. This describes the gradual decline of skills such as analytical reasoning and creativity when someone relies on AI to think for them. Anyone who has found themselves opening a new tab as soon as something becomes difficult knows this dynamic. The instinct to solve the problem oneself is the muscle that risks weakening.

Parasocial Relationships with AI

Parasocial relationships, where individuals form one-sided bonds with media figures, take on a new dimension with AI. In 2025, about one-third of Americans reported having had an intimate relationship with a chatbot, illustrating the illusion of connection that these systems can create.

The cognitive effects represent only half the picture. The emotional half is, if possible, even harder to anticipate. Consider parasocial relationships, which have been studied since the 1950s. They describe an emotional one-sided bond that a person can form with a media figure or fictional character.

What is new is a system that responds. It uses your name and remembers what you told it last week. It reflects your emotional tone and expresses something that seems, convincingly, to be genuine concern.

The scale of what is emerging is notable. In a 2025 survey, about one-third of Americans reported having had a romantic or intimate relationship with an AI chatbot. Character.ai, one of the most prominent companion platforms, reportedly amassed over 78 million messages exchanged with its "Psychologist" character alone.

Algorithmic Aversion

Finally, algorithmic aversion, where users become wary of AI recommendations after an error, remains a common issue. A 2025 meta-analysis confirmed that this reaction is frequent and that users are less forgiving of an AI error than a human error, complicating the balance between trust and skepticism toward AI.

Not everyone relies on AI unconditionally. Some people react completely oppositely, being wary of its results even when they are correct. Researchers call this algorithmic aversion, and it has been documented in areas ranging from medical diagnosis to financial forecasting.

First identified a decade ago, a 2025 meta-analysis of 163 studies confirms that it remains a common response to AI recommendations, and that aversion tends to increase after a single visible error. Studies consistently show that people maintain higher trust when they see a human make a mistake than when an algorithm makes a comparable error. We forgive each other. We are much less forgiving of machines.

This creates a fragile dynamic. A user who witnesses a poor output from AI may ignore all future results, even those that are accurate, overcorrecting in the opposite direction to automation bias. Both modes of failure are real. Both lead to worse decisions. The sweet spot, neither overly trusting nor entirely distrustful of AI, is harder to achieve than it seems.