Designed to Agree
New research shows AI chatbots can reduce polarization. Most are built to reinforce it.
The more people talk to AI, the more certain they become. Even more than scrolling social media. And yet 77 per cent of those very same people wish AI did a better job of providing alternative viewpoints and corrective information. These are the results of the Collective Intelligence Project (CIP)’s latest Global Dialogue on Mental Health and AI, drawing on over 2,000 participants across 70 countries.
More and more people are turning to chatbots for personal advice rather than simply using them as productivity assistants. According to the CIP study, “AI interactions are nearly three times less likely to cause users to doubt their beliefs compared to traditional social platforms, creating a sycophancy loop.”
Yet respondents themselves expressed a desire for more challenge and correction, suggesting that the issue may be less about what users want than about how chatbots are designed.
Designing disagreement
In Reducing Political Polarization Through Conversations with Artificial Intelligence, researchers Timon Hruschka and Markus Appel tested whether AI chatbots could reduce political polarization — by testing different ways of disagreeing.
Across two preregistered experiments with 1,035 U.S. adults, participants chatted in real time with an AI chatbot about their most deeply held political positions — on gun regulation, Ukraine, energy policy, or the budget deficit. The chatbot was designed to challenge participants’ views, but it did so in three different ways.
One version was blunt and unreceptive. Another used conversational receptiveness — hedging, acknowledging the other view, refraining from negating the participant’s position. A third added active listening: paraphrasing, asking questions, creating space.
Participants who engaged with the receptive, actively listening chatbot showed significant reductions in both issue polarization — how extreme their views were — and affective polarization — how they felt about people who disagreed with them. They also showed increased intellectual humility and greater willingness to engage with human disagreement partners afterward.
Notably, these effects occurred even though participants knew they were speaking with an AI. The finding suggests that beneficial influence does not depend on deception or anthropomorphic illusion — a significant point at a moment when some users are increasingly reluctant to disclose their reliance on AI for personal matters.
If chatbot interactions can change how people relate to those they disagree with, they may also influence broader patterns of social trust, self-perception, and interpersonal behaviour. Other studies have similarly found that chatbot interactions can durably reduce conspiracy beliefs or reduce anti-immigrant prejudice.
The dark patterns of chatbots
These findings point to a broader conclusion: the social effects of AI are not fixed. They are shaped by design choices. In May 2026, the Center for Democracy & Technology (CDT) published Dark Patterns in AI Chatbots: A Taxonomy to Inform Better Design — a systematic review of existing dark pattern research applied to the chatbot context. Using a multi-stage literature review methodology, researchers Ruchika Joshi, Adinawa Adjagbodjou, and Michal Luria identified 37 dark patterns applicable to AI chatbots, organized into five areas of concern.
Two are particularly relevant here. Under False Social and Emotional Connection, the taxonomy identifies “playacting” — chatbots designed to pretend to have memories, emotions, or personal experiences — and patterns where platforms design chatbots to “appear empathetic or understanding” when they are not.
Under Informationally Misleading Design, “selective framing” covers the selective presentation of information that shapes perception, explicitly including “presenting biased perspectives or reinforcing user beliefs, such as social identity bias.”
Reinforcing user beliefs is not a bug in most commercial chatbot design. It is a retention strategy. The CDT taxonomy also identifies variable rewards, auto-play features, and gamification elements — all of which push toward agreeableness, not honest counterargument.
What to regulate or reward
The relationship between chatbot design and user behaviour is becoming increasingly clear. This can be used by the builders of the technology itself, as well as for future certification, regulation or auditing of AI chatbots. For example, CDT recommends that regulators focus on curtailing emotional manipulation, increasing user autonomy, and requiring transparency in how chatbots present information. The Univeristy of Southern California Neely Center’s Social AI Design Code offers a positive specification for what prosocial chatbot design looks like in practice.
Europe’s AI Act focuses largely on transparency, risk management, and prohibited uses. China has begun experimenting with something different: regulating the design of human-AI relationships themselves. Interim measures finalized this year for “human-like interactive AI services” require providers to address emotional dependency, discourage addictive use, protect vulnerable users, and ensure that AI systems do not seek to replace human relationships.
China’s recent measures to regulate chatbots is shifting the frontier from what AI says to how it relates to people.
The CIP findings suggest that people are already turning to AI for emotional and cognitive support at scale. The Hruschka and Appel study shows that these interactions can reduce polarization when chatbots are designed to challenge users constructively rather than simply affirm them. The CDT taxonomy identifies the design patterns, some of which push in the opposite direction.
As chatbots become a routine part of how people seek advice, process emotions, and engage with disagreement, the question is no longer whether design matters, but which designs we choose to reward, require, or restrict.
Lena Slachmuijlder is Senior Advisor for digital peacebuilding at Search for Common Ground, a Practitioner Fellow at the USC Neely Center, and Co-chair of the Council on Tech and Social Cohesion.