The placebo response in psychiatry is a real biological effect of expectation on the brain — one worth studying rather than subtracting away.
Randomized drug trials are built to remove the placebo response, and in psychiatry that design choice quietly buries something worth studying. Everyone enrolled in an antidepressant trial receives the same surrounding treatment: a diagnostic assessment, repeated contact with a clinician, a pill taken on a fixed schedule, and the expectation that it might help. The trial subtracts the improvement in the placebo arm from the improvement in the drug arm and reports the leftover difference as the effect of the medication. In depression that subtracted portion is large. A review of 75 trials found that about 30% of patients assigned to placebo met response criteria, with individual studies ranging from 12.5% to 51.8%, while the best medication arm averaged about 50%. Meta-analyses that compare effect sizes rather than response rates find that roughly three-quarters of the improvement seen in medication arms also appeared in placebo arms, though that figure is computed differently across analyses and remains contested. These numbers are usually read as evidence that the drugs barely beat an inert pill. They raise a different question, which is what the placebo arm was doing while most of the recovery happened there.
Part of the improvement in a placebo arm comes from natural history and measurement artifacts. Depression fluctuates, so many people enter a trial near a low point and improve on their own. Symptom scores drift back toward the mean, and being measured repeatedly nudges what people notice and report. If that accounted for everything, the placebo arm would be a bookkeeping artifact and subtracting it would be exactly right. Part of the response tracks expectation directly, though, and that part shows up in the body in ways an artifact cannot explain. The clearest evidence comes from experiments that manipulate expectation and then measure the pharmacology it engages.
Early findings concerned pain, where expectation directly engages the body’s own pain-control chemistry. In 1978 Jon Levine, Newton Gordon, and Howard Fields studied patients recovering from dental surgery and gave some of them a placebo they believed was a painkiller, which reduced their pain. When the researchers then gave naloxone, a drug that blocks opioid receptors, the pain relief was reduced in the patients who had responded to the placebo. Because naloxone reverses the effect of morphine and of the body’s own endorphins in the same way, its ability to cancel placebo analgesia points to a concrete mechanism: expecting relief prompts the brain to release its own opioids, the same chemistry a pain drug would recruit. Martina Amanzio and Fabrizio Benedetti later dissected this pharmacologically and confirmed that expectation-driven placebo analgesia runs through the endogenous opioid system.
Neuroimaging later linked placebo analgesia to changes in the brain’s pain-processing and expectation systems. Tor Wager and colleagues found in 2004 that placebo analgesia was associated with reduced activity in pain-sensitive regions such as the thalamus and anterior cingulate cortex, together with increased prefrontal activity while people anticipated pain. Jon-Kar Zubieta and colleagues used PET imaging in 2005 to show that placebo directly activated mu-opioid neurotransmission in the brain, and that this activation tracked lower pain ratings and reduced negative affect. Placebo analgesia can recruit the specific systems that regulate pain.
These responses scale with the strength of the manipulation. Stronger verbal suggestion produces stronger placebo analgesia, and prior experience of a real drug working raises the response to a placebo given afterward. This dose-like structure is why the encounter around a treatment functions as an active ingredient in ordinary care. The confidence a prescriber projects and the ritual of taking something on a schedule both scale the response, so every genuine prescription already carries an expectation component that no one measured or set out to isolate.
The same kind of evidence exists for dopamine. Parkinson’s disease comes from the loss of dopamine-producing neurons, and its motor symptoms respond to drugs that restore dopamine. In 2001 Raúl de la Fuente-Fernández and colleagues gave patients a placebo they were told was their medication and used PET imaging to watch dopamine release; the placebo triggered a substantial release, comparable to a dose of the active drug. A few years later Benedetti’s group recorded from single neurons in the subthalamic nucleus during surgery and found that a placebo changed the firing of individual cells in a way that tracked whether the patient’s rigidity improved. These studies used small samples and stand as proof of mechanism rather than as measures of population effect size, yet they show expectation moving the exact circuit the disease depletes.
The response can also be learned. Pairing an inert substance with an active drug repeatedly lets the inert substance start producing the drug’s effect on its own, the way a bell that has accompanied food comes to trigger salivation. Robert Ader and Nicholas Cohen showed in 1975 that even immune suppression could be conditioned this way in animals, and Marion Goebel and colleagues reproduced it in humans in 2002 by pairing an immunosuppressive drug with a distinctive drink. Benedetti and colleagues separated the two routes directly in 2003: verbal suggestion changed pain and motor performance, while conditioned hormonal responses appeared even when the verbal suggestion pointed the other way. A conditioned response runs on the same associative machinery as any other reflex, which helps explain why the effect is stubborn and reproducible.
With this in hand, the antidepressant debate reads differently. In 2008 Irving Kirsch and colleagues analyzed FDA-submitted trial data for several newer antidepressants, including unpublished trials, and found a mean drug–placebo difference of under two points on the Hamilton depression scale, a clinician-rated measure that runs from roughly 0 to 52 and on which a change of about three points is usually taken as the smallest clinically meaningful one. The difference grew with baseline severity and reached that threshold only at the very severe end. The reason for the gradient is the informative part: the gap widened because the placebo response shrank in the most severely depressed patients, while the drug response changed little. Jay Fournier and colleagues reached the same conclusion from patient-level data in 2010. While these findings are often read as a verdict against antidepressants, the medications do carry a specific effect. A 2018 network meta-analysis by Andrea Cipriani and colleagues, covering 522 trials and more than 116,000 participants, found that all 21 antidepressants studied beat placebo; patients taking them were consistently more likely to respond than those given placebo. Selective publication had inflated the earlier picture, as Erick Turner and colleagues showed in 2008, but even the corrected record leaves the drugs with a specific effect. Both results hold at once: the medications carry a real pharmacological benefit, and a substantial, reproducible placebo response accompanies it, part natural recovery and measurement and part the expectation-linked effect the design is built to remove.
Regulators isolate the molecule to test its specific action, while clinicians and patients rely on the combined effect, since treatment always arrives inside a context. This raises the practical question of how to make the placebo-responsive part of care more reliable, honest, and safe. Ted Kaptchuk and colleagues tested it in irritable bowel syndrome by separating three ingredients: being observed and assessed, a placebo ritual delivered as sham acupuncture, and a warm, confident patient–practitioner relationship. At three weeks, adequate relief was reported by 28% of patients on a waiting list, 44% who received the placebo ritual with limited interaction, and 62% who received it with the added relationship. The structure of care could be varied like an experimental factor, and the variation changed outcomes.
Using this on purpose raises an obvious ethical objection, because amplifying expectation seems to require deceiving the patient about what they are taking. Open-label trials weaken that objection. In 2010 Kaptchuk and colleagues gave IBS patients a bottle of pills labelled as placebo, told them plainly the pills contained no active ingredient, and explained that placebos can still produce real effects through mind–body pathways. The open-label group did better than patients given no pills, with about 59% reporting adequate relief against 35%, so the effect survived being described honestly. The result is modest and specific to a condition with a strong symptom-report component, and it removes the sharpest reason to leave the mechanism unstudied.
This reframing changes how to judge therapies that work mainly through the interaction itself, including AI chatbots used for mental health. A placebo response draws on attention, a plausible account of what is wrong, and the expectation that the encounter will help, and a conversation with a responsive chatbot supplies all of these. When someone reports feeling better after talking to such a system, that relief can be genuine and can run through the same expectation pathways as any credible treatment. Expectation is what moved opioids and dopamine in the studies above, and a credible conversation is one way to create it. A chatbot is therefore not inert simply because it is only talking. Short-term relief also does not show that the tool works, because the informative comparison is a convincing sham interaction, not an absence of contact. This is where most published trials of mental-health apps fall short. The small Woebot study that reported a two-week symptom reduction used an information-only control, and many other app trials compare against a waitlist or no support. A waitlist controls for the passage of time but not for the expectation and ritual the app itself supplies, so the reported benefit blends any specific effect with a placebo effect the design cannot separate. These context effects could also do harm, and that possibility deserves measurement rather than assumption.
The limits matter, because the placebo response is easy to over-sell. It is best documented in pain and other subjective symptoms, with proof-of-mechanism findings in Parkinson’s and in conditioning studies; a broad review by Asbjørn Hróbjartsson and Peter Gøtzsche found little evidence for powerful placebo effects on objective outcomes. Its size varies widely between people, conditions, and even trials of the same drug, which is part of why Walsh and colleagues found the placebo response in depression trials to be substantial and rising over the decades. No one can yet dose or predict it, and using it on purpose still raises a real question, because managing what a patient expects shades toward managing what they believe. Studying placebo as a mechanism calls for stronger controls and higher standards of evidence.
The evidence establishes that the placebo arm of a psychiatric trial contains a real treatment: one that recruits the brain’s own opioid and dopamine systems, that can be conditioned like a reflex, and that survives being described honestly to the patient. Subtracting it away keeps trials clean, and it also discards the data that would show how expectation changes the brain.