91探花

What was once elusive is becoming somewhat inescapable.

AI used to be an incredible concept, the very thought of which was challenging to grasp but inexplicably exciting to consider. It created opportunity where there was none and opened doors to a future few ever dreamed they would live to see. AI was simultaneously exciting and scary, and for all intents and purposes, it still is.

What鈥檚 changed, however, is the fact that it鈥檚 no longer a figment of our excited imaginations and anxious longing; a mere component of complex, advanced technology far out of the reach of ordinary people. Now, it鈥檚 everywhere 鈥� whether we like it or not. An unavoidable feature of every app, an addendum to document, a division at every tech company and a hot topic in every conversation.

At least, that鈥檚 been my experience, and I hardly think I鈥檓 alone in feeling this way.

So, what鈥檚 the problem? I鈥檓 not against AI nor am I participating in AI fear mongering that crowds online chatrooms and floods Baby Boomers鈥� WhatsApp chats. I use it all the time and I encourage others to do the same.

But, AI does make me nervous. Just not for the reasons you may think.

听

AI Dooms Day Anxiety: Legitimate Fear or Tired Trope?

听

The whole 鈥淎I apocolypse鈥� trope is very dystopian in nature, latching onto tired old technology-related panic from years past combined with sci-fi-esque 鈥渆nd of the world鈥� imagery. And sure, these things certainly do make one stop and think 鈥� they make us wonder what we鈥檙e creating and how much control we鈥檒l have over it.

Of course, the primary fear on the minds of AI laymen (and by that, I mean non-AI-experts) is Artificial General Intelligence (AGI) and the development of consciousness. This concern is promptly followed by visions of angry robots turning on humans and the end of the world as we know it 鈥� sure, we鈥檝e all seen the movies.

However, I believe this fear is misplaced. I鈥檓 not saying we should totally neglect these concerns 鈥� please don鈥檛 call me when the robots turn on us, I鈥檝e been saying 鈥減lease鈥� and 鈥渢hank you鈥� to ChatGPT. Rather, I think the dystopian future we all picture so vivdly is not our most immediate concern (if one at all). But I鈥檓 not saying don鈥檛 worry 鈥� I鈥檓 saying we need to worry about something else.

听

听

What If the Real Risk Is The One That Looks Harmless?

听

Sometimes, the real threat, or the most significant threat, is the one we don鈥檛 see coming. And yes, I鈥檓 well aware that I sound like Obi-Wan handing off an earnest piece of advice in a deleted scene from 鈥淪tar Wars: The Phantom Menace鈥�.

But I stand by the sentiment. While the world chatters about AGI, GPUs and compute, rushing to create and implement national and even global AI regulation in record time 鈥� all the big and most important issues and concerns 鈥� we鈥檙e letting the seemingly smaller issues slip through the cracks. And the problem is, they may seem like small issues compared to things like data sovereignty and mass redundancies, but these 鈥渟mall鈥� issues could have serious implications.

They already are. They may not be bold and obvious just yet, we may not be seeing the effects everywhere we go, but the consequences are beginning to crop up increasingly often.

听

An 鈥淎ssumption Failure, Playing Out at Civilisational Scale鈥�

听

Much like the problem may seem 鈥渟mall鈥�, it鈥檚 also quite simple. It鈥檚 about implicit bias and disproportionate representation, and the effects that these things can and will (and already are having) on society. Of course, this argument could and should be made for demographics and representation in all contexts, but for now, I鈥檓 focusing on gender 鈥� because:

1. There is a decent amount of research that鈥檚 already been done on the topic.
2. It offers a case study, an example, to make a broader argument.

So, I鈥檒l put it quite simply. The issue at hand, the one I鈥檝e alluded to and frustratingly danced around for the better part of 500 words, is this: women are disproportionately involved in artificial intelligence in comparison to men.

The phrase 鈥渋nvolved in AI鈥� is,听 agreeably, awkward and vague, but bear with me 鈥� there鈥檚 good reason. Indeed, when I say 鈥渋nvolved in鈥�, I鈥檓 talking about many things, but most of all, women who are actively working on developing actual AI technology (the complicated stuff), women working 鈥渋n鈥� AI (the AI industry, shall we say) and actually using AI technology themselves.

According to a report published by Forbes in April 2025, women both adopt and actively use AI tools far less than men who formed part of the study. The Financial Times revealed in a June 2025 article that a Danish study involving 100,000 workers showed that women are 20 percentage points less likely than men to make use of common AI tools like ChatGPT. LeanIn exposed that according to another study, women are not only less likely to use AI both at work and at home, but they also feel more anxiety around using AI tools, they鈥檙e less likely to be encouraged to use AI and overall, the feel significantly less positive about AI in general.

There have been countless studies conducted across time, regions and age groups, and they all seem to tell us the same thing. Men use AI more than women.

And this is only AI use. As you may have already guessed, the industry itself reflects the same pattern. There are significantly more men than women hired to fill AI-related professional roles.

So what does this mean?

Well, first and foremost, women are obviously underrepresented, and there are plenty of reasons for this 鈥� the types of jobs women do compared to men (ie. the continuation out of traditional gender norms), general gender perceptions, access and so much more. Honestly, this is a massive debate in itself, and I鈥檓 going to bypass it not because it鈥檚 unimportant, but becasue I think that it often becomes a roadblock that prevents us from reaching the next issue. And the next issue, the one I鈥檓 focused on, is what this underrepresentation of women actually means for us 鈥� 鈥渦s鈥� being society as we know it.

Because when I started considering this issue and looking into it, I became alarmed quite quickly. And I鈥檓 sorry to say it, but further research and discussion with experts has escalated that concern exponentially. Inequality and gender issues are already massive problems we face today, and it seems as though we may, unknowingly, be exacerbating the problem in epic proportions.

I don鈥檛 want to be dramatic, but Sayali Patil, AI Infrastructure Reliability Specialist, has described the situation as an 鈥渁ssumption failure鈥� that is currently 鈥減laying out at civilisational scale鈥�. And if that, coming from a technical AI expert of epic proportions, doesn鈥檛 concern you, it really should.

听

The Implications of Female Underrepresentation In AI

听

The problem we鈥檙e facing is that if more men than women are building AI systems, shaping datasets and using these tools daily, then we have to ask a simple question 鈥� can AI really serve everyone equally?

After all, AI learns patterns. It learns what 鈥済ood鈥� looks like, what 鈥渟trong鈥� sounds like and what 鈥渃onfidence鈥� reads like. However, those patterns don鈥檛 just appear out of nowhere 鈥� they come from data, and that data is shaped by human behaviour. It鈥檚 data that we feed it, data that comes from humans and is selected by humans, and if that behaviour is disproportionately male, then the definition of 鈥渘ormal鈥� risks becoming male too.

I鈥檓 not saying that AI will intentionally favour men 鈥� at least not in general terms 鈥� and that鈥檚 actually what makes the problem more concerning. The bias is rarely deliberate; it鈥檚 subtle, it鈥檚 structural and often, it鈥檚 invisible.

Syed Asif Ali shared with me a striking example while testing an AI hiring tool in Dubai. He noticed the system repeatedly downgraded some female candidates, despite strong skills and relevant experience. The only noticeable difference, after curious analysis, was tone. That is, candidates wrote in a less aggressive, less self-promotional style.

鈥淲hen we looked closer, it made sense,鈥� he explained. 鈥淭he model had been trained on data where that more direct style was treated as a signal of confidence. So anything outside that just鈥� looked weaker to it.鈥�

Technically, nothing was broken. The system wasn鈥檛 programmed to discriminate, yet, the outcome still skewed in one direction. As Ali put it, 鈥渂ias in AI isn鈥檛 always loud or obvious. Sometimes it鈥檚 just one style quietly becoming the default, and everything else getting pushed down without anyone noticing.鈥�

And I think that this example highlights the risk I鈥檓 raising better than I could explain, because it鈥檚 real. AI doesn鈥檛 just reflect data; it standardises it. Once a particular communication style, career path, behaviour pattern or language becomes associated with success (in this example), the system begins reinforcing it. And over time, that pattern scales.

This is where representation becomes critical. If the people building AI systems come from similar backgrounds, then the definition of what looks 鈥渞ight鈥� becomes narrower. Ali warns that 鈥渋f the people building these systems all come from similar backgrounds, the definition of 鈥榥ormal鈥� gets very narrow. And once that鈥檚 baked into the system, it scales fast.鈥�

The challenge is that we may not even know what鈥檚 causing the bias. It might not be obvious, such as variables like gender labels. It could be tone, phrasing, career gaps, communication style or behavioural patterns that correlate with one group more than another in ways that we haven鈥檛 necessarily noticed yet. These signals are subtle, and once embedded in large datasets, they become extremely difficult to isolate.

Indeed, once we create this problem, it becomes harder to fix. If we can鈥檛 clearly identify the bias, we can鈥檛 easily remove it, and if AI systems are deployed at scale before we address it, those patterns risk becoming normalised across hiring, performance reviews, education, finance and beyond.

听

The Gender Gap In AI Could Shape Society

听

This is why the gender gap in AI matters 鈥� it鈥檚 not just about fairness in the workforce (albeit really important), it鈥檚 about shaping the technology that will increasingly shape society as a whole. If women are underrepresented in building and using AI, then their experiences, behaviours and communication styles may be underrepresented in the data that trains it.

The result isn鈥檛 necessarily overt discrimination; it鈥檚 most likely going to be something quieter and potentially more unintentionally sinister. It鈥檒l be systems that subtly favour one way of working, tools that reward one tone over another or models that learn from patterns that don鈥檛 fully reflect everyone.

And because AI operates at scale, those small biases don鈥檛 stay small. They grow, and they grow quickly.

This is the uncomfortable reality in which we find ourselves. The problem isn鈥檛 that AI will intentionally exclude women 鈥� rather, it鈥檚 that it may unintentionally optimise around male-dominated patterns simply because those patterns appear more often in the data. And once that happens, the technology we rely on every day could quietly reinforce the very inequalities we hoped it would help solve. The inequalities we鈥檝e been working to mitigate and challenge for centuries.

If AI is going to shape the future, then representation in building it matters 鈥� not just for fairness, but for accuracy. Because when half the population is underrepresented, the system isn鈥檛 just biased, it鈥檚 incomplete.

听

Our Experts:

听

I spoke to a group of experts on the topic:

听

• Sarah Hoffman:听Director of AI Thought Leadership at AlphaSense
• Jenny Briant: Director of Talent Strategy at Ten10
• Maria Nugroho: AI Enterprise Strategist
• Ana-Maria Badulescu: VP of AI Labs at Precisely
• Charlotte Wilson: Head of Enterprise Business at Check Point Software
• Sayali Patil: AI Infrastructure Reliability Specialist
• Syed Asif Ali: Founder and Digital Identity Architect at Point Media
• Michael Ferrara: Technology Contributor and Legal IT Practitioner at Conceptual Technology
• Edward Tian: CEO of GPTZero
• Emma Irwin: Director of Sales Engineering at Dataiku
• Faye Ellis: Principal Training Architect, Pluralsight
• Kristyna Vlckova: VP of Growth Marketing at GoodData

听

Here鈥檚 what they had to say.

听

Sarah Hoffman, Director of AI Thought Leadership at AlphaSense

听

听

鈥淎I has the potential to move us forward in extraordinary ways. But even as systems become more autonomous, they are still shaped by human decisions. If those perspectives are too narrow, the outcomes will be too. We鈥檝e already seen how bias can surface in areas like hiring and healthcare.

鈥淎I is quickly becoming foundational infrastructure for how work is done and how decisions are made. If AI is going to positively impact the future of work for women, then women across cultures and communities need to help shape it. Without women鈥檚 contributions, we risk carrying yesterday鈥檚 assumptions into tomorrow鈥檚 digital infrastructure.鈥�

听

Jenny Briant,听Director of Talent Strategy at Ten10

听

听

鈥淎I is not being introduced into a neutral environment, and that鈥檚 where the risk lies. Recent data from the International Labour Organisation shows that female-dominated roles are almost twice as likely to be exposed to generative AI than male-dominated ones, which means the people most affected are not always the ones influencing how these tools are designed, tested or governed.

鈥淭hese systems are built on past decisions and behaviours. If those reflect uneven access to opportunities or progression, AI can end up reinforcing those same patterns in how people are assessed, hired or supported at work.

鈥淭he issue is not the technology itself, but who is shaping it. If women are underrepresented in the teams designing, testing and governing AI, their perspectives are missing from critical decisions.

鈥淚f we want AI to close gaps rather than widen them, we need diverse representation, strong governance and consistent human oversight built in from the start.鈥�

听

Maria Nugroho,听AI Enterprise Strategist

听

听

鈥淭his isn鈥檛 just a diversity issue; it鈥檚 a technology and commercial quality issue, and we need to say that more loudly.
Only 22% of the global AI workforce is female (WEF, 2025). Just 14% of AI research papers have a female first author (Stanford AI Index, 2025). These aren鈥檛 abstract statistics; they represent who decides which problems AI solves, which datasets get used, and whose experiences get encoded into systems shaping enterprise decisions worldwide.

鈥淭he commercial consequence is measurable: AI products built by gender-diverse teams show 15% fewer bias-related errors (McKinsey, 2024). Homogeneous teams produce homogeneous outputs. When AI misrepresents half the population, organisations deploying it inherit that blind spot and pay for it in failed adoption and missed value.

鈥淲e are not just building technology. We are building the infrastructure of future economies. If women aren鈥檛 in the room, we aren鈥檛 building for the full market and no amount of post-deployment patching fixes a foundation built without us.鈥�

听

Ana-Maria Badulescu, VP of AI Labs at Precisely听

听

听

鈥淭he future of AI and data innovation is dependent on the diversity of the perspectives shaping it. Now that AI is increasingly embedded across every sector, we have a responsibility to make AI truly reflective of the society it serves. We must act now if we want AI to be a force for equity and innovation, rather than exclusion, and that begins with diversity amongst those who work on the technology itself.

鈥淲hen we create space for more women 鈥� and for people of all backgrounds and experiences 鈥� we build technology that is stronger, more creative, more equitable, and ultimately more impactful.鈥�

听

Charlotte Wilson, Head of Enterprise business at Check Point Software

听

听

鈥淭his is not a new problem, and it is a very real one. We know that diversity drives growth and profit, yet 95-97% of developers in the AI space are men. That means unconscious bias is being baked into the technology from the ground up. It鈥檚 not intentional, but it is happening. We only need to look at the data. The lack of women, and particularly senior women, in tech costs the UK economy somewhere between 拢2.5 and 拢3 billion per year. We still have an 88% gender pay gap in the UK, meaning women are earning less, and yet when it comes to accessing AI education, upskilling courses and tools, the cost is exactly the same. There is no dispensation, no adjustment made to try to rebalance that dynamic. That sends a message.

鈥淭he danger here is significant. As reports like those from the LSE and others have shown, when AI is deployed with unintentional bias embedded in its outputs, it doesn鈥檛 just disadvantage women, it actively harms society. And with the current political backlash against DEI initiatives, and some of the more populist movements pushing hard away from a diversity agenda, this is probably going to get worse before it gets better. The cost of inaction won鈥檛 just be felt in boardrooms. It will also show up in taxation, in public services, and in the broader burden on society. We cannot afford to be so dazzled by what AI can do that we ignore who it鈥檚 being built for.鈥�

听

Sayali Patil, AI Infrastructure Reliability Specialist听

听

听

鈥淚 have spent several years building the kind of enterprise AI infrastructure that quietly shapes how millions of people experience technology every day. And what I can tell you from that vantage point is this: the gender gap in AI is not a cultural problem waiting for a cultural solution. It is an engineering problem already embedded in production systems, and it is getting harder to fix with every model that ships.

鈥淗ere is what that looks like in practice.

鈥淚ntent classification, the core mechanism that determines how an AI system understands what a person is asking, is shaped by the data it is trained on and the assumptions of the people who designed it. When those people are overwhelmingly male, the system learns to recognize and prioritize patterns of communication that reflect male experience. Not because anyone decided to discriminate. Because nobody in the room knew what was missing.

鈥淭hat gap does not show up in a product demo. It shows up six months after deployment, in support tickets, in satisfaction scores, in the quiet frustration of users who feel like the system never quite understands them. By that point the model is already in production. The assumptions are already downstream. Rolling them back is not a checkbox exercise. It is an architectural undertaking.

鈥淚 hold a USPTO granted patent in intent based chaos engineering (US12242370B2), which is fundamentally about how AI systems behave when their assumptions about the world do not match reality. The gender gap in AI development is exactly that kind of assumption failure, playing out at civilisational scale.

鈥淭he models being trained today are not products. They are infrastructure. Infrastructure lasts decades. And we are building it right now, in rooms where one perspective is dramatically overrepresented, at the exact moment when course correction is still possible.

鈥淭hat is the story worth telling.鈥�

听

Syed Asif Ali, Founder and Digital Identity Architect at Point Media

听

听

鈥淚 ran into this last year while testing an AI hiring tool in dubai. It kept downgrading some female candidates and I couldn鈥檛 figure out why. Skills were solid. Experience was fine. The only difference was how they wrote 鈥� less aggressive, less 鈥淚 did this, I led that鈥� kind of tone.

鈥淲hen we looked closer, it made sense. The model had been trained on data where that more direct style was treated as a signal of confidence. So anything outside that just鈥� looked weaker to it.

鈥淣othing was technically wrong. But the output still felt off.鈥�

鈥淭hat鈥檚 the part people miss. Bias in AI isn鈥檛 always loud or obvious. Sometimes it鈥檚 just one style quietly becoming the default, and everything else getting pushed down without anyone noticing.

鈥淚f the people building these systems all come from similar backgrounds, the definition of 鈥渘ormal鈥� gets very narrow. And once that鈥檚 baked into the system, it scales fast.鈥�

听

Michael Ferrara,听Technology Contributer and Legal IT Practitioner at Conceptual Technology

听

听

鈥淎I is becoming part of everyday life, so it matters who helps create it. Right now, many AI teams are still led mostly by men. That can create problems, even when no one means harm. People usually build products based on what they know, and they may miss challenges others face.

鈥淲e have seen examples of this before. Some AI image tools once showed mostly men when asked for pictures of business leaders or entrepreneurs. Now some tools seem to push harder for balance by showing more women in top jobs. That is interesting, but changing pictures alone does not solve the bigger issue.

鈥淩eal progress happens when we see more women that are involved in building the technology itself. That includes writing code, testing products, leading teams, and making final decisions. Different backgrounds bring different ideas and help spot mistakes sooner.

鈥淚f AI can help shape jobs, schools, healthcare, and business, then it should be built with many voices at the table, not only a few.鈥�

听

Edward Tian,听CEO of GPTZero

听

听

鈥淔rom my experience developing GPTZero, I have learned that AI bias appears in many inconspicuous forms. It includes what the model considers 鈥渘ormal,鈥� the types of items flagged for anomaly detection, and those that are omitted from consideration. If mostly male data sources are used to create training data or products, then it is likely that women will be impacted by AI in a different way than men. This is particularly frequent in an employment context (e.g., hiring, evaluating performance, and moderation).

鈥淔rom our detection research, we also see examples where models operate differently based on writing style, tone, and method of communicating. This is important because language is created by societal and gender norms, so if the design and testing personnel represent fewer than complete demographic diversity, then they will probably not see the failure modes until they have caused widespread damage as more people use them.

鈥淭he answer to this problem is not only 鈥渕ore women in AI,鈥� but is also the development of stronger evaluation and testing practices that account for disparate impact based upon demographic groupings and real world contexts.

鈥淭he key takeaway is that AI will reflect its creators unless there is pressure to make them reflect all groups of people it serves.鈥�

听

Emma Irwin, Director of Sales Engineering at Dataiku

听

听

鈥淭o guarantee AI success at scale, in a way that is trusted and mitigates bias, we must first combat the issue of female representation in AI. If AI models are shaped by the viewpoints of the engineers that build them, how can we avoid bias when an AI engineering team is made up entirely of men? And if that is the case, what can we do to ensure the decisions they shape are representative of the female half of the population?

鈥淭he data that AI models are trained on must be representative of a range of diverse demographics to avoid in-built bias 鈥� including a balance of female contributions. Having more varied perspectives shaping AI will improve the outputs from both an equality standpoint as well as the quality of AI outputs overall. Embedding inclusion in the AI development lifecycle will also reduce blind spots and help create AI that is more credible and widely accepted.

鈥淯ltimately, not only will having more female voices present drive more inclusive and impactful innovation, but it will also improve the quality of AI models overall, while strengthening trust, credibility and creativity.鈥�

听

Faye Ellis, Principal Training Architect, Pluralsight

听

听

鈥淗istorically, women are overrepresented in roles like coordination, documentation and support 鈥� exactly the tasks which AI is most likely to automate or devalue. Unless roles are intentionally redesigned, women face higher risks of redundancies than men do.

鈥淭his reflects the fact that AI 鈥� largely developed by men 鈥� cannot effectively serve the needs of both men and women in the workplace. Roles must be redesigned so that women can move into work involving judgement, oversight, decision-making work.

鈥淲omen must also be present for decisions across the entire AI lifecycle, including product, architecture, governance and procurement.

鈥淲omen are less likely to have time to learn outside of work hours than men, and learning how to use AI is no exception. Organisations must build dedicated time to learn how to use AI models at work into the working day, otherwise women will fall behind in their knowledge compared with men. AI capacity must be built into roles, not bolted on, otherwise the gender gap will inevitably widen.鈥�

听

Kristyna Vlckova,听VP of Growth Marketing at GoodData

听

听

鈥淚 think there鈥檚 a clear difference in how men and women adopt AI. Men often integrate it across both personal and professional life, whereas women tend to separate the two, creating a psychological barrier. Helping teams see AI as a tool for both spheres can make adoption much smoother.

鈥淥ne of the biggest challenges I see is helping women feel confident experimenting with AI. Often, they only feel ready when they鈥檙e overqualified, which is why creating a supportive environment where people can learn at their own pace is so important. It鈥檚 not about ability 鈥� it鈥檚 about psychological safety.鈥�

听

Post-publishing note: Due to a significant volume of feedback received on this topic, a part two will be published in the coming days.