Manspl-AI-ning: The evolution of AI and its effects on women

by Bonny Chung

Over the past few years, Artifical Intelligence (AI) has taken over our screens and mind’s eye at lightning speed. It’s here and it’s not going anywhere anytime soon! But what do we really know about it? Researchers from Remi AI App Ltd (UK) have been looking into the effects of AI on marginalised groups. This article is focused specifically on women.  

Gender Biases in AI 

We first started looking at biases, which often start when the AI is being trained, that is to say when the data is being fed in to the AI model. AI takes the biases, already baked into its training data, turns them into mathematics, and applies them consistently and at scale. Bias is scaled to everyone, everywhere, all the time. A human recruiter might have a bad day and make one biased decision. An AI system makes that same biased decision millions of times, without hesitation, and calls it objectivity. 

As AI quietly migrates into recruitment, performance management, and the tools women use every day at work, the evidence is mounting that gender bias is being automated, amplified, and normalised. The consequences can be economic — eroding women’s professional trajectories and pay — and psychological, compromising their confidence and agency in the workplace. The Fawcett Society has identified this automated discrimination as a significant and under-recognised threat to the economic equality that allows women the ability to live independent lives. 

A male-authored internet 

A 2023 study by Pitt Cyber, “Who Authors the Internet?”, analysed the publicly disclosed training resource behind GPT-3 e.g. Wikipedia, and found that just 26.5% of the training data was written by women. This tells us that the “voice” of the foundational internet knowledge base is notably male-dominated”.

“The ‘voice’ of the foundational internet knowledge base is notably male-dominated.” 

If the knowledge base is male, the worldview is male. Word embeddings, the way AI represents meaning place “strategic” and “analytical” closer to male pronouns, “empathetic” and “supportive” closer to female ones. Even the latest models, given identical candidate profiles, still describe women in emotional terms and men in strategic ones. 

The problem runs deeper. Stripping out names and pronouns, the “gender scrubbing” approach, does not work. Other details on a CV, from extracurricular activities to employment gaps to word choice, act as hidden proxies. The AI reconstructs gender from the remnants. Research with synthetic datasets like FairCVdb confirms it: scrubbing reduces leakage but never eliminates it. 

AI systems learn from past performance ratings and promotion decisions, ratings produced by humans who had their own biases. If women historically scored lower or were passed over, the AI treats those outcomes as ground truth. It does not know the labels were wrong. It just replicates them. 

And the tendency is that the system feeds on itself. A biased AI screens out diverse candidates. The homogeneous workforce that results generates the data that trains the next model, which becomes more confident in the same preferences. When qualified women notice the pattern, they stop applying. The pool gets less diverse. The cycle consolidates. 

The tools women can’t escape 

Beyond hiring, AI-powered monitoring tools “bossware” erode digital autonomy, and this disproportionately affects women, whose caregiving responsibilities more often disrupt standard work schedules. The metrics these systems generate can encode the same historical biases, automating discrimination at every stage of the employment lifecycle. 

Women are also more likely than men to enter e-meetings off-camera and muted, surrendering visual presence entirely. Add to this the tendency of men to speak over women during meetings, the possibility of glitches, cameras not working, plus a host of other technical difficulties and we women are left feeling deflated, unseen and unheard. 

When the way you ask changes the answer 

The way we talk to AI, prompt engineering, is not a neutral technical skill. It is a sociolinguistic exercise, and research shows that men and women prompt differently. Female-authored prompts tend to be more indirect, contextual, and involved. Male-authored prompts are more direct, brief, and transactional. 

If a model has been fine-tuned through Reinforcement Learning from Human Feedback by a predominantly male cohort of raters, it may respond best to command-line-style imperatives and stumble on the conversational, context-rich prompts women tend to write. The model’s performance degrades, and women are left assuming the problem is theirs. It is not. Digital upskilling must recognise these differences, not paper over them with a one-size-fits-all approach.

A plausible solution 

Currently, all the leading models come from America, closely followed by China, and Europe’s Mistral project falls far behind on benchmarks. The consolidation of power by a few American AI companies means that there is a very slim chance of fighting against the internalised bias. However, a world with many digital intelligences provides hope. The more democratised AI model training becomes, the more chances we have to train egalitarian models fairly and correctly.

How do we incentivise plurality rather than a monopoly?

Under the EU AI Act, AI used to advertise, filter, shortlist or score job applicants is classified high-risk. This brings written duties: human oversight, logging, data quality, and communication between recruiter and applicant. Emotion recognition in hiring is prohibited, and since February 2025 the law has also required that people operating AI understand what they are operating. 

How can Remi App help