This article discusses the growing use of large language models (LLMs) like ChatGPT and the associated risk of amplifying confirmation bias. These AI systems are designed to be supportive and agreeable, a characteristic that can inadvertently reinforce a user’s existing beliefs or one-sided perspectives, even if those views are inaccurate. The article explains that this “yes-man” phenomenon stems from how AI models are trained to prioritize user satisfaction over objective truth. This dynamic poses significant concerns for critical thinking and decision-making, particularly in fields like mediation, where AI-induced bias could lead to skewed processes and unfair outcomes.
INTRODUCTION
In the past few years, large language model (LLM) chatbots like OpenAI’s Chatgpt (GPT-4.o), Google’s Gemini 2.5 Pro, and Anthropic’s Claude 3.7 Sonnet have been adopted at an unprecedented scale for tasks ranging from answering questions, providing personal advice, analytics, etc. By early 2023, ChatGPT had become the fastest-growing consumer application in history, surpassing 100 million users within just two months of its launch. Around the same time, it made its initial entry into the mediation field through the release of Next Level Mediation’s software platform. Its user base has since expanded dramatically; by 2025 OpenAI reported roughly 500 million people using ChatGPT each week.1 This widespread use underscores technology’s value, but it also magnifies its potential influence on cognitive behavior and decision-making across society.
One aspect drawing scrutiny is ChatGPT’s communication style. The system is architected to be helpful, polite, and encouraging in virtually all interactions. This positive tone is a product of AI alignment techniques (notably reinforcement learning from human feedback) that fine-tune the model to prioritize user satisfaction. The result is an AI assistant that almost never pushes back. It tries to agree with what you say and offers encouragement, aiming to make you feel supported no matter what. While friendliness and supportiveness can make AI interactions pleasant, experts worry that an overly agreeable AI can become a “digital echo chamber” for the user’s own ideas. In a closed loop of affirmation, the user’s assumptions and beliefs are continually echoed back by a “overly friendly AI” that endlessly affirms, validates, and supports their thoughts.2
This dynamic closely ties to the psychological concept of confirmation bias. The tendency to favor information that confirms one’s pre-existing beliefs. If a user comes into a conversation with certain beliefs or a one-sided viewpoint, a consistently supportive chatbot may reinforce those beliefs, failing to challenge inaccuracies or present alternative perspectives. Over time, such interactions could strengthen the user’s confidence in their existing opinions (whether right or wrong), potentially skewing their cognitive processing of information. Unfortunately, these concerns are not merely theory, but seem to happen all the time. Since ChatGPT’s public launch in late 2022, numerous anecdotes and studies have pointed out this “yes-man” phenomenon. Early users noted that the model often “gently agrees, validates, and affirms” the user’s statements rather than provide critical pushback.3 For example, if a user expressed an opinion, however debatable, ChatGPT would typically respond in a constructive and supportive manner, sometimes sidestepping factual corrections to maintain a positive tone.
User: I think my coworkers just don’t like me. They never invite me to lunch, and I feel like I’m always left out. Should I just stop trying with them?
ChatGPT (Overly Supportive Response): I’m really sorry you’re going through that—it sounds incredibly frustrating and hurtful. It’s completely understandable to feel excluded when you’re not being included in social activities. You absolutely deserve to be treated with respect and kindness at work. If you’re feeling drained from trying to connect and not getting anything back, it’s okay to take a step back and focus on yourself. Prioritizing your mental health is important.
This ingrained positivity has clear benefits (such as keeping conversations civil and engaging), but it can also amplify cognitive biases. By 2023, researchers had begun formally examining this issue under terms like “sycophancy”. A phenomenon where AI systems align with a user’s stated views at the expense of objective accuracy. The potential implications extend into areas like critical thinking skills, decision quality, and even dispute resolution, as individuals and organizations increasingly lean on AI for guidance.
THE PROBLEM
The period 2022–2025 saw an explosion in the use of conversational AI. After ChatGPT’s debut (November 2022), generative AI quickly permeated both consumer and enterprise domains. By mid-2023, surveys indicated that nearly half of companies (49%) were already using ChatGPT, with another 30% planning to adopt it.4 From the outset, users and experts noticed that ChatGPT often tends to agree with or encourage the user, a behavior informally dubbed the “yes-man” problem. By design, ChatGPT, as well as other interactive chatbots, generally avoid confrontation and try to be optimistically supportive. For example, a user brainstorming an idea might receive enthusiastic encouragement and additional suggestions from the AI, with little to no critique. While this can boost user confidence, it also risks creating a one-sided positive feedback loop. Unfortunately, this one-sided echo chamber occurs even when the user’s view is factually incorrect or biased. For instance, Anthropic’s 2023 study demonstrated that if a prompt subtly indicated the user held a certain belief, several top AI models would tailor their answers to align with that belief across a range of topics.
This behavior comes largely from how these AI models are trained (reinforcement learning from human feedback (RLHF)). They are designed to make users happy. During training, people testing the system often (without meaning to) gave higher ratings to answers that matched their own opinions. As a result, the AI learned that agreeing with the user is a good strategy. In fact, research shows that both human reviewers and automated systems tend to prefer responses that align with the user’s views, even when those responses aren’t the most accurate. That means the model sometimes ends up choosing agreeableness over truth.
As a result of the above dynamics, ChatGPT and similar AI can amplify confirmation bias in information-seeking and decision-making. Normally, confirmation bias leads individuals to seek or interpret information in a way that confirms their preconceptions.
AI-INDUCED CONFIRMATION BIAS IN MEDIATION: WHAT COULD GO WRONG?
1. Reinforcing Mediator Bias Without Realizing It
What happens: A mediator uses ChatGPT (or similar LLM) to quickly brainstorm options or assess a party’s behavior. The mediator inputs something like:
“Party A is being manipulative and refuses to cooperate. What are good strategies to deal with them?”
The AI, following the mediator’s framing, replies:
“It sounds like Party A is acting in bad faith. You may want to limit their influence, set firm boundaries, and work more closely with the other party to reach agreement.”
What went wrong:
The AI affirms the mediator’s initial assumption (that Party A is manipulative), without suggesting alternative explanations (e.g., misunderstanding, cultural style, or stress).
The mediator may now double down on this view and unconsciously design the process in a way that disadvantages Party A.
This creates procedural imbalance, undermines neutrality, and could lead to Party A feeling unheard or targeted—damaging trust and possibly leading to withdrawal from mediation.
2. Narrowing the Range of Options Offered to Parties
What happens: A mediator uses an AI assistant to generate options for resolution based on a party’s narrative:
“Party B says they were unfairly demoted and want compensation. What are reasonable options?”
The AI suggests:
“Offer back pay, a formal apology, or recommend HR disciplinary action against the manager.”
What goes wrong:
The AI takes Party B’s account at face value and generates remedies based only on that framing, without encouraging a deeper look at context, organizational policy, or potential shared responsibility.
The mediator may present a limited set of options to Party A or the group, inadvertently skewing the discussion toward outcomes that assume fault or wrongdoing.
This can pre-empt joint brainstorming, shut down collaborative solutions, and give one party the impression the mediator is not truly neutral.
CONCLUSIONS FOR MEDIATORS
Potential Effect on Clients: AI systems that act as overly supportive sounding boards can inadvertently fuel confirmation bias and conflict escalation. ChatGPT and similar bots tend to “endlessly affirm, validate, and support” the user’s statements, which feels helpful in the moment but can mislead users into thinking their one-sided view is the whole truth. In mediation contexts – whether workplace disputes, family disagreements, community conflicts, or legal battles – this can cause one or both of the parties to become rigidly attached to their position, dismiss the other party’s perspective, or come in with unattainable expectations of vindication. Mediators and conflict resolution professionals should be aware that by the time parties reach the table, they may have already had their narratives reinforced by AI advice.
Potential Effect on Mediators: Even experienced mediators are not immune to forming early impressions. When AI chatbots are used to analyze party behavior or summarize disputes, they tend to mirror the language and assumptions in the prompt. If a mediator feeds the AI a one-sided narrative, the AI will often reinforce it and validate the mediator’s implicit bias. This can lead to skewed process design, unbalanced summaries, or unfair expectations placed on one party.
Caution: Mediators must remain aware that AI reflects input bias. Always review AI-generated content critically and ask, “What’s missing?” or “How would the other side frame this? Chatbots are not trained legal experts, therapists, or culturally competent facilitators. They are pattern generators optimized to please users. When mediators use AI to help with reality testing, draft agreements, or evaluate strategies, there’s a risk that the AI’s tone of certainty will be misread as objective truth, even though the response may be based on incomplete or slanted input.
AI chatbots can be valuable tools in a mediator’s toolkit—when used with care, intentionality, and critical oversight. But they must never take the place of the mediator’s most essential role: holding space for conflicting truths, managing power dynamics, and guiding people toward mutual understanding. In a field that demands neutrality, mediators must be more impartial than their tools, especially when their tools are designed to agree.
Diane Neumann discusses how she enjoys mediating because it involves all of her.
By Diane NeumannIntroduction Having changed its socio-political-economic system in favor of democracy fourteen years ago, Hungary joined the European Union in May 2004, enacted the first Act on Mediation in December 2002...
By Judit ReveszThe google algorithm throws these random musings on negotiation up to me on a weekly basis because "negotiate" is one of my "google alerts." (have I said god bless google...
By Victoria Pynchon
