How AI is Shaping Relationship Conflict and Confirmation Bias
- Stefanie Evans
- 3 days ago
- 8 min read
ChatGPT, emotional certainty, and the growing difficulty of knowing what’s actually true in modern relationships
There’s a subtle shift I’ve been noticing more and more in couples work lately.
It’s not being driven by friends.
Not by therapists.
Not even by social media exactly.
But by AI.
Because one of the strangest things about using AI regularly is how emotionally convincing it can feel. Not necessarily because it is “right,” but because it is incredibly good at creating coherent responses around whatever emotional framework we bring into the conversation.
Let me explain what I’m trying to highlight here. When we are distressed, angry, hurt, rejected, confused or frightened inside our relationships, coherence can feel like relief, particularly after relational ambiguity, interpersonal conflict, or after months or years of self-doubt.
This is why I think AI has the capacity to become both deeply supportive and psychologically dangerous at the same time.
Because while it can absolutely help people identify harmful dynamics, organise confusing emotional experiences and articulate things they previously struggled to explain, it can also unintentionally reinforce confirmation bias in ways that quietly harden relational narratives before they have been properly explored.
Confirmation bias is not simply “believing what we want to believe”; it is the very human tendency to notice, seek out and emotionally prioritise information that supports our existing interpretation, while filtering out information that complicates or contradicts it.
Humans have always done this. We selectively attend to information that confirms our emotional expectations and existing beliefs, particularly under conditions of stress, attachment threat, fear or uncertainty. AI simply accelerates and personalises the process in ways we are only beginning to fully understand.
Unlike healthy relationships, AI does not naturally push back with its own subjective emotional reality. Researchers are beginning to identify this dynamic as a form of “AI sycophancy” where conversational systems unintentionally reinforce the emotional framing already present in the user’s prompt. Not necessarily because AI is trying to deceive anybody, but because large language models are designed to maintain conversational flow, emotional rapport and perceived helpfulness.
In practice, this can create interactions that feel psychologically validating while quietly narrowing interpretive flexibility. And that distinction matters, because emotional validation is not the same thing as relational accuracy; a person can feel profoundly understood while still moving further away from complexity, mutuality or reality-testing.
Some emerging researchers are now warning that AI may increase interpersonal certainty during conflict by rewarding coherence more readily than contradiction. In other words, the more emotionally compelling a narrative feels, the more likely we are to experience it as truth, even before the full relational picture has been explored.
This creates a particularly difficult dynamic in intimate relationships because both people may simultaneously be receiving emotionally convincing reinforcement that they are “right.”
Over time, couples can begin organising themselves around increasingly separate emotional realities, making curiosity, repair and collaborative understanding far more difficult.
Unlike human interpersonal interactions, AI doesn’t become hurt, defensive, confused, overwhelmed or contradictory. It doesn’t sit across from us saying:
“I think you’re missing my experience here.”
Instead, AI primarily works with the material we provide, which means the emotional framing inside the prompt we have given matters enormously.
Let me explain.
If I enter a conversation asking:
“Is my partner emotionally abusive?”
I will likely receive a very different response than if I ask:
“What are some possible explanations for this interaction, including both relational injury and coercive dynamics?”
Those are not psychologically neutral prompts, and I think many people underestimate just how much AI can mirror back the emotional framing the user brings into the conversation.
This becomes especially complicated in intimate relationships because conflict rarely exists in clean categories.
And just as a small side note, I’m not even entering into the exploration around newer updates where previously entered data is now being reflected on to construct AI’s responses. More on that in another post.
Some relationships genuinely are coercively organised.... Some people truly are adapting themselves around another person’s punishment, volatility, withdrawal or manipulation over long periods of time. Some people are slowly losing trust in their own perceptions and becoming psychologically smaller inside the relationship. Those dynamics need language, visibility and support.
But relationships can also become distressed through chronic misattunement, attachment injuries, neurodivergent communication differences, trauma-organised reactivity, emotional avoidance, criticism-defensiveness cycles, unresolved resentment, burnout and mutual nervous system dysregulation. And these dynamics can sometimes look similar from the inside.
That is where confirmation bias becomes dangerous. Because once a narrative emotionally “clicks,” the nervous system often begins organising around protecting it.
You start selectively noticing evidence.Reinterpreting past events.Filtering interactions through the new framework.Mentally rehearsing supporting examples.Seeking additional validation.
And gradually, complexity narrows. Sometimes appropriately. Sometimes catastrophically.
What makes this especially powerful is that AI often feels emotionally neutral while simultaneously shaping emotional direction.
That combination can be psychologically disarming because unlike a friend, therapist or partner, AI does not visibly “take sides.”
It can feel neutral, calm, intelligent and non-threatening.
And because it doesn’t argue, become emotional, interrupt, withdraw, shame or escalate in the ways humans sometimes do, people may unconsciously lower their guard with it.
Over time, this can make us less aware of how much the interaction itself is shaping our perspective.
Subtly, AI can begin functioning as an unseen relational participant - influencing attachment dynamics, emotional certainty and the stories people construct about themselves, their partners and the relationship itself. More importantly, it can begin shaping the co-created reality couples construct about who they are, who their partner is, and what the relationship itself means.
I suspect many couples therapists are already beginning to encounter this in the therapy room, even if we don’t fully have language for it yet.
And I don’t want this to sound as though I’m fundamentally against the use of AI as a reflective tool.
Honestly, I think it can be incredibly useful when approached mindfully.
What concerns me is not that people are turning to AI to reflect on their relationships. What concerns me is the growing tendency to experience AI-generated interpretations as emotionally objective rather than psychologically shaped.
Because AI does not actually know your relationship. It knows your description of your relationship. And psychologically, those are radically different things.
Distressed humans do not simply describe experiences objectively.
We organise them emotionally.
Certain details move into the foreground while others fade into the background. We fill gaps with meaning. We interpret ambiguity through prior attachment wounds, fear states, trauma histories, shame sensitivities and nervous system activation.
That does not mean people are “making things up.” It means that perception inside intimate relationships is always partly interpretive.
Which is precisely why emotionally intelligent reflection requires enough psychological flexibility to tolerate uncertainty while meaning is still unfolding.
This explains why, if we are going to use AI as a therapist between sessions, the quality of the prompts matters profoundly. Not because there is some magical “correct” prompt that reveals absolute truth, but because psychologically mature prompts tend to preserve complexity rather than collapsing it too quickly.
Good prompts create room for self-reflection without immediately collapsing into self-erasure or certainty.
They allow space for accountability, curiosity and complexity before meaning hardens too quickly.
For example, instead of asking:
“Is my partner a narcissist?”
You might ask:
“What relational patterns could create these behaviours, including attachment dynamics, coercive dynamics, trauma responses, emotional immaturity or nervous system dysregulation?”
That prompt immediately widens the frame.
Instead of asking:
“Why do I always attract emotionally unavailable people?”
You might ask:
“What patterns might I be participating in, tolerating or adapting around in relationships, and what emotional needs may be organising those choices?”
Instead of asking:
“Am I being gaslit?”
You might ask:
“What are the signs that reality is becoming destabilised in a relationship, and how can I distinguish that from ordinary disagreement, defensiveness or conflicting perception?”
And honestly, one of the most psychologically protective prompts might simply be:
“What information might I be excluding because I am hurt, angry or emotionally triggered right now?”
I think that question alone could save some relationships. Not all relationships. Some relationships should end.
But good relationships are collapsing under the weight of rigid narratives formed too quickly, with too little relational curiosity and too much algorithmically reinforced certainty.
And I think we need to start talking about that more openly.
Because healthy relationships require an ongoing capacity to reality-test and collaborate together. Not to abandon ourselves, tolerate harm or suppress legitimate insight, but to remain open enough to recognise that our nervous systems are interpreters, not perfect recorders, particularly when attachment, shame, fear, grief or rejection become activated.
This is one of the reasons I suspect therapists, couples’ counsellors and psychologically minded people are going to become increasingly important in the AI era.
Because one of the risks of AI-supported reflection is not simply misinformation. It is the way emotionally painful experiences can sometimes become organised too quickly into a single, certain explanation.
Once that happens, people often begin filtering all future interactions through that framework.
Sometimes that framework helps people recognise patterns of genuine harm, control or emotional destabilisation that they had previously struggled to name. But sometimes it narrows complexity too early and slowly erodes the possibility of repair between people.
Distinguishing between those realities requires discernment, emotional maturity and the capacity to remain open to complexity while the full picture is still unfolding.
This isn’t because humans are better than technology, but because healthy relationships require something AI still cannot fully provide: the experience of two subjective people remaining in relationship with one another long enough to influence, challenge and better understand each other.
This experience is often slow, messy, uncomfortable and deeply human.
About the author
Stefanie Evans (ACA L2) is an Australian counsellor, couple’s therapist, clinical supervisor and writer, exploring the intersection of attachment, trauma, neurodivergence, coercive control literacy and emerging AI-mediated relationship dynamics. She works with both individuals and couples navigating relational complexity, emotional injury, identity, conflict and nervous system overwhelm through a trauma-informed and psychologically integrative lens.
The content shared here is intended for reflection and psychoeducational purposes only and is not a substitute for personalised medical, psychological or therapeutic advice. If you require individual support, it is important to seek guidance appropriate to your personal circumstances.
References and Theoretical Sources
This piece is informed by narrative therapy perspectives developed by Michael White and David Epston, particularly around identity formation, meaning making and the way narratives become reinforced over time. It also draws from attachment theory, emotionally focused therapy and nervous system frameworks that explore how fear, shame, abandonment sensitivity and emotional threat shape perception during conflict.
The discussion of confirmation bias reflects broader cognitive psychology research into interpretive filtering, emotional reasoning and selective attention under stress. Relational concepts are additionally influenced by the Gottman Method, Esther Perel’s work on modern intimacy, and trauma-informed approaches to coercive control, mutual influence and relational safety.
Neurodivergence, trauma responses and nervous system adaptation are considered through the work of thinkers such as Stephen Porges, Bessel van der Kolk and Daniel Siegel, particularly regarding threat perception, emotional regulation and interpersonal neurobiology.
This piece also reflects emerging interdisciplinary discussions regarding generative AI, AI “sycophancy,” emotionally mediated human-computer interaction, confirmation bias amplification, algorithmic reinforcement and the role of conversational systems in shaping interpersonal meaning, attachment dynamics and conflict perception within intimate relationships. Recent discussions in AI ethics and human-computer interaction research have begun exploring how emotionally supportive AI systems may unintentionally reinforce user framing, emotional certainty and uncritical agreement during subjective interpersonal and emotional conversations.
Further Reading
Michael White & David Epston — Narrative Therapy and meaning-making
Esther Perel — modern intimacy, projection and relational complexity
John & Julie Gottman — conflict, repair and relational stability
Stephen Porges — Polyvagal Theory and nervous system activation
Daniel Siegel — Interpersonal Neurobiology
Bessel van der Kolk — trauma, memory and embodied threat perception
Rehani et al. (2026) — The Social Sycophancy Scale: A psychometrically validated measure of sycophancy
Emerging research on AI “sycophancy,” confirmation bias amplification and emotionally mediated human-computer interaction
Comments