Rejecting algorithmic advice even when it outperforms human judgment, especially after witnessing one error.
Imagine you have two friends helping you pick apples. One friend (the robot) picks 95 great apples and 5 bad ones. The other friend (the human) picks 80 great apples and 20 bad ones. But the moment you see the robot pick one bad apple, you say 'I knew it! Robots can't pick apples!' and go back to the friend who's actually worse at it. You forgive your human friend's mistakes because you think they'll learn, but you never forgive the robot for the same mistake.
Algorithmic aversion describes the paradoxical behavioral pattern in which people refuse to rely on statistical or algorithmic predictions after observing them make even minor errors, while simultaneously tolerating far larger and more frequent errors from human experts. This aversion is amplified for tasks perceived as subjective—such as hiring, medical decisions, or moral judgments—where people believe algorithms lack the empathy, intuition, or contextual understanding required. Critically, the bias is asymmetric: people hold algorithms to a standard of near-perfection while granting humans a generous margin of error, because they believe humans can learn and adapt from mistakes in ways algorithms cannot. The effect intensifies as the stakes of a decision increase, creating a tragic pattern in which people are most likely to reject superior algorithmic advice precisely when getting the decision right matters most.
The same glitch looks different depending on the terrain. Finance, medicine, a relationship, a team — same mechanism, different costume.
Investors and portfolio managers frequently override quantitative models after a single losing quarter, reverting to discretionary stock-picking that statistically underperforms index-tracking algorithms. The aversion is strongest after market downturns, when the emotional sting of algorithmic losses feels less tolerable than equivalent human-made losses.
Patients and physicians resist AI-assisted diagnostics—especially in radiology, dermatology, and pathology—despite evidence that these systems match or exceed specialist accuracy. The resistance intensifies for life-or-death decisions, where people feel that only a human can bear moral responsibility for an error.
Teachers and administrators resist algorithmic tools for student placement, grading, or early-warning dropout detection, believing that human intuition captures nuances that data cannot. This persists even when algorithmic predictions of student success are demonstrably more accurate than teacher judgment.
People strongly prefer human matchmakers or their own intuition over dating algorithms, viewing romantic compatibility as inherently subjective and resistant to quantification. Even users of dating apps may distrust the matching algorithm while trusting their own profile-swiping instincts, which are often driven by superficial cues.
Users abandon recommendation engines, spam filters, or autocomplete features after a single salient failure, reverting to manual processes that are slower and less accurate. Product teams struggle with the paradox that making algorithmic errors more visible (for transparency) can accelerate user abandonment.
HR departments resist algorithmic resume screening or performance evaluation tools after any publicized error, preferring interview-based assessments that introduce well-documented biases like the halo effect. Managers distrust algorithmic scheduling and forecasting tools the moment they produce a visibly wrong output.
Voters and the public distrust algorithmic content curation and moderation on social media, perceiving automated decisions about what news to show or which content to remove as lacking nuance and fairness—even when human moderators make similar or more frequent errors. This distrust fuels demands for 'human oversight' regardless of comparative accuracy.
Berkeley J. Dietvorst, Joseph P. Simmons, and Cade Massey at the Wharton School, University of Pennsylvania, coined and formalized the term in 2015, published in the Journal of Experimental Psychology: General.
Humans evolved to calibrate trust through social signals—reading intentions, emotions, and accountability in other agents. Survival depended on judging whether a fellow tribesperson was reliable based on eye contact, tone, and reciprocity. Delegating decisions to an entity that provides no social feedback loop (no explanation, no apology, no body language) violates deeply wired trust-assessment circuits that expect agency and intentionality in decision-makers.
Algorithmic aversion creates a feedback loop that degrades AI systems: when users override or ignore algorithmic recommendations, the resulting human-generated data pollutes training sets, teaching models to mimic suboptimal human decisions. Additionally, organizations underinvest in AI tools because adoption metrics are low, creating a self-fulfilling prophecy where algorithms never get the data or iteration cycles needed to improve. AI systems may also be deliberately 'dumbed down' or made less transparent to match human expectations, sacrificing accuracy for perceived trustworthiness.
50 cards are free to preview. Buyers unlock the rest of the deck plus the interactive training — Spot-the-Bias Quiz unlimited, Swipe Deck with spaced repetition, My Blindspots, Decision Pre-Flight, the Printable Deck + Cheat Sheets, and the Field Guide e-book. $29.50$59.
