By Orhan Onde

Make you imagine say you dey describe your pain to doctor and you dey watch as her eyes shift from you go screen. The screen talk say your risk low. Your body dey tell another story. For between the two, one person gats to believe.

Hospitals for different part of world dey rush to put algorithms inside clinic — from symptom checkers wey dey triage patients to AI tools wey dey help diagnosis. The koko of the matter na efficiency, coordination and better accuracy.

But these systems dey do one thing wey the brochures no dey talk well well. Dem quietly dey decide which kind knowledge get credence, which symptoms matter and how person words suppose dey interpreted.

I dey call this dynamic algorithmic gaslighting.

Gaslighting without a gaslighter

‘‘Medical gaslighting’’ na how people dey call situation wey clinicians dey dismiss or reinterpret patient symptoms, many times based on gender, race or because person get chronic sickness.

Some critics talk say people dey misuse the word because gaslighting dey carry idea of deliberate deception, and most doctors wey dey dismiss patients no dey do am as part of plan. Dem dey act from implicit bias, the normal routines for the institution and the hierarchy wey don solidify for medical world.

But the harm no depend on motive. The harm na wetin happen to the patient — the person go start to doubt hin own perception and dem go cut the person off from the words wey dem need to understand wetin dem dey experience.

Algorithms make this point impossible to avoid. Software no get intentions. E no fit plan or deceive.

But e fit override patient testimony with the authority wey look neutral. E fit pack live human story into tight categories. And e fit make patient feel sey their own experience no steady.

That is gaslighting without a gaslighter.

The first na data violence. Medical datasets many times dey skew towards white, male and socioeconomically advantaged people. Patients wey no follow that mould fit become statistically invisible.

For example, chest pain many times dey show differently for women than men, but cardiovascular data plenty encode male profiles. Pulse oximeters wey dem calibrate mostly on lighter skin fit overestimate blood oxygen for Black patients and hide dangerously low levels. If triage tools accept those readings as ground truth, the mistake go continue waka downstream.

The second na opacity. You fit argue with doctor reason. But you no fit argue with score wey derivation dey closed to you and many times even closed to your doctor.

Take the Epic Sepsis Model for example — na proprietary alert system wey dey hundreds of US hospitals. One external study find say e perform far worse than the developer claim: e catch only about one third of sepsis cases and e dey give plenty false alarms. Neither clinicians nor patients fit see wetin make any single alert ring.

The third na automated authority. When algorithm "talk say otherwise", patients dey feel pressure to doubt their own story, and clinicians dey fear to override recommendation wey suppose be objective.

That authority fit pass the proof. Internal documents later show say IBM’s Watson for Oncology, wey dem market for many countries as decision-support tool, produce unsafe and wrong treatment recommendations and e train plenty on hypothetical cases.

The fourth na the feedback loop. Patients and clinicians dey learn which descriptions the machine go recognise as proper "cases" and dem go change how dem dey speak. The new descriptions then turn new data, and that one go strengthen the same categories wey first fail to capture the patient.

Over time, people go stop to describe their symptoms as dem really feel am. Dem go describe am as the machine fit hear am.

Why this is not a glitch

E dey tempting to treat all this like bug wey better engineering go fit fix. But the paper talk say otherwise.

Algorithms no select themselves. Hospitals and health systems na dem dey choose to adopt dem, and many times dem dey prefer standardised, case-based information pass messy human story, for reasons of efficiency, liability and scale.

Dis na old administrative instinct. Institutions don always prefer wetin dem fit read, count and audit pass wetin dem no fit. Wetin the system no fit read, for real sense, na wetin the system no fit govern.

So the point wey intent enter story dey upstream. No need make person wan disbelieve particular patient. But to choose to change patient testimony to standardised data na deliberate design decision, and person gots to answer for am.

Because algorithmic outputs turn part of infrastructure, the harm no dey end for one bad appointment. E repeat for many cases and e accumulate for patients wey dey for intersection of different marginalisations — for example, Black woman wey get disability wey both her condition and identity no well represent for data.

Whole communities wey dem dey misrecognise again and again fit begin distrust the health system itself.

There dey global dimension too. For one recent Turkish qualitative study, different language models give very different ethical advice for the same clinical cases — dem differ for cultural sensitivity and depth of interpretation even when dem talk the same principles.

Those frames of reference wey dem encode no be neutral, especially for patients wey far from where these systems dem build.

What should change

The answer no be to ban AI for medicine. The answer na make we stop to treat patients as only data sources and begin treat them as people wey sabi their body.

First, patients suppose get right to contest algorithmic output — no be only to ask for second opinion. Dem suppose fit question the categories, proxies and value judgements wey the tool use. Dem suppose also get explanations wey connect to their own experience, not only technical printout of feature weights.

Second, epistemic justice must dey build inside the systems themselves. That one include independent audits of training data and outputs, with ethicists and patient advocates for the table. E include to involve the people wey most likely to be misrecognised during design, not after the architecture don fix. E include to ask, when buying tools, whether the tool invite patient story or e reduce am to checkboxes.

Participatory design no be cure-all, and the paper frank about this. Proprietary vendors dey resist outside scrutiny, privatised systems no get single funder, and emergency care no get time for co-design. E dey work best for chronic-care settings and where patient organisations don already exist. Where e no fit full, lighter safeguards better pass nothing.

Third, responsibility suppose dey shared. Developers suppose document limits and design so outputs fit contest. Clinicians suppose weigh algorithmic output against wetin patient tell them. Institutions suppose create bodies to monitor deployment and settle dispute. Regulators suppose write epistemic justice into standard of care.

One caveat dey necessary. My argument be conceptual and ethical. I no dey claim to measure how often these harms happen; empirical work go need. But problem wey you no fit name na problem you no fit study.

The question no be again whether healthcare go dey digitally mediated — e go. The question be whether that mediation go widen or narrow the space wey patients fit dey heard.

Health system wey no fit believe the person wey dey front of am no longer be real health system. Medicine dey run on trust, and no model, no matter how accurate, fit replace the patient voice.