Ethics and Equity in Community Health AI

There is a lot of loose talk about AI in healthcare right now, and most of it is written for hospitals in wealthy cities. We work somewhere different: a border district in West Bengal where many of the people we screen have never had a routine blood test in their lives. When you bring algorithm-assisted screening to people with few alternatives, the ethics are not an add-on. They decide whether the work helps or quietly causes harm.

Here is how we actually think about it, in plain terms.

The people with the least power deserve the most caution

A patient who can choose between five clinics can walk away from a bad one. A patient who can reach exactly one screening visit, once, after a long trip, cannot. That imbalance means the burden of getting it right sits entirely on us, not on the patient's ability to shop around. We treat every shortcut as something the patient will pay for, not us.

Consent has to be real, not a signature

Informed consent is easy to fake and hard to do well. Someone can sign a form they did not understand, in a language that was not theirs, under social pressure to go along. We try to do the harder version: the operator explains what the screen does, what it does not do, that it is a screen and not a diagnosis, and that the person can decline any part of it and still be treated with respect. Consent that a person could not have refused is not consent.

Bias is a measurement problem, so we measure it

The most discussed failure of health AI is that it works well on the populations it was tested on and badly on everyone else. We have written about this with pulse oximetry and skin tone. Our response is not to promise we are unbiased. It is to test for bias and publish what we find, broken down by skin tone rather than hidden inside an average. If a tool reads worse for darker skin, the only honest options are to fix it or to say so. Reporting one comfortable average number across everyone is how the problem stays invisible.

Over-claiming is its own kind of harm

It is tempting, when you are trying to do good and raise support, to let a screening tool sound like more than it is. We hold a hard line on this. The screen flags people who should see a clinician. It does not diagnose. A follow-up monitoring band tracks trends for people already identified; it is not a diagnostic device and we do not let it borrow the screen's validation. Tuberculosis, for instance, is something we can raise a flag about for referral, never something we claim to detect. Every time a tool is described as more capable than its evidence supports, a patient somewhere is being given false reassurance or false alarm. Both cost something.

Data belongs to the community, not to us

We hold health data about real people in a real place. We keep it anonymised and aggregated, we do not sell it, and we treat the map of local health needs it produces as a resource for the community's own planning, under India's data protection rules. The test we apply is simple: would the person whose data this is be comfortable with what we just did? If we are not sure, we do not do it.

Equity is the whole point, not a feature

It would be possible to build a screening business that goes where paying customers already are. That is the opposite of why we exist. The point is to reach the people the system has been skipping, in the tones device-makers left out of testing, in the villages too far from a lab. If a decision makes the tool more profitable but less reachable for those people, it is the wrong decision for us. That is not a marketing position. It is the line we use to settle hard choices.

None of this makes us immune to getting things wrong. It just means that when we do, the mistake shows up against a standard we set out loud, where someone can hold us to it. That is the most honest version of this work we know how to do.

FAQ

Does SamaHealth use AI to diagnose patients? No. Algorithm-assisted tools help flag who should see a clinician. Diagnosis is made by qualified clinicians, not by the screening tool.

How do you handle bias in screening accuracy? By testing accuracy across skin tones and reporting results per group rather than as a single average, so a failure on darker skin cannot hide inside an overall number.

What happens to my health data? It is kept anonymised and aggregated, is never sold, and is used in line with India's data protection rules to understand community health needs.