Algorithmic gaslighting: when the number on the screen outranks how you feel

When my watch started telling me how I slept

In July 2025, I took off my smartwatch for good.

For four years, the first thing I did every morning was check my sleep: duration, cycles, minutes of REM, and the score out of 100 the watch gave my night. I had almost four years of unbroken data. One morning I noticed something odd. The number on the screen had become more important than how I felt about my night. If someone asked how I had slept, I no longer said “I slept well, but I woke up tired” or “it was a rough night.” I said I’d had three REM cycles of about forty minutes each, and I quoted the score. My own testimony had been demoted, and I had switched my vocabulary without noticing.

So I freed my wrist and started paying attention to my own experience again.

I was already familiar with the literature on medical gaslighting and Miranda Fricker’s work on epistemic injustice. As I thought about my sleep, the two came together: what if this is gaslighting, but one where the thing that undermines your self-trust is an algorithm? That question became this paper.

What the paper argues

Algorithmic systems are now woven into healthcare, from symptom checkers and triage tools to AI decision support. They are usually evaluated on accuracy, efficiency, bias, and transparency. I argue they also do something less visible: they shape whose knowledge counts as credible, which symptoms matter, and how a patient’s testimony is interpreted.

I introduce algorithmic gaslighting to name a patterned configuration in which:

1. epistemic authority is redistributed within a doctor–patient–technology triad;
2. patient testimony is systematically subordinated to algorithmic outputs;
3. this subordination is stabilized through routine reliance and infrastructural feedback; and
4. patients come to doubt the reliability of their own experiential knowledge.

The concept does not require anyone to intend harm. An algorithm has no motives, and the clinician using it may be well-meaning. The harm lies in the relational effect: patients are left epistemically disoriented and estranged from their own experience. I also argue that “no malicious intent” doesn’t mean “no intent anywhere.” Healthcare organizations deliberately choose standardized, case-based data over narrative accounts, and that upstream choice is where responsibility properly sits.

Algorithmic mediation adds three structural differences to what we already know about interpersonal medical gaslighting:

• It turns a dyad into a triad, adding a third authority that is answerable to neither party.
• Opacity removes even the “reasoning-with” a dismissive clinician still allows. You can argue with a person, but you cannot argue with a score no one can explain.
• Algorithmic outputs are infrastructural, not episodic, so credibility displacement is reproduced across cases rather than confined to one encounter.

Four mechanisms

• Data violence: Training data and devices underrepresent certain bodies, so their symptoms become illegible to the system. The pulse oximeter overestimating oxygen saturation in Black patients is the clearest example.
• Opacity: Reasoning that can’t be interrogated can’t be contested. The Epic Sepsis Model’s proprietary logic is the example here.
• Automated authority: Outputs are treated as more credible than human judgment, even when the evidence doesn’t support that. IBM Watson for Oncology shows how authority can outrun reliability.
• Feedback loops: Patients and clinicians adapt how they describe symptoms to fit system categories, and those adapted reports then feed future models. This resembles Ian Hacking’s “looping effect,” except the loop runs through retraining rather than conscious reclassification.

Together these produce harms that go beyond misdiagnosis: erosion of self-trust, collapse of institutional trust, epistemic objectification, and a recalibration of patients’ sense of themselves as knowers. These effects are magnified for patients at the intersection of several marginalized identities.

What to do about it

Fairness and transparency fixes are necessary but not sufficient. The framework I propose works on three levels:

• Individual agency: a right to contestation (to interrogate categories, proxies, and value judgments, not just appeal a decision) and a relationally meaningful right to explanation, one that connects to patients’ embodied experience and supports clinicians’ judgment.
• Structural embedding: critical audits that include ethicists, patient advocates, and people with lived experience; participatory design; and “epistemic justice checks” in procurement. I am also candid about where participatory design is hard to achieve, such as proprietary vendor tools and acute care, and suggest lighter-weight alternatives there.
• Shared responsibility: developers, clinicians, institutions, and regulators each hold a distinct duty, coordinated rather than diffused.

The paper is deliberately conceptual. It offers a vocabulary, not prevalence data. Empirical work is the natural next step.

The road to publication

I drafted the first version and submitted it to a journal where, after revisions, it was rejected. A later submission led to a recommended transfer, and the paper finally found its home at Philosophy, Ethics, and Humanities in Medicine. Along the way I presented it at the 11th Bioethics Symposium organized by the Turkish Bioethics Association. The audience’s engagement and comments were so encouraging that I kept revising and resubmitting after each rejection. The paper is stronger for every round.

As for my wrist, it has been free ever since. I still sleep, and I’m fairly sure I do it well. I just trust my own account of it.