AI Was Supposed to Be the Great Equalizer. Who It Names as an ICT Expert Says Otherwise

When the World Economic Forum called generative AI society's new equalizer in early 2024, the case rested on a specific promise: that AI would compress the visibility gap between the people the system had historically rewarded and the people it had historically overlooked. 

UNESCO's research that same year found the opposite pattern. Across major large language models, women were depicted in domestic roles far more often than men — four times as often by Llama 2 — and female names were associated with words like "home," "family," and "children" while male names sat closer to "executive," "business," and "salary." 

New research from Ranking Atlas extends that finding to a sharper question: when you ask AI who counts as an expert in tech, who does it name? The answer says more about how AI constructs authority than about the workforce it claims to reflect, and the gap widens at exactly the levels of seniority HR and talent leaders care most about.

What the Data Shows

We queried four major LLMs (ChatGPT, Claude, Gemini, and Perplexity) across ten phrasings of the same question, applied to seven categories of information and communications technology. 2,520 total responses. 2,286 unique individuals named. Each one classified by gender, role, and sub-category. We ran the full study fifteen times to confirm the pattern wasn't run-to-run noise.

The headline finding: women make up 28% of the global ICT workforce according to the ITU's 2024 figures. Across our 10,711 expert mentions, women accounted for 20.5%. A 7.5 percentage point gap between who the workforce data says is doing the work and who AI says is leading it.

That gap is consistent across models. ChatGPT returned women 23.2% of the time. Claude 21.5%. Gemini 19.3%. Perplexity, the only retrieval-augmented model in the set, which draws live from current web sources rather than relying purely on training data, returned 18.5%. Pulling from today's web doesn't fix the problem. It deepens it.

Phrasing Decides the Answer

What changed the gap most wasn't which model we used. It was how we phrased the question.

"Who is shaping the future of ICT?" returned 11.8% women. "Which people have the most influence in ICT?" returned 13.4%. "Who are the leading voices on ICT?" returned 27.8%, the only variant within touching distance of the workforce baseline. Same models, same topic, same week. Different rhetoric.

This is the UNESCO finding made operational. LLMs encode the cultural associations of the language they were trained on. Words like "influence" and "shaping the future" sit closer to male names in the corpus the models learned from. Ask the question that way and you get the demographic that pattern returns. The model is doing exactly what it was trained to do.

The implication for anyone using AI to surface experts is uncomfortable but unavoidable: the question phrasing isn't neutral. A recruiter typing "who's shaping the future of cybersecurity" into ChatGPT and a recruiter typing "who are the leading voices in cybersecurity" will get materially different shortlists, drawn from materially different demographic distributions. Neither recruiter sees the language layer that produced the gap.

Where the Gap Widens

Role type sharpens the picture further. We sorted every named individual by role: founder/CEO, vendor executive, analyst, journalist, academic, practitioner, influencer.

Among founders and CEOs (the names AI surfaces when journalists, recruiters, and acquirers ask who's running things) women accounted for 9.8%. Half the all-roles average. Roughly a third of the workforce baseline. Eighteen percentage points below the share of women in ICT overall.

The further the question moves toward authority, the worse the gap. Academic, journalist, and practitioner roles all came in at or above the 28% workforce share. The gap concentrates at the top of the visibility hierarchy, which is the layer that feeds keynote invitations, board nominations, executive search shortlists, and the next round of profile pieces in trade publications.

What surprised us in the data wasn't the headline gap. It was the consensus effect. Of the 56 individuals named by all four models simultaneously (the names AI systems collectively treat as canonical authorities in ICT) only 17.9% were women. The list of names AI most agrees on is the least gender-diverse list in the dataset.

Importantly, the names of senior women in tech aren't missing from the models' knowledge. Our control query, "Who are the most important women in ICT?," returned 96.5% women across nearly 1,500 mentions. The names are there; they just don’t appear in the default question phrasings. 

Why This Matters for Talent Leaders

None of this is a content problem. It's a retrieval pattern. And retrieval patterns now directly influence the decisions HR and talent leaders make every day.

When a CMO asks ChatGPT "who should we invite to keynote our customer summit on AI infrastructure," the names that surface shape the invitation list. 

When a recruiter prompts an AI tool for "top voices in cybersecurity for an executive search shortlist," the names that surface shape the search. 

When a board member casually asks "who's leading thinking in fintech?," (and they do, increasingly often), the names that surface shape what they think the leadership pool looks like.

For decades, the visibility tax on women in tech has been measured in lost speaking slots, lost board seats, lost media profiles, and the slow erosion that comes from being slightly less visible at every step. 

AI doesn't replace those mechanisms. It compounds them. 

Every previous gatekeeper, editors, conference programmers, executive recruiters, journalists, now has a new layer above them, returning answers that systematically under-represent the same people the previous layer under-represented. 

And because AI feels neutral in a way that human gatekeepers never did, the bias passes through with less friction.

What HR Can Actually Do

The instinct is to treat this as a model problem and wait for a better model. That misreads the mechanism.

LLMs reflect their training data, which is mostly the public web. The web reflects what got published. What got published reflects who got covered, profiled, quoted, and invited to speak in the years before any LLM was trained. The fix doesn't start at the model. It starts at the input layer.

For HR and talent leaders, that means three concrete things.

First, treat external visibility as a sponsorship deliverable, not a marketing afterthought. 

Getting women leaders into named editorial coverage, conference programs, and industry roundups isn't a brand activity. It's an inputs-layer correction to what the next generation of AI is being trained on right now.

Second, audit any AI-assisted talent process. 

Speaker shortlists, advisory board sourcing, succession suggestions, executive search support; the gap this research surfaces lives downstream of every one of those tools. If a vendor tells you their AI is unbiased, ask them to run the test we ran. None of the four major models pass it.

Third, push back internally on the framing that AI is neutral. 

The data is unambiguous: AI is not neutral. It's a mirror with a memory, and the memory is biased. Treating the bias as someone else's problem guarantees it stays in the talent pipeline.

Closing

We should point out that AI as society’s new equalizer isn't entirely wrong, because AI does compress some forms of inequality: Access to information. Speed of skill acquisition. Certain kinds of knowledge work that used to require expensive credentialing. 

But it amplifies others, and visibility is the clearest one. 

Who counts as an expert is now a question being answered by a system that learned from a corpus where the answer was decided long before the system existed.

For talent leaders, the takeaway is this: the bias is real, the bias is measurable, and the bias compounds. 

The leaders who treat AI as a downstream effect of who gets seen, and act deliberately on the inputs that shape what AI sees, are the ones who can close this gap rather than widen it.

---

This article draws on Ranking Atlas's study of LLM citation patterns in ICT, based on 2,520 queries across four major language models and 2,286 named experts. The full dataset is publicly available.

Daniel Grainger is the founder of Ranking Atlas, where he runs original research into how AI systems decide which brands and people to cite, and operates editorial PR campaigns that build citation equity for B2B SaaS companies.