Orhan OnderImagine describing your pain to a doctor and watching her eyes move from you to a screen. The screen says you are low risk. Your body says otherwise. Between the two, someone has to be believed.
Hospitals around the world are racing to put algorithms into the clinic, from symptom checkers that triage patients to AI tools that support diagnosis. The pitch is efficiency, coordination and better accuracy.
But these systems do something the brochures rarely mention. They quietly decide whose knowledge is credible, which symptoms matter and how a patient’s words should be interpreted.
I call this dynamic algorithmic gaslighting.
Gaslighting without a gaslighter
“Medical gaslighting” describes patients whose symptoms are dismissed or reinterpreted by clinicians, often along lines of gender, race or chronic illness.
Critics say the word is misused because gaslighting implies deliberate deception, and most dismissive doctors are not scheming. They are acting out of implicit bias, institutional routine and ingrained hierarchies of expertise.
But the harm is not defined by motive. It is defined by what happens to the patient, who is left doubting their own perceptions and cut off from the words they need to make sense of their experience.
Algorithms make this point impossible to avoid. A piece of software has no intentions. It cannot plot or deceive.
Yet it can override patient testimony with the authority of apparent neutrality. It can squeeze a lived narrative into rigid categories. And it can leave a patient feeling that their own experience must be unreliable.
That is gaslighting without a gaslighter.
The first is data violence. Medical datasets often skew towards white, male and socioeconomically advantaged populations. Patients outside that mould become statistically invisible.
Chest pain, for example, often presents differently in women than in men, yet cardiovascular data disproportionately encode male profiles.
Pulse oximeters, calibrated mostly on lighter-skinned patients, can overestimate blood oxygen in Black patients and hide dangerously low levels. If triage tools treat those readings as ground truth, the error travels downstream.
The second is opacity. You can argue with a doctor’s reasoning. You cannot argue with a score whose derivation is closed to you and, often, to your doctor too.
Consider the Epic Sepsis Model, a proprietary alert system used in hundreds of US hospitals. An external study found it performed far worse than its developer claimed, catching only about a third of sepsis cases while producing frequent false alarms. Neither clinicians nor patients could see the logic behind any single alert.
The third is automated authority. When the algorithm “says otherwise”, patients feel pressure to doubt their own account, and clinicians hesitate to override a supposedly objective recommendation.
That authority can outrun the evidence. Internal documents later showed that IBM’s Watson for Oncology, marketed worldwide as an authoritative decision-support tool, produced unsafe and incorrect treatment recommendations and had been trained substantially on hypothetical cases.
The fourth is the feedback loop. Patients and clinicians learn which descriptions register as legible “cases” and adjust how they speak. Those adjusted descriptions then become new data, entrenching the very categories that failed to capture the patient in the first place.
Over time, people stop describing their symptoms as they experience them. They describe them as the machine can hear them.
Why this is not a glitch
It is tempting to treat all this as a bug that better engineering will fix. The paper argues otherwise.
Algorithms do not select themselves. Hospitals and health systems choose to adopt them, and they often prefer standardised, case-based information over messy human narrative, for reasons of efficiency, liability and scale.
This is an old administrative instinct. Institutions have always favoured what they can read, count and audit over what they cannot. What is illegible to the system is, in a real sense, ungovernable by it.
So the point at which intent enters the story is upstream. No one needs to want to disbelieve a particular patient. But choosing to replace patient testimony with standardised data is a deliberate act of design, and someone is accountable for it.
Because algorithmic outputs are infrastructural, the damage is also not confined to one bad appointment. It repeats across cases and accumulates for patients who sit at the intersection of several marginalisations, such as a Black woman with a disability whose condition and identity are both thinly represented in the data.
Whole communities that are repeatedly misrecognised may come to distrust the health system itself.
There is a global dimension too. In a recent Turkish qualitative study, different language models gave strikingly different ethical guidance on the same clinical cases, differing in cultural sensitivity and interpretive depth even when they invoked the same principles.
Those frames of reference that get encoded are not neutral, particularly for patients far from where these systems are built.
What should change
The answer is not to ban AI from medicine. It is to stop treating patients as data sources and start treating them as knowers.
That means, first, giving patients the right to contest an algorithmic output, and not merely to request a second opinion.
They should be able to question the categories, proxies and value judgements behind it. They should also get explanations that connect to their own experience, rather than a technical readout of feature weights.
Second, epistemic justice has to be built into the systems themselves. That includes independent audits of training data and outputs, with ethicists and patient advocates at the table.
It includes involving the people most likely to be misrecognised in design, not after the architecture is fixed. And it includes asking, at procurement, whether a tool invites patient narrative or reduces it to checkboxes.
Participatory design is not a cure-all, and the paper is candid about this. Proprietary vendors resist outside scrutiny, privatised systems offer no single funder, and emergency care leaves little time for co-design.
It works best in chronic-care settings and where patient organisations already exist. Where it cannot be done fully, lighter safeguards are better than none.
Third, responsibility has to be shared. Developers should document limits and design for contestability.
Clinicians should weigh algorithmic output against what the patient tells them. Institutions should create bodies that monitor deployment and settle disputes. Regulators should write epistemic justice into the standard of care.
A caveat is due. My argument is conceptual and ethical. It does not claim to measure how often these harms occur, and empirical work will be essential. But a problem you cannot name is a problem you cannot study.
The question is no longer whether healthcare will be digitally mediated. It will be.
The question is whether that mediation will widen or narrow the space in which patients are heard.
A health system that cannot believe the person in front of it is no longer a health system. Medicine runs on trust, and no model, however accurate, can replace the patient’s voice.
