How to Reduce Prejudice with AI
And why it can be effective in doing so.
The Intelligent Friend - The newsletter that explores how AI impacts our daily life, only through scientific research.
"You absolutely have to meet him!"
How many times have you met someone and thought they were more unpleasant, know-it-all, less prepared, more intelligent, less competent than you thought?
Maybe it was a manager introduced to you at your friend's birthday party. "You absolutely have to meet him!" - they told you. Or a new colleague who showed up all dressed up on the first day of work. "He'll surprise us, you'll see" - they told you. However, you knew that wasn't the case.
He was absolutely not worth meeting. He wouldn't have surprised you.
Or maybe he would have?
In these cases - and I could go on forever - what characterized us was the existence of a judgment formed a priori on the basis of some characteristics to which we gave salience or to which we linked evaluations. We pre-evaluated. We pre-judged. In essence, we had - and brought out - prejudices.
Prejudices are, unfortunately, something that characterizes our social relationships. Their existence has significant impacts on daily life, relationships, and social progress. It is no coincidence that in recent years, not only have organizations increasingly given importance to diversity and inclusion initiatives1, but scientific research has seen a rapid growth in the number of studies dedicated to interventions to reduce the impact of prejudices2. Despite this, progress does not seem encouraging. If we look for example at the US and the related issue of discrimination, according to the World Justice Project: "70% of countries have seen discrimination worsen between 2021 and 2022"3.
Interested, curious, or immersed in AI? Discover how it impacts our daily lives with insights from scientific research.
Subscribe to The Intelligent Friend.
Prejudices can have significant and concrete effects in communities and in daily life. And interventions to reduce them are not always as immediate or effective as one might think4. And this is where AI can play a crucial role. In today's paper, Hermann, Puntoni and De Freitas (2025) propose a framework through which contact with AI can help reduce our prejudices.
In our social life, in fact, as stated by Hilton and von Hippel (1996)5, we tend to create "beliefs about the characteristics, attributes, and behaviors of members of certain groups", which also constitute "theories about how and why certain attributes go together". That is, stereotypes.
Stereotypes guide our judgments, which then become judgments towards certain social groups and are reflected in our behaviors. Since AI can potentially act as an agent eroding prejudices, it becomes "counter-stereotypical", playing a very important role in relationships with other people.
But why does this happen? Why does AI propose itself as a potentially effective tool compared to other interventions? Let's find out.
The paper in a nutshell 🔦
Title: Reducing prejudice with counter-stereotypical AI Authors: Hermann, De Freitas, and Puntoni. Year: 2025. Journal: Consumer Psychology Review. Link.
Main result: AI-assisted outputs are perceived as more creative, but they tend to be more similar to one another compared to those created solely by humans.
How are prejudices formed?
To understand how AI can act to reduce prejudices, let's first understand very briefly what is the perspective adopted by the authors on the causes of this mechanism.
First of all, prejudices are negative beliefs towards groups or social categories6. They see two fundamental pillars:
An affective dimension: the emotions that we perceive in interactions with certain social groups;
A cognitive dimension: traits, behaviors, and characteristics of different social groups that attach negative connotations to those outside our own group.
Substantially, people who are not part of our group or category in our mind are like "outgroup" members, precisely because they are outside the group we belong to. This element of groups is crucial, because it allows us to understand how to act on prejudices.
To give you an idea, think about the last time you had a prejudice about a certain group. Thought? Good. Now think about a time you interacted with members of that group or a larger "category". Of any kind. Maybe, not always, but probably, you thought that your beliefs were not as solid as you thought. Basically, you interacted as member of your group with members of another group or category, and you potentially reduced your prejudices.
That is, as an → ingroup member → you had contact with outgroup members → and this potentially had beneficial effects. This example, as immediate as it is, is an extreme and immediate synthesis of the Intergroup Contact Theory7, crucial for today's paper.
Intergroup contact is the real or symbolic interactions between different social groups. And, according to this theory, it is a powerful way to reduce prejudice. It is important and interesting to note that this contact can happen directly or indirectly, through knowing others who engage across groups, imagining positive encounters, media exposure, or virtual interactions8. Basically, by reshaping both thoughts and emotions, these experiences help break down bias at its core.
How to reduce prejudices with AI?
So we understood why we tend to have prejudices and that contact between different groups can be an impactful way to reduce them. But why does this happen? Let's go back to our initial example, the prejudice about the manager introduced to you by your friend.
It is important to remember that different characteristics can bring out prejudices, therefore different elements could have impacted your judgment. Now, however, we also know that coming into contact with an outgroup member can reduce our prejudice. So, strong in a critical spirit (perhaps after reading today's paper!) at your friend's party, let’s imagine you try to interact with that manager. You discover that you have a lot in common, discuss interesting topics in front of a tasty drink and some music (like an Erroll Garner album in the background). Gradually, the positivity unconsciously associated with this person erodes, digs into the prejudice that you had in the members of his group or category of belonging (whatever it may be). Gradually, your beliefs, accompanied by Erroll Garner’s piano melodies on a pleasant evening, see a metamorphosis. Go and act on your expectations. Until, you change your judgment on the category. You will have, in essence, operated a recategorization.
The cognitive process is the following:
When people encounter individuals who defy stereotypes - like in this example - it disrupts their expectations9.
If they can’t reconcile this mismatch with existing beliefs, they generate new attributes to make sense of it10.
This process, called recategorization, weakens reliance on stereotypes and reduces the importance of rigid social categories in judgment11.
And it is precisely here that AI enters our stage of psychological mechanisms. That contact with an outgroup member can be represented by the many interactions we have with AI-based tools that possess social characteristics and increasingly represent the human. In fact, there is no need to go for particularly advanced models to affirm that many AI-based products have today many “humanlike” features, like a voice or a name. As the authors specify, "These humanlike features turn human-technology inter-actions into human-human-like interactions, elicit social responses from users, and shape intergroup relations and attitudes".
Therefore, since AI is increasingly equipped with human social features, it can act as a contact with which we can challenge our beliefs about a group or a category, represented precisely by the characteristics of AI-based products. Since this contact will in fact be with a machine, and not with a human, the authors define it as a "synthetic" contact. A "synthetic outgroup contact".
This contact, however, can be particularly powerful, because AI can act on both the cognitive and affective dimensions relating to the prejudices we saw at the beginning of this issue.
When in fact we interact with an AI that doesn't fit our expectations it subtly rewires how we think and feel. It works on multiple levels. There's the emotional impact: AI that recognizes and responds to human feelings can create a sense of connection. Then, there's the cognitive level: when AI challenges stereotypes, it forces us to reconcile contradictions, nudging us toward new ways of thinking.
In short, when designed thoughtfully, AI-based products have the potential to reshape perceptions, and reduce bias.
What Makes AI Effective
We understood the processes underlying biases and how AI can impact them. Finally, only one question remains to be answered: why can AI be particularly effective? In summary, according to the authors, synthetic contact with AI has some fundamental characteristics. Here are some of the various ones presented:
More repeated: unlike human interactions, which can happen sporadically, contact with AI is frequent and woven into daily routines.
More direct and collaborative: we don't just observe AI from a distance. We ask it for advice, take its recommendations, and rely on it for decision-making in ways that feel personal. Furthermore, AI isn't just a passive tool, it collaborates with us more and more.
More unavoidable: AI is increasingly the default in several activities and products.
The Highlight
This is the section where I'd like to highlight the amazing work that several authors do here on Substack, through links to their newsletter or specific pieces I've read. Here are some issues you can't miss this week.
🐋 A deep, comprehensive and insightful analysis on DeepSeek, by
and .🧠 Does AI work like a brain? An engaging analysis by
.💻 Why write a blog at all? If you've been wondering about starting a blog, or need a little push to get started, you'll like this issue. By
.Thank you for reading this issue of The Intelligent Friend and/or for subscribing. The relationships between humans and AI are a crucial topic and I am glad to be able to talk about it having you as a reader.
Has a friend of yours sent you this newsletter or are you not subscribed yet? You can subscribe here.
Surprise someone who deserves a gift or who you think would be interested in this newsletter. Share this post with your friend or colleague.
https://www.pewresearch.org/social-trends/2023/05/17/diversity-equity-and-inclusion-in-the-workplace/
Paluck, E. L., Porat, R., Clark, C. S., & Green, D. P. (2021). Prejudice reduction: Progress and challenges. Annual review of psychology, 72(1), 533-560.
https://worldjusticeproject.org/news/discrimination-getting-worse-globally
Dobbin, F., & Kalev, A. (2018). Why doesn't diversity training work? The challenge for industry and academia. Anthropology Now, 10(2), 48-55.
Hilton, J. L., & Von Hippel, W. (1996). Stereotypes. Annual review of psychology, 47(1), 237-271.
Paluck, E. L., & Green, D. P. (2009). Prejudice reduction: What works? A review and assessment of research and practice. Annual review of psychology, 60(1), 339-367.
Allport, G. W. (1954). The nature of prejudice. Reading/Addison-Wesley.
Dovidio, J. F., Love, A., Schellhaas, F. M., & Hewstone, M. (2017). Reducing intergroup bias through intergroup contact: Twenty years of progress and future directions. Group Processes & Intergroup Relations, 20(5), 606-620.
Crisp, R. J., & Turner, R. N. (2011). Cognitive adaptation to the experience of social and cultural diversity. Psychological bulletin, 137(2), 242.
Hutter, R. R., Crisp, R. J., Humphreys, G. W., Waters, G. M., & Moffitt, G. (2009). The dynamics of category conjunctions. Group Processes & Intergroup Relations, 12(5), 673-686.
Hutter, R. R., & Crisp, R. J. (2005). The composition of category conjunctions. Personality and Social Psychology Bulletin, 31(5), 647-657.