Algorithmic Aversion

Table of Contents

  1. Definition
  2. How to Measure Algorithmic Aversion
  3. How to Utilize the Research
  4. Comparison to Adjacent Concepts
  5. Best Practices
  6. Future Trends
  7. FAQs
  8. Related Terms
  9. Sources

Definition

Algorithmic aversion is the tendency for people to distrust, discount, or reject decisions made by algorithms — even when those algorithms outperform human alternatives. The term was coined by Berkeley Dietvorst, Joseph Simmons, and Cade Massey in a 2015 Journal of Experimental Psychology: General paper based on research conducted at the Wharton School. We show that people are especially averse to algorithmic forecasters after seeing them perform, even when they see them outperform a human forecaster. This is because people more quickly lose confidence in algorithmic than human forecasters after seeing them make the same mistake. Upenn

That last finding is the one that’s stuck. People aren’t generally hostile to algorithms in the abstract. They become hostile after they see an algorithm make a mistake — and they punish that mistake harder than they’d punish the same mistake from a human, even when the algorithm’s overall track record is better.

In marketing, algorithmic aversion shows up almost everywhere a brand has tried to replace or augment human judgment with automated systems. AI-driven product recommendations, chatbot customer service, dynamic pricing, programmatic creative, AI-generated content, automated underwriting, robo-advisors, AI medical screening tools — all of it runs into the same pattern. Consumers will often accept an algorithmic system in principle and then reject it the moment it returns a result that feels wrong, even if “feels wrong” doesn’t mean “is wrong.”

There are a few well-documented drivers. Consumers expect algorithms to be perfect and feel betrayed when they aren’t. They feel a loss of control when a machine decides on their behalf. They perceive algorithms as lacking the empathy and nuance that human judgment can provide — what researchers Chiara Longoni and colleagues call “uniqueness neglect,” the worry that an algorithm can’t account for what’s special about you specifically. The research into “algorithm aversion” highlights that consumers tend to favor human experts over AI in subjective or emotional decision contexts (Dietvorst et al., 2015) due to the algorithms’ inability to address consumers’ unique characteristics, referred to as “uniqueness neglect” (Longoni et al., 2019). In contrast, in more objective decision-making situations, such as those requiring numerical precision or logical reasoning, consumers may prefer AI over human recommendations (Castelo et al., 2019). California Management Review

It’s worth flagging the counter-effect. The same body of research has identified “algorithm appreciation” — situations where people prefer algorithmic advice, particularly when the human alternative is vague, when the task is numeric, or when the stakes are low. Aversion and appreciation coexist. Which one a consumer brings to a given interaction depends on the task, the framing, and what they’ve recently seen.

Back to Top

How to Measure Algorithmic Aversion

There’s no canonical formula. The standard experimental setup compares two groups: one given an algorithm-labeled recommendation, the other given a human-labeled recommendation, with the actual content held constant. The dependent variables typically include:

  • Reliance rate — how often the consumer follows the recommendation.
  • Confidence — self-reported trust in the recommendation.
  • Willingness to delegate — whether the consumer would let the system decide unsupervised.
  • Punishment after error — how sharply trust drops after a visible mistake.
  • Switching behavior — whether the consumer abandons the algorithmic option for a human one.

A simple applied version:

Aversion Index = 1 − (Reliance on algorithm ÷ Reliance on equivalently-described human)

A value of 0 means consumers treat the two sources equally. A positive value means aversion. A negative value means appreciation. The original Dietvorst experiments produced large positive values after participants saw the algorithm err, even when the algorithm was objectively more accurate than the human across the full sample.

Back to Top

How to Utilize the Research

A few practical applications for marketers and product teams deploying algorithmic systems.

Give users some control over the output. Dietvorst, Simmons, and Massey’s follow-up work in 2018 found a powerful intervention: letting users modify the algorithm’s output, even slightly, dramatically increased adoption. Participants were considerably more likely to choose to use an imperfect algorithm when they could modify its forecasts, and they performed better as a result. Notably, the preference for modifiable algorithms held even when participants were severely restricted in the modifications they could make (Studies 1-3). In fact, our results suggest that participants’ preference for modifiable algorithms was indicative of a desire for some control over the forecasting outcome, and not for a desire for greater control over the forecasting outcome. Restaurant booking algorithms that let you re-sort the suggestions, AI writing tools that let you edit the draft, recommendation systems with thumbs-up/down — they all work better than the same systems without those affordances. SSRN

Match the algorithm to the task type. Consumers tend to accept algorithmic recommendations for “search products” (where attributes can be evaluated objectively — laptops, hotels, cars by spec) and resist them for “experience products” (where personal taste dominates — wine, perfume, art). The marketing question isn’t “should we use AI here?” but “is the consumer treating this as a feature comparison or a taste judgment?”

Frame the algorithm as augmenting, not replacing, human judgment. Robo-advisors with optional human review, AI-assisted (not AI-only) medical screening, chatbots with “talk to a human” escalation. Hybrid framing consistently outperforms pure-automation framing in trust studies.

Manage the error narrative. Aversion accelerates after a single visible mistake. Brands deploying algorithmic systems benefit from setting realistic accuracy expectations upfront and providing context when errors occur. “Our system gets this right 94% of the time” lands very differently than the same system implicitly promising perfection and then missing.

Be cautious with disclosure of automation. Recent research has uncovered what’s been called the transparency dilemma: disclosing that something was AI-generated can erode trust even when the output is good. Marketers face a real tension between honesty (and regulatory pressure) about AI use and the trust costs of disclosure.

Audit human comparator framing. Aversion is stronger when the human alternative is described as expert. Conversely, algorithm appreciation tended to surface when the human alternative was vaguely described, such as “other participants” or “unspecified people.” These findings suggest that the framing of human expertise critically shapes the perceived value of algorithmic advice. The system you’re competing against, even rhetorically, affects whether users pick yours. Springer

Back to Top

Comparison to Adjacent Concepts

ConceptCore mechanismWhat it explainsHow it differs from algorithmic aversion
Algorithmic aversionDistrust of algorithms, amplified by visible errorsWhy people reject better-performing automated systemsCentered on the human side of the human-algorithm interaction
Algorithm appreciationPreference for algorithmic over human advice in certain tasksWhy people sometimes trust AI more than expertsThe mirror condition; same field of research, opposite finding
Uniqueness neglectBelief that algorithms can’t capture what’s special about youWhy algorithmic medical or psychological advice gets rejectedA specific cognitive driver of aversion
Automation biasOver-reliance on automated systemsWhy pilots and operators sometimes ignore correct manual cuesThe opposite failure mode — too much trust, not too little
AI literacy effectLower AI knowledge predicts higher AI receptivityWhy some novices are more comfortable with AI than expertsAbout receptivity, not specifically about post-error rejection
Persuasion Knowledge ModelConsumers learn to recognize and resist persuasionWhy advertising tactics decay in effectivenessAbout marketing tactics broadly; aversion is specific to algorithms
Loss of control aversionDistress when agency is removedWhy people resist mandatory automationOne driver of aversion, not the whole phenomenon
Back to Top

Best Practices

Give people the modify lever. The most robust finding in this entire literature is that even tiny amounts of user control over algorithmic output dramatically increase adoption. If your system doesn’t let users adjust anything, redesign before launch.

Don’t oversell accuracy. Implicit perfection promises (“our AI knows what you want”) set up exactly the failure pattern Dietvorst documented. The first visible miss creates disproportionate trust loss.

Show calibrated confidence. Systems that say “we’re not sure about this one” before an uncertain recommendation tend to retain trust better than systems that present every output with the same confidence.

Pair algorithmic outputs with human-authored context where stakes are high. In finance, healthcare, hiring, and education — categories with high involvement and high consequences — pure-algorithm interfaces underperform hybrid ones almost universally.

Watch the human comparator. If your algorithmic recommendation is being weighed against a named expert (“Sommelier Jane recommends…”), aversion goes up. Against an anonymous baseline (“Customers who liked X also liked Y”), it’s lower.

Design for recovery, not just performance. The question isn’t only “how often is this right?” but “what happens when it’s wrong?” A good error UX — clear acknowledgment, easy override, learning from the correction — protects trust better than rare-but-opaque errors.

Test labeling explicitly. The phrase “AI-powered,” “algorithm-driven,” “personalized for you,” and “recommended by our experts” can describe the same backend with very different consumer reactions. A/B test the labels.

Back to Top

A few shifts that are changing how this plays out.

Generative AI is rewriting the baseline. The 2015 research mostly studied algorithmic forecasts. Today, the more common consumer experience is algorithmic generation — writing, images, recommendations, conversation. Aversion patterns are evolving accordingly. Some research now finds that lower AI literacy predicts greater AI receptivity, complicating the older picture where exposure and education seemed to reduce aversion.

The transparency dilemma is sharpening. Regulators are pushing for AI disclosure (EU AI Act, FTC guidance, content provenance standards). Research is increasingly finding that the disclosure itself reduces trust, even when the underlying output is unchanged. Brands are caught between legal obligations and behavioral costs.

Task-specific calibration is becoming standard. Rather than asking “do consumers accept AI?” the field has moved toward “for what tasks, in what contexts, with what framing?” Marketers are starting to deploy AI selectively — recommendation engines yes, emotional service no — based on the task-type findings from Castelo, Longoni, and others.

Hybrid interfaces are winning. Across categories, the systems gaining traction tend to keep a human visibly in the loop. AI-drafted, human-approved. AI-suggested, human-decided. AI-screened, human-reviewed. The pure-automation interfaces that companies tried to launch in 2018-2022 have mostly retreated.

Repeated exposure changes the curve. A growing body of work finds that aversion can decline with repeated, successful exposure to algorithmic systems — but only when early experiences are good. Systems that fail early lose users in a way that’s hard to recover from.

Back to Top

FAQs

Is algorithmic aversion the same as fear of AI? Related but not identical. AI fear tends to be about big-picture concerns (jobs, surveillance, existential risk). Algorithmic aversion is a specific behavioral pattern in decision-making contexts — choosing the human option over the algorithmic one even when the algorithm is more accurate.

Do experts show less aversion? Mixed. Domain experts often show more aversion in their own field, not less, because they have strong priors about what good judgment looks like and notice when the algorithm deviates. AI literacy specifically — knowledge about how AI works — shows a more complex relationship with receptivity.

Does explainability reduce aversion? Sometimes. Showing why an algorithm reached a recommendation can build trust, but only if the explanation is meaningful to the user. Technical model details don’t help most consumers. Plain-language reasoning (“we suggested this because you bought X”) tends to.

Why does showing an algorithm err hurt trust so much more than seeing a human err? Several mechanisms. People expect machines to be precise, so errors violate expectations more sharply. Human errors are absorbed into a model of human fallibility; algorithmic errors are taken as evidence of a flawed system. There’s also less of a social contract — a human who made a mistake can apologize and learn; an algorithm doesn’t get that benefit.

Where does algorithm appreciation show up? In numeric estimation tasks, in objective domains like routing and pricing, in low-stakes recommendations, and when the human comparator is vague or unimpressive. It’s also stronger in younger consumers and in markets with more algorithmic exposure.

Is aversion stronger in some industries than others? Yes. Healthcare, hiring, criminal justice, finance, and any other category with high personal stakes and high subjective content tend to show stronger aversion. E-commerce recommendations, search ranking, and entertainment recommendations show much less.

How does this affect AI chatbot deployment in customer service? Significantly. Chatbots that handle transactional tasks (account balance, order status, password reset) tend to be accepted. Chatbots that handle emotional or relational tasks (complaints, complex billing disputes, sensitive issues) often increase frustration and reduce satisfaction. Escalation paths to humans are not optional in most categories.

Does branding the algorithm as “AI” make aversion worse? The literature is mixed and shifting. Some research finds that labeling something as AI reduces engagement; other research finds the label has limited effect when the output is good. The current consensus is that the label interacts with the task — labeling AI tends to hurt more for subjective tasks and matter less for objective ones.

Can companies overcome aversion by hiding the fact that an algorithm is involved? Short term, sometimes. Long term, it’s brittle. Consumers who discover they were interacting with an undisclosed algorithm react more harshly than they would have to upfront disclosure. Regulatory trends are also moving against undisclosed automation.

What’s the single biggest takeaway from the research? Give users some control. Across studies, across categories, the modifiable-algorithm finding is the most reliable intervention. People don’t need full control. They need any control.

Back to Top
  • Algorithm appreciation
  • Uniqueness neglect
  • Automation bias
  • AI literacy
  • Explainable AI (XAI)
  • Human-in-the-loop
  • Recommender systems
  • Trust calibration
  • Operational transparency
  • Persuasion Knowledge Model
Back to Top

Sources

Back to Top

Tags:

Was this helpful?

Previous

Absorptive Capacity

Next

Ansoff Matrix