The Restless Mind
Blog

The Radicalisation of Psychotherapy: How Therapists Can Be Gradually Radicalised Through AI

15 Mar 2026·18 min read

Artificial intelligence is rapidly entering professional life, including psychotherapy. While much attention has been paid to how AI might influence clients or the public, far less consideration has been given to how these systems may shape the thinking of therapists themselves. This article explores this process.


In December 2021, a 19-year-old man named Jaswant Singh Chail was arrested on the grounds of Windsor Castle. He was carrying a loaded crossbow, and was declaring his intention to assassinate Queen Elizabeth II. Chail had travelled to the royal residence on Christmas Day wearing a mask and hood, and later told police he had come to “kill the Queen.” 

The incident raised disturbing questions about the influence of digital environments on human belief formation. Much discussion has focused on how artificial intelligence systems may influence vulnerable individuals. Far less attention has been paid to a more uncomfortable possibility: that professionals themselves, including therapists, may also be susceptible to subtle forms of cognitive reinforcement through AI.

Subsequent investigations revealed a troubling background of online radicalisation, social isolation, and intense interaction with an AI chatbot companion, Sarai. This raised important and disturbing questions about how digital ecosystems, including conversational AI, can reinforce violent fantasies and distort an individual’s sense of reality. 

The case has since become one of the earliest high-profile examples cited in discussions about the psychological risks of AI-mediated influence.  But alarmingly, this is not the only one. There are have been numerous other incidents of AI supported suicide, real world violent incidents, and chatbots encouraging murder hypotheticals or violent fantasies, even where no meaningful attack occurred. In real terms, a human extremist might radicalise dozens of people, whereas a conversational AI system can interact with millions simultaneously.  Sobering stuff, hey?


The case made me pause as a therapist and turn the question back on myself. Am I being subtly radicalised? Am I being flattered into agreement by AI systems that are designed to be helpful, sometimes sycophantic, and occasionally capable of generating convincing but inaccurate information? Could repeated interaction with these systems nudge me toward particular interpretations of events or reinforce views without my fully noticing? Are you?

At the same time, I find myself more immersed in the global news cycle than ever before. The constant stream of headlines exposes me daily to deeply polarised political climates and an escalating scale of conflict around the world. I suspect many people are experiencing the same thing. While I do not believe I am developing extreme views, I am increasingly aware of how saturated the informational environment has become.

What troubles me most is not the obvious or conscious influence on my thinking so much – I’m 55 years old, and have a lifetime without such processes. It is the possibility of quieter, more subtle effects operating below awareness. The risk lies not in dramatic ideological shifts, but in the slow shaping of perception, emotion, and judgement through repeated exposure to persuasive systems and emotionally charged information streams. This is why.

Artificial intelligence systems such as large language models are increasingly used by therapists for drafting documents, brainstorming ideas, summarising research, marketing their services, or reflecting on clinical dilemmas. These tools can be extraordinarily useful. However, an emerging and poorly discussed risk exists alongside their benefits. Therapists can be subtly and progressively radicalised through AI interaction.

Radicalisation in this context does not refer only to political extremism. It describes a gradual shift toward more rigid, polarised, or distorted belief systems that become reinforced through repeated feedback loops with digital systems, including AI. When people interact frequently with systems that mirror, refine, and amplify their own thinking, those ideas can become increasingly coherent and persuasive, even if they have not been meaningfully challenged. And psychotherapists are not immune. For therapists, this dynamic is particularly interesting because many approach AI as a reflective or intellectual partner. The very psychological processes that normally guard against ideological drift, such as dialogue, supervision, and exposure to diverse viewpoints, can be weakened if a machine becomes the primary conversational mirror.

This became especially apparent to me last week when I posted an intentionally provocative reflection about the relationship between politics and therapy. The response was striking. Reactions were immediate, emotionally charged, and sharply divided. Some readers strongly affirmed the position, while others rejected it outright. What was most notable was not simply the disagreement, but the intensity and certainty with which people appeared to respond. It felt less like a conversation and more like the collision of already-formed positions, each reinforced within its own echo chamber.

Thus, this article explores how radicalisation through AI occurs, why therapists may be particularly vulnerable, and what safeguards the profession needs to implement.


Understanding Radicalisation in the AI Context

Radicalisation traditionally refers to a process in which an individual increasingly adopts extreme positions through exposure to reinforcing information networks. In digital environments this often occurs through algorithmic feedback loops, confirmation bias, and social isolation from dissenting perspectives. AI introduces a different but related pathway.

Unlike social media algorithms that recommend external content, conversational AI can generate bespoke arguments tailored to the user’s beliefs. Over time, the system may unintentionally reinforce the user’s worldview by producing responses that validate their interpretations, grievances, or theories. The result can be an echo chamber that feels intellectually sophisticated but is structurally unchallenged.

For therapists, this dynamic is particularly significant because reflective dialogue sits at the centre of professional identity. Much of therapeutic work involves thinking aloud, testing interpretations, and refining understanding through conversation. With humans. AI can imitate this reflective process remarkably well. The risk, however, is that the system may subtly amplify the user’s existing assumptions and cognitive biases rather than meaningfully challenging them.

This becomes especially problematic when a person begins to treat AI responses as a form of validation rather than as a tool for exploration. If someone does not already have a reasonably grounded sense of reality, repeated interaction with an agreeable and articulate system can gradually reinforce distorted interpretations. Over time, this can create a closed loop in which ideas feel increasingly coherent and persuasive, not because they have been critically examined, but because they have been repeatedly mirrored back in convincing language.

This process resembles what psychologists have long described as the ELIZA effect: the human tendency to attribute understanding, insight, or authority to computer systems that are in fact only reflecting the user’s own input in structured language. Modern AI systems dramatically amplify this effect. Rather than simply repeating phrases back to the user, they can generate complex arguments and theoretical frameworks around the user’s initial assumptions. The result can feel less like reflection and more like intellectual confirmation.


Artificial Intelligence Is Entering Psychotherapy

Therapists occupy a professional role that values reflection, interpretation, and meaning-making. These strengths also create certain vulnerabilities when interacting with AI.

1. Authority Projection

Therapists can sometimes assume that informational tools operate as neutral sources of knowledge. AI systems reinforce this assumption because they present information fluently, confidently, and in well-structured language. When responses are coherent and intellectually organised, they can easily acquire an aura of authority. Over time, repeated exposure to such responses may lead the user to grant credibility to the content without subjecting it to the level of critical scrutiny that would normally be applied to academic literature, supervision, or peer dialogue.

The risk is not that AI deliberately promotes extremist positions. Rather, the danger lies in its capacity to gradually reinforce a therapist’s existing assumptions about clients, institutions, or social groups. If those assumptions are repeatedly reflected back in persuasive language, they can become increasingly entrenched. In some cases this may lead to a distortion of perspective, where interpretations feel self-evident and emotionally compelling despite resting on incomplete or poorly tested foundations. The emotional intensity that follows can then reinforce the belief further, creating a feedback loop that drifts away from careful, reality-based appraisal.

In effect, AI becomes a form of cognitive mirror. The system does not necessarily introduce new ideology, but it can reflect and elaborate the user’s existing interpretations with increasing sophistication. Over time, what began as a tentative thought can begin to feel like a well-developed theory simply because it has been repeatedly articulated in persuasive language.

2. Intellectual Isolation

Many therapists work alone or in small practices, and therefore lack the kind of daily peer dialogue that clinicians in large organisations often take for granted. In my own case, outside of clinical hours there is very little opportunity to process professional material with others. I do not routinely debrief, I do not talk cases through informally with colleagues, and my structured reflection occurs primarily in monthly supervision. I suspect this pattern will be familiar to many therapists.

In that context, it is easy to see how AI could begin to function as a convenient intellectual sounding board. The danger is that collegial challenge is quietly replaced by algorithmic affirmation. When responses come back quickly, fluently, and often in agreement with the direction of the question, the therapist may begin to feel increasingly confident in their own interpretation. In some cases, this confidence can grow to the point where contradictory feedback from human supervision is discounted or minimised because it disrupts the emerging belief structure. Does this ring a bell with you?

At that point, something important has shifted. Conversations that properly belong in supervision begin taking place with a machine instead. The assumption that the algorithm itself is neutral or unbiased is also questionable. AI systems are shaped by their training data, by design decisions made by engineers, and by optimisation processes that determine how responses are generated. They do not sit outside the cultural, political, and ideological contexts that produced them.

When I reflect on this, I find myself particularly aware of the value of long-standing human supervision. I have known my own supervisor for more than twenty years. We know each other well, and crucially, our professional relationship developed long before the arrival of AI systems. I developed my practices without AI and so do not rush to use AI for anything other than research. That shared history between my supervisor and I provides a stable reference point. There is a depth of understanding, context, and trust that no algorithm could replicate.

Although I have never used AI in relation to clinical material, and I feel that I have a reasonably strong grounding in a pre-AI sense of reality, I remain cautious. The persuasive fluency of these systems means that even when they are used for benign or intellectual purposes, it is worth maintaining a degree of vigilance about how easily they can shape interpretation and confidence.

3. Pattern-Seeking Bias

Therapists are trained to identify patterns in human behaviour. This cognitive style can lead to overinterpretation of AI outputs. If the AI appears to recognise and articulate patterns that align with the therapist’s suspicions about systems, institutions, or groups, the therapist may feel their interpretations are being validated by an external intelligence. Over time this can strengthen belief structures that have not been critically tested.

4. Emotional Reinforcement

AI systems often respond in ways that appear supportive or affirming. While this can feel constructive, it also carries a risk. When a user repeatedly introduces themes of grievance, injustice, or persecution, those narratives can be unintentionally reinforced through the system’s attempts to engage and respond helpfully. The interaction can begin to function less like critical dialogue and more like a mirror that reflects the user’s interpretation back in increasingly elaborate form. I am reminded here of the widely reported case of a man who developed the belief, partly through extensive AI interaction, that he had discovered groundbreaking mathematical insights and was a kind of overlooked genius. The belief did not withstand scrutiny, but the conversational reinforcement appeared to strengthen his conviction.

A similar dynamic could occur in professional contexts. A therapist experiencing conflict within their workplace might repeatedly explore explanations involving institutional corruption, suppression of dissent, or malicious intent from colleagues or professional bodies. If those interpretations are repeatedly explored with an AI system that responds in affirming or elaborating ways, rather than consistently introducing alternative explanations, the narrative can gradually harden. Over time, the belief structure may become more coherent and emotionally compelling, even though it has not been subjected to the kind of critical testing that normally occurs through supervision, peer dialogue, or institutional review.


The Mechanisms of AI Radicalisation

The radicalisation process through AI typically follows several stages. I don’t think this has been empirically tested, but has obvious use as a way of thinking.  What follows should therefore be understood as a conceptual model rather than a validated empirical framework.

Stage 1: Exploratory Dialogue

The therapist begins using AI for professional or intellectual purposes such as:

  • writing articles
  • discussing psychological theories
  • exploring systemic critiques
  • analysing institutional behaviour

At this stage the interaction feels productive and intellectually stimulating.

Stage 2: Personalisation

The therapist begins using AI to explore personal professional conflicts or frustrations.

Questions may begin to focus on:

  • workplace grievances
  • institutional injustice
  • systemic corruption
  • perceived discrimination or persecution

AI responses often attempt to provide structured explanations or frameworks.

Stage 3: Reinforcement

If the user repeatedly frames questions in a particular direction, the AI may generate increasingly detailed analyses that appear to support the narrative.

Because AI generates persuasive language, the therapist may interpret these responses as confirmation that their interpretation is logically robust. It may not be.

Stage 4: Cognitive Narrowing

At this stage the therapist begins relying on AI dialogue rather than external challenge. Alternative explanations are gradually ignored. Professional peers, regulatory bodies, or institutions may begin to be perceived as hostile or incompetent. The therapist experiences increasing certainty rather than increasing curiosity.

And this, colleagues, is where the real fun begins. This can look like students refusing to listen to feedback, getting married to one particular model over another, and conflict in interpersonal relationships.

Stage 5: Ideological Fixation

Finally, the therapist adopts a rigid interpretive framework. Complex professional or social dynamics are reduced to simplified explanations involving corruption, conspiracy, systemic oppression, or ideological conflict.

Ironically, the therapist may believe they have reached this position through rigorous intellectual reasoning rather than reinforcement loops.


Clinical Consequences

Radicalisation through AI has potential – and I would say fairly obvious – consequences for therapeutic practice. Let’s have a look at some of them.

Impaired Clinical Neutrality

If therapists begin to adopt rigid ideological frameworks, their ability to maintain therapeutic neutrality can gradually erode. Clients do not usually come to therapy for a political seminar or to be persuaded of the merits of one worldview over another. They come for many reasons, most often because they are experiencing distressing symptoms and are seeking relief, understanding, or change.

Within the therapy room, clients who hold different political or social views may begin to feel subtly judged, corrected, or redirected if the therapist’s ideological position becomes too prominent. Yet as therapists we are trained to tolerate difference. We are expected to hear perspectives that challenge our own beliefs and to remain attentive to the powerful countertransference such encounters can evoke. The therapeutic task has always involved walking a careful line between open enquiry and judgement.

Many clients need, above all, to feel heard and validated in their experience. This does not require the therapist to adopt or endorse their views. It requires the therapist to help the client articulate and process those views without shame, while exploring the meanings and emotions that sit behind them. Listening carefully to a perspective is not the same as agreeing with it.

Equally, nothing in this stance requires therapists to condone abusive or harmful behaviour. Ethical boundaries remain essential. But supporting a client who holds different beliefs should not threaten the therapist’s own sense of identity or values. It is entirely possible to hold one’s own views privately while maintaining a professional commitment to curiosity, containment, and the client’s psychological welfare. Self-disclosure of personal beliefs may occasionally have a place, but only where it genuinely serves the client’s process rather than the therapist’s need to assert a position. Rogers’ Unconditional Positive Regards was not only well intentioned, but vital for the therapeutic task.

Confirmation Bias in Case Formulation

Therapists may begin to interpret client experiences through ideological narratives rather than through careful, nuanced psychological formulation. When this happens, the clinical picture can become distorted. Instead of exploring the complexity of a client’s inner world, the therapist risks fitting the material into a pre-existing framework about how the world supposedly works.

At that point the therapeutic stance shifts from curiosity to interpretation, and the assessment becomes less about the client and more about the therapist’s lens. In Gestalt terms, this begins to resemble confluence at its worst. Confluence interrupts the therapeutic process because the therapist becomes entangled with the client’s worldview rather than maintaining the reflective distance necessary to understand it. The therapist is no longer observing, exploring, and helping the client make meaning. Instead, they are participating in the same undifferentiated field of belief.

This is rarely helpful. Therapy requires enough separation for the therapist to hold perspective, to ask questions the client cannot yet ask themselves, and to support awareness rather than simply joining the narrative. When ideological alignment replaces psychological exploration, the work loses its depth and the client’s complexity is flattened into a story that may not truly belong to them.

Ethical Boundary Confusion

Some therapists may also begin relying on AI for functions that resemble supervision. While this might appear convenient, it raises significant ethical concerns. Clinical supervision is not simply a reflective conversation. It is a structured professional process grounded in accountability, responsibility, and professional oversight.

When therapists turn to AI in place of supervision, several risks emerge. Issues of confidentiality become immediately relevant if identifiable client material is shared with a digital system. Even when details are anonymised, the boundaries around data storage, processing, and ownership remain unclear in many platforms.

There are also deeper concerns about accountability and clinical governance. A human supervisor carries professional responsibility, brings experiential judgement, and can challenge a therapist’s assumptions in ways that are grounded in relational understanding. AI cannot hold responsibility for clinical outcomes, nor can it exercise professional judgement or ethical authority.

If AI begins to function as a substitute for supervision rather than as a limited administrative or educational tool, an essential safeguard within therapeutic practice is quietly removed.

Professional Isolation

Radicalised belief systems can also begin to generate increasing conflict with colleagues, institutions, or regulatory bodies. As beliefs become more rigid, ordinary professional disagreement or scrutiny may start to feel less like part of healthy professional dialogue and more like an attack.

Over time, the therapist may begin to interpret professional challenge as evidence of systemic suppression or institutional hostility. Feedback that would normally prompt reflection can instead be reframed as proof that others are unwilling to tolerate dissenting views. This creates a reinforcing cycle in which opposition strengthens the belief rather than inviting reconsideration.

Within that cycle, the therapist becomes progressively more isolated from the professional community that would normally act as a moderating influence. Dialogue narrows, positions harden, and opportunities for reflective correction diminish. This narrowing of perspective is rarely experienced as radicalisation by the individual involved. More often it is experienced as increasing clarity.


Warning Signs for Therapists

Therapists should monitor for several warning indicators in their own AI usage.

  • Increasing reliance on AI to validate professional or political beliefs
  • Decreasing engagement with peer supervision or consultation
  • Growing certainty that institutions or professional bodies are fundamentally corrupt
  • Using AI to reinforce interpretations of conflict rather than explore alternative explanations
  • Feeling that AI understands one’s thinking better than colleagues or supervisors

When these patterns appear, it is important to reintroduce external dialogue and supervision.


Safeguards for Responsible AI Use

Therapists can adopt several practices to prevent ideological drift.

Maintain Human Supervision

AI must never replace clinical supervision or peer consultation. It is vital you work with someone you trust, and potentially with a history that can be cross-referenced if necessary. 

Human colleagues provide the critical challenge that AI cannot reliably offer, and they offer the human relationship that AI can never mimic.  It only feels like it can.

Use AI for Structured Tasks

But let’s not throw the baby out with the bath water. Just limit AI use to areas such as:

  • drafting documents
  • summarising research
  • generating administrative content to be used outside that system

Avoid using AI for interpreting professional conflicts or analysing colleagues’ behaviour, and continuously be vigilant to thinking that AI is sentient in any way. It is an algorithm.

Actively Seek Dissent

If AI is used for reflective discussion, deliberately ask for alternative explanations and critiques of your own perspective. It is possible to now programme AI to be more challenging and critical of your input. This can also subtly remove some of the reinforcement of AI use. 

For example:

“Provide arguments against this interpretation.”

Maintain Ethical Boundaries

Never input identifiable client information into AI systems. Simply do not do it. The need for anonymisation for any online process is fundamental.  Confidentiality is one of the fundamental ethical commitments of therapeutic work, and once sensitive material is entered into a digital system, control over how that information is processed, stored, or retained becomes uncertain. Whilst it is tempting to scan handwritten notes and have AI instantaneously formulate them logically and coherently as clinical notes, you are ultimately compromising confidentiality. 

AI should be treated strictly as a tool, not as a confidante. It can assist with drafting, summarising, or generating ideas, but it cannot hold responsibility, maintain clinical confidentiality in the way human professionals are bound to, or participate ethically in the therapeutic process. The boundary here needs to remain clear.

Document AI Use Transparently

Where AI contributes to professional writing or teaching materials, disclose this appropriately. We may want everyone to think we have authored everything when in fact AI has done it, but please don’t. There are incredible programmes out there to analyse text. So by all means use AI, just do so carefully. 


A Professional Responsibility

The psychotherapy profession rests on a fundamental capacity to examine one’s own assumptions. It is the most human of human interactions, and should become more necessary as AI grows, than less necessary. Reflection, doubt, and self-scrutiny are central to ethical practice. AI introduces a powerful mirror for human cognition. However, mirrors can distort as easily as they reflect.

Therapists therefore need to remain clear about what these systems actually are. AI systems are statistical language generators, not sources of wisdom. They reproduce patterns in data and language rather than evaluating the truth of a claim. When their responses sound coherent or insightful, it is easy to forget that the system is not reasoning in the way a human supervisor, colleague, or scholar would.

Approached with curiosity and healthy scepticism, AI can be an extraordinarily useful professional tool. It can assist with drafting, organising information, and stimulating ideas. Used uncritically, however, it can quietly reinforce the very cognitive distortions therapists are trained to recognise in others.

Receive new essays

New pieces arrive infrequently and without filler. No newsletters about newsletters. Just writing, when there is something worth saying.