Imagine that you’re the Director of Marketing at a creative agency. You’re in search of a new graphic designer and have received almost 200 applications. The creative agency has recently adopted a new AI-powered hiring tool that screens resumes for relevant skills, experience, and cultural fit. To save yourself time, you use the tool to review the resumes and identify five top candidates. Just as you’re about to send out interview invites, your coworker asks if you’ll be interviewing their acquaintance, Lily. Lily wasn’t one of the candidates the AI tool suggested, but based on your coworker’s referral, you go back to find her resume. 

Lily’s resume looks great! She has supported major brands with graphic design and has strong academic qualifications. She’s someone you think you would have identified as a top candidate, but the AI tool rejected her. Now, after losing trust in the screening tool, you decide to manually review all 200+ resumes. Your loss of trust even leads you to dismiss the five candidates flagged by the tool. When you interview your top five candidates, including Lily, you’re left more confused. The people whose resumes you liked the best performed badly in the interview, unable to answer simple technical questions and lacking passion, except for Lily, who was a superstar. This led you to conclude that the algorithm had made a mistake, so the next time you need to hire someone, you bypass the AI tool altogether.

In this instance, you’ve fallen victim to algorithm aversion. Often, algorithms outperform people and make more optimal decisions than humans. Their decisions are less likely to be influenced by bias, but people generally distrust technology and believe they are better equipped to make decisions. People will forgive human errors more quickly—like the fact that you ignored that your judgment was incorrect for four of the five candidates you interviewed—but may completely lose faith in algorithms if they make just one mistake, in this case, not identifying Lily as a good candidate. Part of the reason for algorithm aversion is that we often don’t understand how algorithms arrive at certain decisions, and the lack of transparency leads to distrust.¹

Artificial intelligence (AI): A machine or computer system trained through the analysis of vast amounts of data, to mimic human intelligence. Artificial intelligence allows computers to problem-solve and make decisions that historically could only be made by humans.³

Algorithm: A set of rules or instructions that are followed in order to solve a problem. Artificial intelligence develops algorithms by finding patterns in data, and subsequently uses them to make decisions when analyzing new data.⁴ 

Transparency/Explainability: Transparency refers to the degree to which information about how AI systems operate and make decisions is available to users. Explainability refers to our understanding of the algorithms or rules an AI tool uses to arrive at a decision. Often a lack of transparency and explainability in AI tools causes algorithm aversion.⁵ 

Decision Fatigue: The tendency for the quality of our decisions to decline when we have to make many decisions or are presented with too many options. As our cognitive capabilities are limited, after making too many decisions, we may feel fatigued or overwhelmed and make poorer decisions. Decision fatigue is a human phenomenon and supports the case for using AI tools to make decisions for us. 

Automation Bias: The opposite of algorithm aversion, in which we favor the decisions made by algorithms and AI and ignore contradictory information. It is an overreliance on automated systems that can lead to poor decisions.⁶

In 1954, clinical psychologist Paul Meehl published his book Clinical vs. Statistical Prediction: A Theoretical Analysis and Review of the Evidence, in which he compared the efficacy of clinical versus statistical prediction methods used by psychological clinicians. Historically, in clinical psychology, practitioners often relied on subjective judgment to diagnose patients and predict treatment success, but there was growing demand for psychologists to become more empirical and objective at this time. In the book, Meehl suggested that simple, statistical methods often outperformed clinicians across a wide range of prediction tasks. Although Meehl didn’t use the term “algorithm aversion,” the book highlighted that clinicians often preferred using their judgment instead of statistical prediction tools, even though the latter were often superior.⁷

Meehl’s book was highly controversial. People resisted the idea that statistical models and formulas could make better decisions than humans, especially in such a human-centric field. Similarly, in the medical field, doctors were resistant to having machines diagnose patients, even when there was evidence suggesting automated systems could make better decisions than humans.⁸ However, as computers became more prominent and technological advancements made them more sophisticated, the resistance towards algorithms lessened in the 1960s and 1970s, with algorithms mostly being used by managers in business. Still, computer-based solutions were not fully trusted. Instead, they were conceptualized as decision support systems that could help people make decisions by analyzing data and modeling. Ultimately, the final decision still came from a human.⁹ 

In 2000, as machines had become deeply embedded in most fields, psychologist William M. Grove and his colleagues decided to revisit people’s aversion to using algorithmic decision tools. They compared the accuracy of clinical and mechanical data-combination techniques for making predictions about human health and behavior. Through a review of 136 studies, they found that algorithms outperformed human forecasters by an average of 10%.¹⁰ Although there was evidence of the superiority of statistical prediction models, people still showed a lack of trust in these decision-making techniques. Researchers explored various reasons for the distrust: the desire for perfect forecasts, the perceived inability of algorithms to learn, the presumed ability of human forecasters to improve with experience, the perception of algorithms as dehumanizing, the belief that algorithms could not incorporate qualitative data into their decision-making, and ethical concerns about the reliance on algorithms.¹¹

In 2015, professors Berkeley Dietvorst, Joseph Simmons, and Cade Massey finally put a name to this phenomenon: algorithm aversion. They conducted a series of experiments demonstrating that when someone sees an algorithm make an error, they become less likely to use it and favor a human forecaster. For example, in one study, participants were given admissions data from past students and asked how well they had performed in the MBA program. They had to decide between making the decision themselves or using a statistical model that had been built on that data. Those who had observed the algorithm make an error were less likely to rely on it than those who had not seen it make a mistake, despite the fact that the model outperformed participants in all studies.¹¹ 

As algorithms and AI become more embedded in our day-to-day lives and work tasks, it’s more important than ever to confront algorithm aversion. Algorithm aversion can hinder progress and cause people to use suboptimal decision-making processes, like relying on human judgment, if they encounter even a single mistake by the statistical tool.

Paul Meehl 

An American clinical psychologist, well known for his work on the limitations of human judgment in predicting behavior. In his influential 1954 book Clinical vs Statistical Prediction: A Theoretical Analysis and Review of the Evidence, he advocated for the use of statistics and computational modeling in clinical psychology, showing that these methods outperformed human judgment.¹² Meehl also served as president of the American Psychological Association (APA) in 1962.¹³

William M. Grove

An American clinical psychologist whose research focused on statistics, psychiatry, and internal medicine. Grove was an advocate for using applied statistics in psychiatry research, for which he provided evidence through a meta-analysis of over 100 studies with his colleagues in 2000, demonstrating mechanical data-combination techniques were better at predicting outcomes than human judgment. He was also interested in schizotypy traits and demonstrated how they could be indicators of family susceptibility to schizophrenia.¹⁴

 Berkeley Dietvorst

An American professor of marketing whose doctorate, focused on decision-making processes, led to his interest in studying algorithm aversion. Dietvorst, along with his colleagues, coined the term in their 2015 paper, “Algorithm Aversion: People Erroneously Avoid Algorithms After Seeing Them Err.” He believes that within business, marketers most often use algorithms to help them decide how to price products, which ads to run, and which markets to target. His follow-up research explored why algorithm aversion exists and how it can be overcome.¹⁵ 

Joseph Simmons

An American professor of applied statistics and operations, information, and decisions, who, alongside other academic colleagues, coined the term “algorithm aversion.” Simmons’ research explores how psychological processes and biases lead to suboptimal decision making. Simmons is also a strong advocate for transparency in research to preserve the integrity of the field, and his work has changed the way behavioral scientists report their research.¹⁶

Cade Massey

An American professor of operations, information, and decisions, who alongside Berkeley Dietvorst and Joseph Simmons, coined the term “algorithm aversion” in 2015. Massey’s research focuses on how people make decisions in the face of uncertainty, examining real-world contexts like financial investment, football league drafts, and graduate school admissions. Currently, Massey is the faculty co-director of the Wharton People Lab, a University of Pennsylvania research and education hub that focuses on data-driven workplace decision-making.¹⁷

Algorithm aversion significantly influences how AI and automated tools are integrated—or resisted—across various sectors. Understanding these impacts reveals both missed opportunities and potential risks in healthcare, hiring, and financial decision-making.

Medical algorithm aversion

The COVID-19 pandemic showed us that, globally, we are experiencing a shortage of doctors. Doctors and nurses are overworked, which can often lead to decision fatigue, where suboptimal decisions are made for patients. While AI and algorithm tools can help decrease the burden, helping practitioners make better, and quicker, decisions, and treat a greater number of patients, algorithm aversion has prevented these tools from being meaningfully embedded in healthcare.¹⁸ This means patients receive a quality of care lower than what could be achieved through statistical modeling tools to diagnose and suggest treatment outcomes.

It’s not just doctors that mistrust AI. A 2025 study showed that patients do not trust health systems to use artificial intelligence, with 65.8% reporting that they do not trust them to use AI responsibly, and 57.7% saying that they were not confident that the system would ensure an AI tool did not harm them.¹⁹ People perceive these tools as impersonal and fear that they lack the nuance to adequately treat them. 

Algorithm aversion in hiring practices

Over the years, human bias in hiring decisions has been under scrutiny. While people have advocated for strategies like blind resume reviews (removing names to avoid racial or gender discrimination), people still continue to challenge the use of AI tools for candidate evaluation. This resistance isn’t without merit—famous cases like Amazon’s AI recruiting tool, which showed bias towards women,²⁰ show that algorithmic practices aren’t without fault either. Part of the problem is that these algorithms are often trained on human data, which tends to include bias.

However, some research has shown that AI tools can reduce bias in hiring practices. Frida Polli, co-founder of Pymetrics, a game-based recruiting tool that assesses social, cognitive, and behavioral skills, has discussed how AI can focus on objective criteria, reducing subjective judgments that humans make due to biases and focusing on what really matters. Instead of making judgments based on a person’s resume, where information such as name, address, and education appears, AI can assess communication and problem-solving skills from video interviews and approach evaluation from a cultural-fit perspective. Companies like Unilever have resisted algorithm aversion and partnered with Pymetrics, which led to the hiring of 30,000 people in just one year.²¹

It seems that both algorithm aversion and automation bias are dangerous for hiring practices. Algorithm aversion causes people to rely on their own judgment to make hiring decisions, which is both time consuming and prone to bias, while automation bias can cause people to blindly accept algorithm-driven hiring decisions without scrutiny. 

Bias in financial decisions

In many financial institutions, the decision of whether or not to approve someone’s loan often relies on simple statistical models. For example, models like Fair Isaac Corporation (FICO), which look at someone’s credit history, debt, and income, are used in the U.S. for credit decisions. Financial institutions rely on these simple models because they are transparent, and lenders can understand why the model arrived at a particular decision. They avoid using more complex, nuanced AI tools because of algorithm aversion.

However, this system is biased against people who lack standard credit backgrounds. Young people, newcomers, individuals in the gig economy, and low-income individuals who haven’t yet built a credit file, will often be denied credit through traditional scorecards. Research has shown that more sophisticated AI-driven models are able to analyze alternative data sources, such as e-commerce activity and utility and rent payment history,  to gain a more complete understanding of a person’s financial behavior and make credit decisions. One fintech company found that using AI to assess credit decisions led to a boost in approval rates of up to 30%, without increasing risk for the financial institution. ²²

By resisting these more accurate tools due to algorithm aversion, financial institutions may unintentionally perpetuate exclusion and missing opportunities to extend fair and responsible credit to underserved populations.

The rise of algorithms in decision-making has sparked important debates around trust and reliance. Balancing human judgment with automated recommendations remains a complex challenge with significant consequences.

Automation Bias

While being overly distrustful of algorithms and AI can negatively impact decision-making, being overly confident can lead to the same outcome. Automation bias describes an overreliance on automated systems, causing us to not apply our critical thinking skills and human intuition to problem-solving. While algorithms can help facilitate decision-making, it’s important to ensure transparency and the ability to understand how the system arrived at a decision so as not to be complacent. 

In some situations, automation bias can have disastrous outcomes. In the aviation industry, for example, pilots rely on AI tools and computers that fly the aircraft, calculate fuel-efficient paths, and diagnose system malfunctions. While these can support pilots, they may also lead to overreliance. Studies conducted on pilots in simulators found that over 50% of the time, pilots would ignore crucial information, such as anomalies in weather patterns, altitude, or communication from air traffic control, if their automated systems did not alert them. The pilots placed too much trust in the automated systems, believing that if something was really wrong, it would alert them, rather than trusting their own expertise or experience.⁶

Can algorithms really minimize bias?

Although there are instances where algorithms can make better predictions than people, because they are often trained on human data, they can perpetuate the same biases. In some cases, algorithm aversion is prudent. 

For example, AI tools are sometimes used in criminal justice and policing, but because of bias in criminal profiling, they can incorrectly predict the likelihood of reoffense based on demographic characteristics such as race. In one case, a computer program used to predict the likelihood of recidivism for two individuals who committed similar crimes made an error. In 2014, Brisha Borden was arrested for stealing a kid’s bike. A year before, Vernon Prater was arrested for shoplifting tools from Home Depot. The algorithm predicted that Borden was at higher risk than Prater for committing a future crime. Here’s where it gets interesting: Borden was 18 years old and had never committed a crime before, while Prater had already been to prison for armed robbery. The reason the algorithm rated Borden higher was because she was black, while Prater was white. Just two years later, Borden had not committed any other crimes, while Prater was serving an eight-year prison term for stealing again.²³

While in other fields, just one mistake may not warrant complete distrust of algorithms for decision-making, when it comes to high-risk assessments like those in the criminal justice system, just one error can have grave consequences.

Automation Complacency

While AI and algorithms can be useful tools for making decisions more efficiently and optimally, having some reservations can provide a critical safeguard. When people rely solely on these tools to make decisions, they push aside critical thinking and human judgment, which are necessary to scrutinize and evaluate information. AI can make errors, just as humans can. Over time, if we become complacent, we may stop thinking critically, making it more difficult to identify errors.

For example, while studies have shown that using algorithms in healthcare through clinical decision support systems leads to more accurate diagnoses and treatment, over-reliance on them can also lead to critical errors. In a 1995 study, doctors were asked to interpret patient cases that included electrocardiogram (ECG) tests. Half of the doctors saw ECGs with computer-generated interpretations attached, while the other half only saw the ECGs. Doctors were asked to give their own diagnoses. The researchers found that doctors who saw the computer-generated interpretations were more likely to agree with them, and spent approximately 25% less time reviewing them. This was true even when the diagnosis was wrong.²⁴ This study highlights a critical form of automation bias: the loss of accurate expert judgment when trust is placed in the system over personal expertise.

Mistrust in self-parking systems

These days, most cars come with some kind of assistive parking technology, but, we often ignore the parking camera, turning our heads over the seat instead to park. We believe we can see better than the camera system. 

In 2015, automakers were increasingly integrating self-parking features into vehicles and testing their ability. With parallel parking being one of the most difficult tasks to execute, the American Automobile Association (AAA) conducted studies to see how self-parking systems in five different car models performed compared to the same models without these systems. The study showed that drivers who used the self-parking systems had 81% fewer curb strikes and parked the car 10% faster than those without, using 47% fewer maneuvers. Sometimes, the system was able to successfully park the car with just one maneuver—a feat I’ve certainly never accomplished while parallel parking. 

What AAA also found, though, was that the majority of Americans do not trust self-parking technology. A total of 80% of American drivers are confident in their independent parallel parking abilities, whereas only 25% would trust technology to park their vehicle behind them. Due to biases like algorithm aversion, we become overconfident in our own abilities and undervalue technology that can improve our driving.²⁵

Robo-advisors could help you get rich—if you let them!

AI and algorithms can be powerful tools for managing finances, with more and more banks investing in tools to support their customers. Studies have shown that tools like robo-advisors can optimize investments based on desired return and risk tolerance, and periodically adjust your portfolio for you as markets fluctuate, but people still prefer to use human advisors. 

People’s preference for having humans handle their money has been consistent throughout generations. Even though millennials have grown up in a digital world and tend to be more comfortable with technology, a study conducted by Vanguard, an investment company, found that millennials were partial to human advice just like Gen X and baby boomers. Vanguard even qualified the preference by surveying 1,500 customers. Clients believed that a human advisor added $160,000 in value to a $1 million portfolio goal, whereas a robo-advisor was thought to only add $50,000.²⁶ However, this is only a perceived value. Vanguard found that clients who used robo-advisors had an average annual return of 24%, whereas those who used a human advisor had an average annual return of 15%. In this case, algorithm aversion and distrust in AI could actually be costing people money.²⁷

Algorithms that Run the World with Cathy O’Neil 

Although algorithms are thought to be objective, bypassing human judgment, this isn’t always the case. Often, algorithms are developed or trained on historical human data, which means they can continue to perpetuate bias. In this podcast, our Research Director, Dr. Brooke Struck, sits down with Cathy O’Neil, an expert in arithmetic algebraic geometry and author of the New York Times bestseller Weapons of Math, to discuss the political nature of algorithms and how we can create more responsible algorithms. 

Algorithms for Simpler Decision-Making (½): The Case for Cognitive Prosthetics

In the age of Big Data, where vast amounts of data are available, algorithms can assist our decision-making processes by filtering information and suggesting what is most relevant to our problem. Most of us are quick to turn to Google or ChatGPT to help us solve a question. In some ways, these tools act like a cognitive prosthetic, extending our capacity for critical thinking. However, they can also limit cognitive autonomy and influence what information we consume. In this article, our writer, Jason Burton, explores the trade-offs of using algorithms, recognizing that they are both empowering and constraining. 

1. Slattery, P. (2024, May 1). Algorithm aversion: Who trusts algorithms to make decisions? Psychology Today. https://www.psychologytoday.com/ca/blog/transformative-ai/202404/algorithm-aversion-who-trusts-algorithms-to-make-decisions
2. Harrell, E. (2016, September 7). Managers shouldn’t fear algorithm-based decision making. Harvard Business Review. https://hbr.org/2016/09/managers-shouldnt-fear-algorithm-based-decision-making
3. Innovation, Science and Economic Development Canada. (n.d.). Artificial Intelligence (AI). In Government of Canada. Retrieved March 7, 2025, from https://www.canada.ca/en/services/science/innovation/artificial-intelligence.html
4. Gillis, A. S. (2024, July 29). What is an algorithm? TechTarget. https://www.techtarget.com/whatis/definition/algorithm
5. Macvittie, L. (2024, July 16). Crucial concepts in AI: Transparency and explainability. F5 Networks Blog. https://www.f5.com/company/blog/crucial-concepts-in-ai-transparency-and-explainability
6. Hoffman, B. (2024, March 10). Automation bias: What it is and how to overcome it. Forbes. https://www.forbes.com/sites/brycehoffman/2024/03/10/automation-bias-what-it-is-and-how-to-overcome-it/
7. Grove, W. M. (2005). Clinical versus statistical prediction: The contribution of Paul E. Meehl. Journal of Clinical Psychology, 61(10), 1233–1243. https://doi.org/10.1002/jclp.20179
8. Davenport, T. H., & Harris, J. G. (2005, July 15). Automated decision-making comes of age. MIT Sloan Management Review, 46(4), 83–89. https://sloanreview.mit.edu/article/automated-decision-making-comes-of-age/ 
9. The History of Automated Decision-Making. (2023, December 10). Symbio6. Retrieved March 7, 2025, from https://symbio6.nl/en/blog/history-of-automated-decision-making
10. Grove, W. M., Zald, D. H., Lebow, B. S., Snitz, B. E., & Nelson, C. (2000). Clinical versus mechanical prediction: A meta‑analysis. Psychological Assessment, 12(1), 19–30. https://doi.org/10.1037/1040-3590.12.1.19 
11. Dietvorst, B. J., Simmons, J. P., & Massey, C. (2015). Algorithm aversion: People erroneously avoid algorithms after seeing them err. Journal of Experimental Psychology: General, 144(1), 114–126. https://doi.org/10.1037/xge0000033
12. Ryu, J. (2023). Remembering Paul E. Meehl: Historical contributions to predictive modeling in human behavior. Harvard Review of Psychiatry, 31(2), 92–95. https://doi.org/10.1097/HRP.0000000000000360
13. American Psychological Association. (n.d.). Paul E. Meehl: 1962 APA President. American Psychological Association. Retrieved March 7, 2025, from https://www.apa.org/about/governance/president/bio-paul-meehl
14. Research.com. (2025). William M. Grove. In World’s best scientists 2025 revealed: Best scientists – Psychology. https://research.com/u/william-m-grove
15. Nief, T. (Host). (2018, November 9). Berkeley Dietvorst, PhD (University of Chicago) [Audio podcast episode]. In Todd Nief Podcast. Todd Nief. https://toddnief.com/berkeley-dietvorst-interview/
16. Simmons, J. (n.d.). Joseph Simmons, PhD. Wharton Executive Education. Retrieved August 4, 2025, from https://executiveeducation.wharton.upenn.edu/faculty/joseph-simmons/
17. Massey, C. (n.d.). Cade Massey. The Wharton School, University of Pennsylvania. Retrieved August 4, 2025, from https://oid.wharton.upenn.edu/profile/cadem/
18. Zhao, Y., & Xiao, C. (2024). Overcome medical algorithm aversion: Conditional joint effect of direct and indirect information. International Journal of Human–Computer Interaction, 41(7), 3889–3900. https://doi.org/10.1080/10447318.2024.2344146
19. Nong, P., & Platt, J. (2025). Patients’ trust in health systems to use artificial intelligence. JAMA Network Open, 8(2), e2460628. https://doi.org/10.1001/jamanetworkopen.2024.60628
20. Goodman, R. (2018, October 12). Why Amazon’s automated hiring tool discriminated against women. American Civil Liberties Union. https://www.aclu.org/news/womens-rights/why-amazons-automated-hiring-tool-discriminated-against
21. Hale, M. (2023, October 12). Next gen AI in action: Unilever’s AI-powered recruitment revolution. Global Skill Development Council. https://www.gsdcouncil.org/blogs/next-gen-ai-in-action-unilever-s-ai-powered-recruitment-revolution#
22. Duran, N. (2025, February 10). How AI can improve credit scoring for greater financial inclusion. ISU Corp. https://www.isucorp.ca/blog/how-ai-can-improve-credit-scoring-for-greater-financial-inclusion
23. Angwin, J., Larson, J., Mattu, S., & Kirchner, L. (2016, May 23). Machine bias: There’s software used across the country to predict future criminals. And it’s biased against blacks. ProPublica. https://www.propublica.org/article/machine-bias-risk-assessments-in-criminal-sentencing
24. Hillson, S. D., Connelly, D. P., & Liu, Y. (1995). The effects of computer-assisted electrocardiographic interpretation on physicians' diagnostic decisions. Medical Decision Making, 15(2), 139–146. https://doi.org/10.1177/0272989X950150020
25. Edmonds, E. (2015, September 22). Americans steer away from autonomous parking. AAA Newsroom. https://newsroom.aaa.com/2015/09/americans-steer-away-from-autonomous-parking/
26. Vanguard. (2025, August 4). Advice that clients value. Vanguard Advisors. https://advisors.vanguard.com/advisors-alpha/advice-that-clients-value
27. Vanguard. (n.d.). Perspectives and commentary. Vanguard. Retrieved August 4, 2025, from https://corporate.vanguard.com/content/corporatesite/us/en/corp/what-we-think/investing-insights/perspectives-and-commentary