AI Does Not Remove Historical Bias from Medicine. It Can Convert That Bias into Code.

Where clinical assumptions are incomplete, AI can make them permanent.

Recent examples of AI-supported medical technology have made me pause and ask a difficult question: are we using AI to improve healthcare, or are we at risk of automating the same blind spots that already exist within it?

This is not about singling out individual companies or assessing any particular product. It is about a much wider issue that deserves far more attention as AI becomes embedded in patient intake, symptom assessment and clinical decision support.

Technology is often presented as objective. But AI does not arrive without context. It is shaped by the evidence, definitions, assumptions and clinical priorities used to build it. And medicine's understanding of women has never been complete.

AI can capture symptoms and still miss the context

AI-powered patient intake tools have enormous potential. They can give patients more time to describe what they are experiencing, organise complex information and help clinicians begin appointments with a clearer view of the patient's history. That could be genuinely valuable.

But collecting symptoms is not the same as understanding them.

A woman may describe changes in mood, energy, sleep, weight, concentration, libido, bleeding, pain, bladder function or general wellbeing. There may be several possible explanations. It may be entirely appropriate to consider thyroid conditions, nutritional deficiencies, gastrointestinal problems, mental health conditions and other causes.

The concern is not that those possibilities are explored. The concern is what happens when relevant hormonal, reproductive or gynaecological explanations are not considered at all. Not as a definite diagnosis. Not as the only explanation. Simply as part of the wider clinical picture.

Perimenopause and menopause are one example. The same concern applies when conditions such as endometriosis, adenomyosis, polycystic ovary syndrome, menstrual disorders, pelvic health conditions or other aspects of women's health are absent from the system's view.

A system can record every symptom accurately and still fail to ask the questions needed to understand how those symptoms may connect.

AI is built on the medicine we already have

Clinical technology is shaped by many things:

• the research available to developers
• the way conditions are defined
• the symptoms associated with those conditions
• the patient groups represented in the evidence
• the questions the system is designed to ask
• the clinical assumptions used to interpret the answers

None of these choices is neutral. They reflect the healthcare system, evidence base and medical understanding that already exist. That matters because women's health has historically been under-researched, inconsistently taught and too often considered through a narrow clinical lens.

Symptoms may be separated rather than viewed as part of a wider pattern. Pain may be normalised. Hormonal changes may be overlooked. Gynaecological symptoms may be treated as isolated complaints. Mental and physical symptoms may be divided even when they are closely connected.

If important experiences have not been adequately studied, recorded or connected within the evidence base, AI will not automatically discover what is missing. It may reproduce the gap.

Women's hormonal and gynaecological health is often understood too narrowly

Perimenopause and menopause have frequently been reduced to a familiar set of symptoms: hot flushes, night sweats, irregular periods, periods stopping. Those experiences matter, but they are not the whole picture.

Women may also report changes in:

• mood
• sleep
• memory and concentration
• energy
• sexual desire
• bladder and vaginal health
• joints and muscles
• headaches
• skin
• body composition
• overall wellbeing

The same narrow thinking can affect other areas of women's health. Endometriosis may be reduced to painful periods, despite its wider effects. Polycystic ovary syndrome may be viewed primarily through fertility or weight. Pelvic pain may be normalised. Heavy or irregular bleeding may be treated as something women are simply expected to tolerate. Sexual, vulval and vaginal symptoms may be poorly explored because patients find them difficult to raise and clinicians may have limited time to ask.

Not every symptom experienced by a woman is caused by hormonal or gynaecological change. That would be an equally unhelpful assumption. But those possibilities should not become invisible simply because a woman does not present with a stereotypical textbook picture. The right response is not automatic attribution. It is appropriate consideration, relevant questioning and clinical judgement.

Bias does not always look obvious

A system does not have to make an openly discriminatory statement to reproduce inequality. Bias can appear through omission. It can happen when:

• women's health conditions are defined too narrowly
• menstrual, reproductive or hormonal context is not collected
• pain, mood or sleep symptoms are considered in isolation
• changes in cognition, libido or bladder function are not explored
• age and life stage do not trigger relevant follow-up questions
• symptoms are mapped more strongly to better-documented conditions
• the system does not recognise patterns that sit across several clinical domains

Each decision may appear reasonable on its own. Together, they can create a clinical picture that looks complete while something important remains unseen.

Confidence is not the same as completeness

One of the most powerful features of medical technology is also one of its greatest risks: the appearance of authority. A polished report, coded clinical terminology or a numerical confidence score can make an output feel more certain than it really is. But a system can only express confidence within the framework it has been given. A confidence score does not tell us:

• whether all relevant possibilities were included
• which questions were never asked
• how broadly a condition was defined
• which groups were represented in the evidence
• whether symptoms were interpreted in the context of sex, age and life stage
• what may have been overlooked before the calculation began

Precision is not the same as accuracy. And accuracy within an incomplete framework is not the same as a complete clinical picture.

The danger is authoritative incompleteness

An obviously implausible output is relatively easy to challenge. A polished, plausible and professionally presented output is harder. Patients may assume the technology has considered every relevant explanation. Clinicians may understandably focus on the information placed in front of them. Organisations may interpret technical sophistication as evidence that bias has already been addressed. But sophistication does not remove the need for scrutiny.

We still need to ask:

• Who shaped the clinical model?
• Which patients were represented?
• How were women's symptoms understood?
• How were hormonal, reproductive and gynaecological conditions defined?
• Which symptoms trigger further questions?
• Which experiences are considered significant?
• What does the system fail to ask?
• How are gaps identified once the technology is being used?
• Who is responsible for challenging the assumptions built into it?

Specialist clinical design matters

There is a growing temptation to believe that a broad medical model can simply add women's health as another category. I do not think it is that simple. Women's health is not one neat specialty or one stage of life. It includes hormonal, reproductive, menstrual, gynaecological, sexual, pelvic and post-reproductive health. These areas overlap with cardiovascular health, mental health, bone health, metabolic health, cancer risk and many other parts of medicine.

That complexity requires specialist knowledge, clinical depth and a willingness to challenge conventional assumptions. It requires input from clinicians who understand the relevant pathways and from women who understand the lived experience. It requires testing beyond neat textbook examples. And it requires recognition that women are not one homogeneous patient group. Age, ethnicity, disability, socioeconomic circumstances, sexual orientation, gender identity, pregnancy history and access to care may all shape how symptoms are experienced, reported and interpreted.

Most importantly, it requires humility. No clinical technology should be treated as complete simply because it is sophisticated. The responsible position is not to claim that bias has been eliminated. It is to recognise that bias is possible, look for it deliberately and build systems that can be questioned, reviewed and improved.

We should not automate what healthcare has failed to see

AI could make healthcare better. It could give patients more time to explain what is happening. It could help clinicians work with clearer, more structured information. It could identify patterns that fragmented consultations currently miss. But none of that happens automatically. AI reflects the quality of the clinical thinking beneath it. Where women's health has been overlooked, narrowly defined or poorly understood, technology must be designed deliberately to avoid carrying those weaknesses forward.

Otherwise, we risk taking old blind spots, embedding them into software and deploying them at scale.

AI does not remove historical bias from medicine. Without deliberate clinical design, specialist scrutiny and meaningful representation, it can convert that bias into code. Then present it as progress.

This article reflects the author's views on the responsible design of clinical technology. It is intended as industry commentary and does not assess any individual product, patient case or clinical decision.