EpistemologyEvaluation biasAI & psychotherapy

Epistemic Double Standard

In brief: in the debate on AI in psychotherapy, LLM-based tools are systematically evaluated against standards that human reference practices do not themselves satisfy. We demand from AI proofs of harmlessness that human psychotherapy has never provided. We count the risks of AI without comparing them to the risks of the status quo. This bias is not an oversight — it is a deeply ingrained cognitive reflex.

Frame of reference

This concept belongs to the epistemology of cognitive biases (Kahneman) and to a comparative, consequentialist rationality: an honest evaluation presupposes that human practices and AI tools are commensurable — comparable against common criteria. That presupposition of commensurability is precisely what other traditions discuss. → See other perspectives

Why this concept matters

When you read a study concluding that “AI is dangerous in mental health”, or when a colleague claims that “chatbots cannot replace a therapist”, a simple question lets you detect a possible double standard:

“Would the criticism levelled at AI also hold for the corresponding human practice?”

If the answer is yes — and we do not raise the criticism against the human — that is a double standard. This does not mean the criticism is wrong. It means it is incomplete: it is missing the comparator that would give it its proper measure.

The four-pan balance

In psychotherapy, a well-informed clinical decision rests on a complete decisional balance — not only on risks. Four pans, not two.

Change (integrate AI)

+Benefits of change

−Risks of change

Status quo (change nothing)

+Benefits of status quo

−Costs of status quo (loss of opportunity)

The epistemic double standard manifests when one fills only one pan: the risks of change. The other three — the benefits of change, the benefits of the status quo, and especially the cost of the status quo (the loss of opportunity) — remain empty. This is exactly the structure of most alarmist articles on AI in mental health: a meticulous tally of potential risks, without a word on what the absence of the tool costs.

The death rate from medical error is counted in millions per year. The one tied to AI in mental health can be counted on the fingers of one hand. This does not make AI harmless — but it makes the comparison mandatory.

Five common double standards in the psy-AI debate

”AI is too sycophantic”

LLM sycophancy is a real and documented risk. But when did you last read a study on the sycophancy of human therapists? Asch’s experiments (1951) show that 75% of humans yield to social conformity on a trivial perceptual question. Therapists are not immune: social desirability, allegiance to one’s theoretical framework, confirmation bias in supervision — all forms of human sycophancy that are rarely named.

The problem is not that AI is sycophantic. It is that we blame AI for it without measuring it in humans.

”AI can trigger suicide attempts”

Isolated cases of suicides linked to chatbot use have been reported. Each is a tragedy. But the comparator is missing: how many suicide attempts occur despite a human therapeutic follow-up? How many patients do not commit suicide because a chatbot was available at 3 AM when the therapist was not? The testimonial of Gaëlle Charlot documents one such case.

Pointing to the risk without comparing it to the cost of absence is doing a benefit-risk analysis with one pan.

”LLMs have no proof of clinical effectiveness”

This is factual: most studies on LLMs in mental health are at the T1 (technical test) level, not T3 (clinical effectiveness). But how many common human practices in psychotherapy have a robust T3 level? Entire swaths of practice — supervision, reformulation, intuitive frame adjustment — rest on no T3-level evidence. To demand from an LLM proofs that the human therapist does not provide is a double standard.

The evidence deficit of LLMs is real. But it is rarely compared to the (just as real) evidence deficit of the human reference practice.

”AI reinforces cognitive biases”

Likely. But it is also what a poorly supervised human therapist does, a well-meaning friend, or a peer group sharing the same priors. Confirmation bias is a universal human phenomenon — AI may amplify it, but it does not invent it. The useful question is not “does AI reinforce biases?” (trivial answer: yes) but “to what extent, compared to what, and with what safeguards?“

”There must be a human in the loop”

A reasonable principle of caution we share. But one that has the status of an article of faith when it is asserted without empirical confrontation. The data that would support “a human in the loop systematically produces better clinical outcomes than an LLM alone” do not yet exist — precisely because the study is not done. The principle could be true. But to declare it true without proof, while demanding proofs from AI, is the double standard itself.

Why the double standard is so easy to commit

It is not a trivial bias nor an attention lapse. It is fed by several well-documented cognitive mechanisms.

•
Status quo bias: we perceive the risks of change as more salient than the risks (which are equivalent or higher) of the absence of change. The negative of the status quo is invisible because it is habitual.
•
Familiarity: we grant implicit trust to familiar systems (the human therapist) and implicit scepticism to unfamiliar systems (the LLM). It is the same mechanism that made the printing press feel threatening in the 15^th century.
•
Professional identity: questioning the absolute superiority of the human therapist touches the very identity of the practitioner. The double standard protects this identity by keeping the comparison out of view.
•
Structure of scientific publication: an article demonstrating a risk of AI is publishable. An article demonstrating that the same risk exists in humans is trivial. The structure of academic incentives favours counting AI risks and ignoring the comparator.

What this concept is not

It is not an argument in favour of AI

Pointing to a double standard does not mean AI is harmless or preferable. It means the evaluation must be complete and comparative — the four pans filled — to be intellectually honest.

It is not a rejection of caution

The precautionary principle is legitimate. What is not legitimate is applying it selectively to AI while suspending it for human practices. Caution is coherent or it is nothing.

It is not an attack on therapists

Human psychotherapy is precious and irreplaceable in many contexts. Saying that the comparison with AI must be honest is not saying that the human is failing — it is saying that the debate is failing when it avoids the comparison.

What it changes for your practice

When you read a study listing risks of AI in psychotherapy, check whether the comparator is present. “AI can do X harm” without “compared to what?” is an incomplete argument.
When you debate among peers, try to fill the four pans before concluding. What are the risks of AI? Its benefits? What are the benefits of the status quo? And above all: what is the cost of the status quo?
With your GAD patients, the double standard will be familiar to you: it is exactly the cognitive structure of anxious avoidance. The risks of change are hyper-salient; the costs of immobility are invisible. The cognitive restructuring that works for the patient also works for the professional debate.

Other perspectives

The concept demands that AI and human practice be compared against common criteria. But is every asymmetry of standards a bias? Several traditions argue it is not.

Incommensurability: the relationship is not a service

For phenomenological approaches and the ethics of care, placing therapist and chatbot on the pans of a single scale already assumes that care is a service with measurable outcomes. Yet part of what heals — the presence of a responsible, engaged other, vulnerable in their own right — cannot be converted into a comparable metric.

For the clinician: Some refusals to compare express less a double standard than an arguable conviction: the two situations are not the same kind of thing. The honest discussion then bears on exactly that point.

Precautionary principle: asymmetry can be rational

For the philosophy of risk (Hans Jonas), treating the new more severely than the existing is not always a bias: when the scale of deployment and the reversibility differ radically, the asymmetry of standards is a deliberate precaution. A failing therapist harms their patients; a failing model deployed at scale harms millions of people at once.

For the clinician: The right question is not “is there an asymmetry?” but “is the asymmetry of standards proportionate to the asymmetry of scale?”

Social epistemology: trust rests on institutions

The trust granted to the human therapist is not merely a familiarity bias: it rests on institutions — training, professional ethics, supervision, legal liability — that make human practice correctable and accountable. The gap in treatment may reflect that real institutional gap, not just a conservative reflex.

For the clinician: Demanding institutional equivalents from AI (audit, recourse, accountability) is not a double standard — it is a demand for institutional symmetry.

These perspectives do not excuse us from filling all four pans of the scale — they remind us that comparison itself has conditions: agreeing on what is comparable, at what scale, and with what institutional guarantees.

Related concepts on this site

Sycophancy of LLMs

Sycophancy is a real LLM risk. But the sycophancy reproach itself is subject to the double standard if it is not compared to human conformity (Asch 1951, social desirability in therapy).

WEIRD Sample

Studies that denounce LLM biases often themselves rely on WEIRD samples. A double standard squared: blaming AI for a training bias from studies that suffer from the same sampling bias.