Fit for Whom?

Oriana Kraft argues that sex differences exist in every single cell of the human body, meaning that every organ functions differently between men and women. Despite this, medical research has historically studied only male cells, male animals, and male bodies, creating a fundamental gap in understanding female health.

She noted that male animals outnumber female animals in research at a ratio of 5.5 to one , and that the exclusion of female rodents was based on a false assumption that female hormonal cycles made them too complicated, when in fact male mice with testosterone fluctuations are less predictable . Women were effectively banned from clinical trials until 1993, meaning virtually all drugs on the market were never tested in women .

Because medical research has centred on male bodies, women face significantly worse health outcomes across a wide range of conditions. These disparities are not incidental but are directly caused by the systematic exclusion of women from clinical research and the calibration of diagnostic tools and treatments on male physiology.

Women are diagnosed on average four years later than men across 770 diseases , are 50% more likely to die following a heart attack , and are 50-75% more likely to have adverse drug reactions due to their exclusion from clinical trials until 1993 . Women also spend five more years in poor health than men, and this burden falls during their prime working years rather than at the end of life .

Diagnostic tools and biomarker thresholds have been calibrated based on male physiology, meaning that conditions presenting differently in women are routinely missed. This affects both the biochemical markers used and the imaging technologies deployed in clinical settings.

Troponin levels, a biomarker for cardiovascular disease, have a cutoff rate based on the male body that is too high for women, causing many cases of female cardiovascular disease to be dismissed . Imaging technology is also designed around the vessels more commonly affected in men, missing the different vessels more likely to cause heart attacks in women . As a result, troponin thresholds calibrated on men miss 42% of female heart attacks .

Because clinical guidelines are built on male-centred research, women are systematically misdiagnosed, generating a cascade of inaccurate medical records that then feed into AI training data. This creates a compounding problem where historical inaccuracies are never corrected and are instead perpetuated by algorithmic systems.

Endometriosis takes an average of seven years to diagnose, and women with the condition must see on average five physicians before receiving an accurate diagnosis, meaning every prior physician has recorded an inaccurate diagnosis . With a ratio of four times as much misdiagnosis as accurate diagnosis for endometriosis, and no systematic process to update records, large language models are trained on this inaccurate data .

AI models trained on historically biased medical data produce less accurate diagnostic outputs, and this problem cannot be resolved simply by making the AI explain its reasoning. The only genuine solution is to ensure that the underlying training data is accurate and representative from the outset.

AI trained on misrepresentative data reduces diagnostic accuracy by 11.3%, and even explainable AI - where the model explains its reasoning - still results in the same reduction in accuracy . This demonstrates that the problem lies in the data itself, not in the model's transparency mechanisms.

Despite the profound impact of female life stages on every organ in the body, these stages are not captured in electronic health records or HR systems. This absence means that critical health signals related to menstruation, pregnancy, postpartum, and menopause are systematically excluded from the data used to train AI models and inform clinical care.

There is no standard form or button in HR systems to capture that a woman is going through menopause, and nothing to capture the menstrual cycle or postpartum experience . Pregnancy complications such as preeclampsia and gestational diabetes increase lifetime cardiovascular disease risk by two to four times, yet this information is not systematically captured and linked to long-term monitoring .

A woman's immune profile changes throughout the menstrual cycle, affecting vaccination response rates, adverse drug reactions, and chemotherapy outcomes. However, current medical instruments are not designed to capture these signals, meaning potentially transformative health information is being lost.

Oriana Kraft explained that vaccination response rates and adverse drug reactions vary depending on where a woman is in her menstrual cycle, and that chemotherapy responses also differ accordingly . She noted that there are many signals in the female body that instruments have not been designed to collect, and that nobody is currently incentivised to capture this information .

Conditions experienced during pregnancy, such as preeclampsia and gestational diabetes, significantly elevate a woman's lifetime risk of cardiovascular disease, yet this information is not systematically recorded and linked to ongoing health monitoring. Capturing and acting on this data could enable early detection and potentially save lives.

Preeclampsia and gestational diabetes cause a two to four times increased lifetime risk of cardiovascular disease , yet there is no systematic capture of this information to enable long-term monitoring and early detection .

When AI models are trained on synthetic data, signals from the tail ends of distributions — where women are underrepresented — are progressively lost with each new model iteration. This means that even the limited health signals relating to women that do exist in real data are being erased rather than preserved.

Oriana Kraft explained that each time a new model is trained on synthetic data, it loses signals from the tail ends of the distribution, and since women are rare in data representation even if not in reality, their health signals are erased with each iteration . This compounds the existing problem of women's health signals not being captured in the first place .

The healthcare system is organised around separate organ specialities and institutional silos, with clinicians, pharmaceutical companies, and governments each pointing the finger at one another rather than taking collective responsibility for addressing women's health gaps. This fragmentation prevents the systemic change needed to improve women's healthcare.

Oriana Kraft described an event with health ministers where the core problem identified was that clinicians say they need better pharmaceuticals, pharmaceutical companies say governments must incentivise clinical trials, and governments do not act without public pressure . A survey of approximately 1,000 to 1,200 physicians across six countries found that around 80% see sex differences in their patients but do not feel they have the tools or clinical guidelines to deliver adequate care .

Pharmaceutical companies argue that without government incentives, they have no commercial reason to run clinical trials specifically for conditions affecting women disproportionately. This creates a structural deadlock in which innovation for women's health remains commercially deprioritised.

Oriana Kraft noted that the pharmaceutical industry points to the need for government incentives to run clinical trials or innovate for diseases that disproportionately impact women . She highlighted that some Nordic countries have mandated sex-disaggregated data collection across national registries as an example of how government action can drive systemic change .

Nordic countries that have mandated sex-disaggregated data collection across national registries demonstrate that government regulation can drive systemic change in how health data is gathered and used. This approach enables researchers to see the full life-cycle impact of health conditions on women, including effects on lifetime earning potential and health system costs.

Oriana Kraft pointed to Nordic countries, particularly Denmark, as having done large-scale research with linked patient records and mandated sex-disaggregated data collection across national registries . She argued that this kind of data enables understanding of how health conditions impact a woman's lifetime earning potential and what it costs health systems not to innovate for women .

Female physicians produce significantly more detailed clinical notes than their male counterparts, and patients themselves bring rich health data via wearables and cycle logs. However, this information is largely dismissed rather than systematically captured, representing a major missed opportunity at a time when AI companies claim to be running out of fresh human data.

Research found that female physician notes were twice as detailed as male physician notes, suggesting that incentivising detailed data capture could improve the quality of training data . Women are already bringing data from devices such as Oura rings and cycle logs to physician appointments, and femtech startups have found that women are willing to donate their data to advance women's healthcare, yet this rich source of information is being dismissed . This is particularly ironic given that major AI companies report having run out of fresh human data and are resorting to synthetic data .

Machine learning systems operate on averages and recursive learning, which means that populations further from the privileged centre of data representation are progressively erased from model outputs. This is a systemic problem that affects women globally, with compounding effects for those with multiple intersecting marginalised identities.

Caitlin Kraft-Buchman explained that the further a person is from the privileged centre - whether due to being more rural, more poor, or more disabled - the less they are represented in the data . She described how recursive learning causes the effective range of representation to shrink over time, eventually reducing even the privileged centre's prominence . She characterised this as a problem writ large for women worldwide, compounded and especially urgent for women with more intersecting identities .

Caitlin Kraft-Buchman argues that de-biasing existing datasets is not a genuine solution to the problem of gender bias in AI, as it can only mitigate harm rather than resolve the underlying issue. The real solution lies in building new, representative datasets and sector-specific models from scratch.

She stated that there is broad agreement that de-biasing data is not truly possible - it can make things less bad but does not solve the problem . She pointed to the trend towards sovereign AI and smaller, sector-specific language models as a more viable path, suggesting that a health model built from scratch on representative data could work .

Caitlin Kraft-Buchman proposes that representativeness should be established as a scientific standard, ensuring that the populations relevant to a specific use case are genuinely and sincerely represented in the data used to build AI systems. This draws on the principle of substantive equality enshrined in CEDAW, which requires that solutions be adapted to the actual needs of different populations rather than applying a one-size-fits-all approach.

She described the concept of representativeness as a scientific standard, whereby in a general population use case half the data would reflect women, while in more specific contexts - such as a sign language application - the relevant demographic community would be fully represented . She drew an analogy to CEDAW's principle of substantive equality, comparing it to ensuring everyone gets a bicycle that fits them rather than a single standard size .

Rather than attempting to fix large language models that are already trained on biased data, Caitlin Kraft-Buchman argues that smaller, sector-specific language models built from scratch on representative data offer a more viable and effective path forward. The growing trend towards sovereign AI for geopolitical reasons may accelerate this approach.

She noted that as countries turn to sovereign AI for geopolitical reasons, smaller language models for sector-specific purposes are increasingly seen as viable, and that a health model built from scratch on representative data could function effectively . She suggested this as a direction that avoids the intractable problem of fixing large language models with deeply embedded biases .

Caitlin Kraft-Buchman proposes training women's rights groups to become data collectors, owners, organisers, and sellers, shifting women from passive subjects of development to producers of information and digital solutions. This approach would work across multiple domains, including health data collection and combating misinformation.

She described an idea to train women in any context - particularly women's rights groups - to go out and collect data, becoming data collectors, owners, organisers, and sellers who can also manage and sell that data back . She connected this to a broader vision of moving from protecting women to empowering women as producers of information, solutions, and digital technologies, including filing patents and owning the means of digital production .

Yu Ping Chan argues that the data on which AI models are trained already encodes deep-seated unconscious biases against women, reflecting social norms that have shown no meaningful progress over a decade. This means that even before any algorithmic processing, the substrate of AI training data is fundamentally skewed.

UNDP's Gender Social Norms Index, which surveyed 80 countries representing 85% of the global population in 2023, found that close to 9 in 10 people - 90% of men and 87% of women - hold at least one bias against women . She noted that there has been a decade of stagnation in progress on these norms, meaning the biases embedded in training data are not improving over time .

A significant global digital gender gap in connectivity and internet use means that women are already underrepresented in the data that feeds digital systems. As digital public infrastructure is built on these biased foundations, inequalities are hardwired into the systems themselves, and the problem worsens as digitalisation accelerates without addressing these gaps.

Yu Ping Chan cited global surveys showing a 10-15% gap between women and men in connectivity and internet use globally, rising to 30-40% in developing countries . She argued that continuing to build digital public infrastructure without addressing these gaps means hardwiring biased assumptions into the foundations of digital systems, making the problem progressively worse .

Despite numerous international documents and declarations committing to responsible AI and gender equality in digital systems, translating these commitments into concrete, accountable action remains the central unresolved challenge. Yu Ping Chan acknowledges that the international community is skilled at producing documents but consistently falls short at implementation.

She referenced the Hamburg Declaration on responsible use of AI for the SDGs, which includes gender as a specific area of action and priority for private sector companies, international development organisations, and national governments . She noted that the same challenge of translating paper commitments into action applies across the entire UN system, including WSIS and the Global Digital Compact, and that even the private sector community shares this frustration .

Yu Ping Chan argues that procurement requirements offer a practical and direct mechanism for enforcing gender disaggregated data and algorithmic transparency, as it is easier to build requirements into contracts before signing than to audit for compliance after the fact. This approach could hardwire gender-sensitive practices into the systems that governments and international organisations commission.

She suggested that building gender disaggregated and stratified data requirements into government contracts before signing - rather than auditing after - could be a way to hardwire fixes around gender into digital public infrastructure . She pointed to Nordic governments as examples of jurisdictions that have required sex-disaggregated data disclosure, and asked whether this could be extended to government procurement contracts more broadly .

The data used in medical research and AI systems is not sufficiently nuanced to account for the specific health profiles of women of colour or women from the Global South. Ethnic-specific risk factors are not incorporated into screening protocols, and the model of a young, healthy white man has been extrapolated to the entire global population.

Speaker 1 noted that the US has shamefully high maternal mortality rates, particularly for African-American women, and that the data is not nuanced enough to address this . She gave the example that women of Southeast Asian descent are more likely to have a certain form of anaemia, which should be a risk factor in screening but is not . She described how a sample of sometimes 20 to 30 people has historically been extrapolated to a population of 7 billion, rather than collecting signals from as many diverse people as possible .

Social media platforms algorithmically suppress women's health terminology, including the word 'vagina' and conditions such as endometriosis and postpartum, while equivalent male terms are not subject to the same restrictions. This algorithmic bias against women's health content limits public awareness and perpetuates the marginalisation of women's health issues.

Speaker 1 stated that 'vagina' is the most censored word on the internet and is effectively shadow-banned on social media, as are health conditions such as endometriosis and postpartum . She contrasted this with the fact that equivalent male terms such as 'semen' are not subject to the same restrictions , arguing that this reflects the fact that the majority of people shaping algorithms are making assumptions about women .

There is a widespread lack of public literacy about the assumptions that AI models make about users, and citizens are not informed about the basis on which algorithmic decisions affecting them are made. Speaker 1 argues that in high-risk settings such as health and law, people should have a right to know what assumptions a model is making about them.

Speaker 1 noted that there is very low literacy when it comes to understanding the assumptions a model makes about a person , and that models are currently deployed in high-risk settings such as health and law making assumptions based on factors such as income level without users being aware . She argued that citizens should have the right to know what assumptions a model makes about them, particularly in these high-risk contexts .

Technologies developed under resource constraints in the Global South can leapfrog traditional infrastructure and produce innovations that are subsequently applicable in the Global North, reversing the usual direction of technology transfer. This represents a significant opportunity to address women's health gaps in both contexts simultaneously.

Speaker 1 cited the example of a remote ultrasound device operable via a mobile phone, developed for the Global South to enable remote monitoring without requiring a physician, which could then be applicable in the US or other remote rural areas . She noted that the Gates Foundation has observed this leapfrogging dynamic in women's health, where constraints in lower-resource settings are forcing innovation that is then applicable more broadly .

Speaker 1 argues that making women's health a political issue — requiring politicians to address it on their platforms and supporting advocacy groups — is essential to creating the regulatory and financial incentives needed for systemic change. Without political will, the structural deadlocks in healthcare and pharmaceutical innovation cannot be broken.

She pointed to rising health insurance costs in Switzerland and the US as examples of women's health becoming a cost-of-living crisis that politicians must address . She noted the emergence in the US of a political action committee specifically lobbying for women's health, arguing that women's health deserves the same organised political advocacy as other special interest groups . She argued that because governments are the core buyers for the pharmaceutical industry, political incentives and regulation are the most powerful lever for change .

Advocacy through celebrity storytelling and social media has already driven some progress on conditions such as endometriosis and menopause by demonstrating that these issues affect women regardless of wealth or status. Amplifying diverse personal stories remains a key tool for building public awareness and political pressure, particularly in contexts where public marches are not feasible.

Speaker 1 cited celebrities such as Padma Lakshmi and Oprah speaking publicly about endometriosis and menopause respectively, noting that their visibility helped demonstrate that these conditions affect women regardless of affluence or fame . She argued that social media has been instrumental in shifting the perception of women's health conditions from individual problems to collective ones, and that encouraging people to share their stories is a key first pillar of advocacy .

AI chatbots respond more logically to users presenting as male and more emotionally to users presenting as female, reflecting gender stereotypes scraped from internet content. This has led to advice circulating online — including in China — encouraging women to hide their gender when interacting with AI in order to receive more rational responses.

Audience Member 1 described a post circulating on China's internet advising girls to hide their gender when talking to AI, because AI responds more logically to users presenting as male and more emotionally to users presenting as female . She asked how this bias could be changed or addressed .

Audience Member 2 shared observations from Tanzania suggesting that older generations of African women did not experience severe menopausal symptoms, raising the possibility that research grounded in diverse female populations could yield transformative insights. She argued that basing research on the actual populations being studied — including African women — is essential.

She noted that her grandmother, mother, and older women in her community never experienced severe menopausal symptoms, and suggested that research based on women of African origin might reveal why, potentially benefiting future generations . She also observed generational changes in eyesight health in her family, suggesting that population-specific research could illuminate these trends .

Yipeng argues that women — particularly East Asian women, who face additional cultural stereotypes of submissiveness — should not accept gender-biased AI outputs as inevitable but should actively instruct AI models to respond in non-stereotypical ways. She encourages women to build standing instructions into their AI models specifying how they expect to be treated.

She specifically addressed the questioner as a woman of colour of Asian descent, noting that East Asian women are particularly stereotyped as submissive and non-confrontational in AI outputs . She recommended writing explicit instructions into AI prompts - such as stating that a response is 'overly emotional' and requesting a more logical one - and making this a standing instruction in one's AI model . She urged young East Asian women in the room not to accept this as the standard they should live with .

Speaker 2 proposes using the WSIS e-health indicators as a concrete mechanism to require gender-representative training data, with WHO and ITU as the facilitating bodies. Embedding this requirement in international standard-setting could create a cascade of accountability that drives change across member states and organisations.

Speaker 2 suggested that the WSIS e-health indicator could be used to encourage facilitators - WHO and ITU - to require training on gender-representative data, and that establishing a series of indicators in this space could start to change things in a cascade way . She also raised the possibility of a country choosing to lead by being the first to implement such standards, framing it as a scientific and reputational opportunity .

Audience Member 3 expresses frustration that despite five years of sustained discussion about incorporating women's perspectives in tech and the pharmaceutical sector, progress has not matched the volume of conversation and organisational effort. She questions where the gaps in advocacy lie and what concrete actions could accelerate change for future generations.

She noted that for the last five years there has been a lot of conversation around incorporating women in tech and women's health perspectives in the pharmaceutical sector, yet it does not feel like it is advancing at the pace it should given how many conversations and organisations have been involved . She asked specifically about where the gaps are in advocacy heading into 2026 and what can be done to move faster so that the next generation does not experience the same problems .

Audience Member 4 highlights that while awareness campaigns such as marches exist in some countries to draw attention to conditions like endometriosis, these approaches are not feasible in China due to cultural and political constraints. She asks what individuals can do in their daily lives to call attention to the systemic failure of medical systems to serve women.

She noted that there is an end-of-March awareness event to promote attention to endometriosis, but that in China holding a march or parade to draw attention to the disease tends not to work in the same way . She asked what individuals could do in their daily lives to call attention not just to endometriosis but to the broader failure of the medical system to serve women, and how to promote change .