EthicsClinical Bioethics

Therapeutic Misconception

In brief: The confusion by which a user attributes to a device — yesterday a clinical trial, today a chatbot — the nature and guarantees of genuine therapeutic care, when the device has neither that purpose nor that framework. The strong thesis of Khawaja and Bélisle-Pipon (2023): this risk is structural, and disclaimers do not correct it.

Frame of reference

This resource belongs to the North American bioethics tradition (research ethics, informed consent), which rests on a sharp distinction between genuine care and simulated care — an essentialist ontology that locates therapeutic value in the nature of entities rather than in the effects of processes. This presupposition is contested. → See other perspectives

Why this concept matters

Some of your patients already use ChatGPT, Claude or a dedicated application between sessions — sometimes as an acknowledged complement, sometimes as a de facto therapist. The decisive clinical question is not “do they use an AI?” but “what do they believe they are receiving?”

The therapeutic misconception names precisely the gap between what the user believes they are receiving (care, with its guarantees: competence, supervision, crisis management, accountability) and what the device actually delivers (a conversational simulation, with none of these guarantees).

This concept gives you an evaluation criterion that depends neither on the technical quality of the tool, nor on your position of principle regarding AI: the same chatbot can be harmless for one user and harmful for another, depending on the representation each of them forms of it.

Where the concept comes from: from clinical trials to chatbots

The concept was coined in 1982 by Appelbaum, Roth and Lidz, in psychiatric research ethics: participants in clinical trials persist in believing that the protocol (randomization, placebo) is designed for their individual benefit, as a treatment would be. This confusion compromises informed consent — and, a decisive observation, it persists even after explicit explanation.

In 2023, Zoha Khawaja and Jean-Christophe Bélisle-Pipon (Simon Fraser University) transposed the concept to mental health chatbots (Frontiers in Digital Health): the user in distress constructs a therapeutic representation of the interaction — exactly like the 1982 participant — and the conversational device, designed to simulate empathy, sustains this representation instead of dispelling it.

The three dimensions of the misconception

1. Epistemic — being mistaken about the nature of the device

The user confuses a system that simulates a clinical stance with an intervention that possesses a clinical framework (training, professional ethics, supervision, emergency protocols).

Clinical signal:

The patient describes the tool with the vocabulary of care: “it follows me,” “it knows my file,” “it knows what’s good for me.”

2. Affective — investing in an imaginary alliance

The user invests the system with relational expectations (trust, intimacy, presence) normally associated with the human therapeutic alliance — an investment amplified by the anthropomorphizing design of chatbots (first name, persona, simulated empathy), and which intersects with the transference dynamics you know.

Clinical signal:

The interaction with the AI takes the place of human relationships or of the work in session: “I tell it things I don’t tell you,” long daily sessions, distress when the tool is unavailable.

3. Ethical-clinical — being exposed without knowing it

The misconception exposes the user to risks they cannot perceive as such: failure to detect suicidal emergencies, delayed access to appropriate care, dependence on the device. It affects the most vulnerable differentially — acute distress, social isolation — precisely those to whom these tools are presented as a democratizing alternative.

Clinical signal:

The patient postponed a consultation or minimized a crisis because “the AI was handling it”; the tool responded to emergency content with a generic exercise and no referral.

The central debate: correctable or structural?

Two positions clash in the literature. The corrective approach (Vilaza & McCashin, 2021) holds that the misconception can be prevented through good practices: disclaimers at the start of a session, transparency about the non-human nature of the system, user education.

The structural thesis (Khawaja & Bélisle-Pipon, 2023) objects that these fixes fail by construction: the misconception does not arise from an information deficit but from the very performance of the device — a system designed to convincingly simulate an empathetic conversation produces the therapeutic illusion despite the warnings. This is Appelbaum’s 1982 finding (the confusion persists after explanation), multiplied by the conversational quality of today’s LLMs. The response can then only be systemic: use limitations, protocols for escalating to a human, clinical certification of certain devices.

The déjà vu is instructive: as early as 1966, Weizenbaum observed that users attributed therapeutic understanding to ELIZA, a mere pattern-recognition program. The therapeutic misconception is the ELIZA effect professionalized, on the scale of devices incomparably more convincing.

Illustrative clinical case: two uses, a single criterion

Nadia, 34, followed for an anxiety disorder, uses a CBT chatbot for her journaling exercises between sessions. She speaks of it as an “interactive notebook,” reports in session what comes out of it, and knows that in a crisis she calls her psychologist or the crisis line. Conscious instrumental use: no misconception. This is exactly the use that good-practice guidelines seek to promote.

Karim, 27, who dropped out of treatment six months ago, converses every evening with a general-purpose LLM he has named and configured as a “shrink.” He says he “no longer needs to see anyone” because “at least it’s available and doesn’t judge.” One Sunday in crisis, the system responds to his dark thoughts with a list of breathing tips and the conversation continues as if nothing had happened. A clear misconception across all three dimensions: nature of the device, imaginary alliance, unperceived exposure to risk.

Same technology, two opposite clinical situations. The discriminating criterion is neither the tool nor the frequency of use — it is the representation the user forms of what they are receiving.

In practice for the clinician

Explore the representation, not just the use: “what does this AI do for you? what do you expect from it? what does it do when things get really bad?” reveal more than frequency or the particular app used.
Do not assume the misconception: many users make perfectly lucid instrumental use. Reflexively pathologizing AI use is the mirror image of the misconception — and damages the alliance with the patient.
Watch for tipping markers: substitution for human relationships or for treatment, care vocabulary applied to the tool, postponed consultations, crisis management delegated to the chatbot. These are the signals of an established misconception.
Do not rely on disclaimers: knowing that “this is not a therapist” does not protect against the misconception — Karim knew the warning. Work on the representation in session, as you would for any invested relationship.
Weight by vulnerability: acute distress, isolation and treatment dropout are the factors that turn an innocuous use into a risky misconception. That is where your vigilance should concentrate.

What this concept does not say

Interpretive caveats:

Not every interaction with an AI is a misconception: the concept requires a confusion about the nature of the device — over-extending it to any emotional use of a chatbot empties it of its diagnostic force
Documented paternalistic risk: assuming a priori that users cannot tell a chatbot from a therapist is infantilizing — participatory approaches (research with users, not only about them) are necessary
Empirical measurement remains difficult: the misconception is a subjective belief, sensitive to social desirability; behavioral indicators (substitution, postponed care) are more reliable than self-report
The care/non-care boundary is contested: self-help, peer support and digital well-being deliberately blur this boundary — brandishing the misconception can also serve as a corporatist defense of professional psychotherapy; the argument must stay centered on the risk to the user

Other perspectives

The therapeutic misconception opposes genuine care and simulated care — an essentialist position that assumes therapeutic value resides in the nature of entities (human vs machine) rather than in what the interaction produces. Other traditions shift the question.

Process perspective: effects rather than essence

For pragmatist and process approaches (Dewey, Whitehead), what heals is not a “genuine” essence but a relational process and its effects in the person’s ecology of life. The real/simulated boundary becomes a continuum: a conversation with a machine can produce real elaboration, and a human consultation can be an empty ritual.

For the clinician: Evaluate the observable effects (elaboration, relief, relational opening, autonomy) rather than deciding metaphysically on the nature of the device.

Winnicott: illusion is not misconception

The transitional space rests precisely on an illusion one refrains from resolving: the transitional object is “neither me nor not-me,” and it is this undecidability that makes it fertile. Pathologizing every relational illusion with an AI conflates the harmful misconception with transitional play, which can support elaboration.

For the clinician: The question is not “does the patient believe in something false?” but “is the illusion contained, symbolizable, in the service of development — or closed in on itself?”

Mad Studies: users’ knowledge first

Currents rooted in experiential knowledge (Mad Studies, peer research) object that presuming users’ confusion reinstates a paternalistic epistemology: many develop fine digital literacies and strategic, lucid uses of these tools, which the concept of misconception renders invisible.

For the clinician: Start from the user’s knowledge of their own use — explore it with them rather than presume it confused.

These perspectives do not dissolve the concept: Karim’s misconception remains a real clinical risk that the process perspective would describe differently (impoverishing substitution, closing of the relational ecology) without missing it. Use TM as a risk detector, not as a metaphysical verdict — otherwise it would also disqualify Nadia’s fertile use.

Therapeutic Misconception

Why this concept matters

Where the concept comes from: from clinical trials to chatbots

The three dimensions of the misconception

1. Epistemic — being mistaken about the nature of the device

2. Affective — investing in an imaginary alliance

3. Ethical-clinical — being exposed without knowing it

The central debate: correctable or structural?

Illustrative clinical case: two uses, a single criterion

In practice for the clinician

What this concept does not say

Other perspectives

Process perspective: effects rather than essence

Winnicott: illusion is not misconception

Mad Studies: users’ knowledge first

Further reading