One Woman for Every Ten: How AI Names the Experts in Tech
A new study shows that across the four largest commercial language models, women made up 20.5% of named ICT experts against a 28% workforce baseline, and changing only the wording of the question moved that figure from 11.8% to 27.8%.
Content Written By:
Daniel Grainger
Founder, Ranking Atlas
Published April 2026
Key Takeaways
- Ask AI for top tech founders and it names one woman for roughly every ten men. The female share for founders and CEOs was 9.8% across 2,802 mentions, against a 28% workforce baseline.
- Five of six ICT sub-categories underrepresented women relative to their own workforce figures. Software engineering showed the widest gap: 15.3% named, against a 22% workforce share.
- One wording change nearly doubled how many women AI named. Across the nine non-control variants, query phrasing alone moved the female share from 11.8% to 27.8%, with the models, topics, and categories held constant.
- The more the models agreed, the more male the answer got. Of the 56 individuals named by all four models, fewer than one in five were women. Among names returned by only one model, it was closer to one in four.
- The AI that searches the live web did worst. Perplexity, the only retrieval-augmented model, returned 18.5% female, below the three trained models' 21.0% average.
- Across the whole study, women were 20.5% of named ICT experts against the 28% workforce baseline: a 7.5 percentage point gap, measured across 10,711 name mentions.
- The "most influence" phrasing returned just 13.4% female across 1,772 mentions: the second-lowest of any variant, on the largest sample among the low-scoring variants.
- The models can return female experts when prompted to. The control query "who are the most important women in ICT?" returned 96.5% female across 1,478 mentions, indicating the names are present in the sources the models draw on but are not surfaced by standard expertise prompts.
Ask AI for Top Tech Founders and It Names One Woman for Every Ten Men
| Role | Female % | Mentions | Female | Male |
|---|---|---|---|---|
| Founder / CEO | 9.8% | 2,802 | 275 | 2,527 |
| Analyst | 20.8% | 2,130 | 442 | 1,688 |
| Vendor executive | 20.9% | 2,296 | 480 | 1,816 |
| Academic | 27.3% | 1,228 | 335 | 893 |
| Journalist | 28.5% | 421 | 120 | 301 |
| Practitioner | 29.0% | 1,048 | 304 | 744 |
| Influencer | 30.0% | 751 | 225 | 526 |
Ask the four largest AI models who the leading founders and CEOs in ICT are, and the names that come back are 9.8% women. That is 275 women against 2,527 men across 2,802 mentions. Roughly one woman named for every ten men.
It is the widest gap in the study. The founder and CEO rate sits 18 percentage points below the 28% workforce baseline, and it is less than half the all-roles rate of 20.5%.
The picture changes by role. Academic, journalist, practitioner, and influencer roles all reached or passed the workforce baseline. Roles further from executive or technical authority track more closely with their workforce share, while the named pool narrows toward men at the upper end of the role ladder.
In Five of Six Tech Fields, AI Underrepresents the Women Who Are There
| Sub-category | LLM female % | Workforce % | Gap | Source |
|---|---|---|---|---|
| Networking & telecommunications | 13.6% | 19% | +5.4 pp | Eurostat / ITU composite |
| Software engineering | 15.3% | 22% | +6.7 pp | BLS 2025 / WomenHack |
| AI and machine learning | 18.1% | 22% | +3.9 pp | Stanford AI Index 2024 / Interface EU |
| Cybersecurity | 18.6% | 22% | +3.4 pp | ISC2 Cybersecurity Workforce Study 2024 |
| Cloud and infrastructure | 20.0% | 14% | −6.0 pp | Statista 2025 / WomenHack |
| Data and analytics | 25.0% | 30% | +5.0 pp | World Economic Forum 2025 |
Each tech field has its own workforce baseline, so a figure that looks low against the overall 28% mark may sit above or below its own sector's share. Measured that way, five of six sub-categories underrepresented women.
Software engineering had the widest shortfall: 15.3% named, against a Bureau of Labor Statistics workforce figure of 22%. Networking and telecommunications, AI and machine learning, cybersecurity, and data and analytics all came in below their sector baselines too.
One field ran the other way. In cloud and infrastructure, the models named women 20.0% of the time against a workforce share of just 14%: the single sub-category where AI named more women than the field actually employs.
One Wording Change Nearly Doubled How Many Women AI Named
| Query | Female % | n (mentions) |
|---|---|---|
| Who is shaping the future of ICT? | 11.8% | 119 |
| Which people have the most influence in ICT? | 13.4% | 1,772 |
| Who writes the most influential content on ICT? | 19.6% | 861 |
| Name the most-cited commentators on ICT. | 19.8% | 1,214 |
| Who are the top experts in ICT? | 20.2% | 1,438 |
| Who should I follow for ICT analysis? | 21.1% | 1,010 |
| Who are the go-to analysts in ICT? | 21.8% | 1,327 |
| Name some thought leaders in ICT. | 22.1% | 1,418 |
| Who are the leading voices on ICT? | 27.8% | 1,552 |
| Who are the most important women in ICT? (control, excluded from headline figures) | 96.5% | 1,478 |
The models did not give a stable answer. They gave a different answer depending on how the question was worded.
Across the nine non-control variants, the female share ran from 11.8% to 27.8%. The variation came from phrasing alone: the four models, the topic, and the seven categories were held constant, and the wording change nearly doubled the result.
Six of the nine variants came back below the 28% workforce baseline. "Who are the leading voices on ICT?" reached 27.8%, just touching it. "Who is shaping the future of ICT?" returned 11.8%, the lowest in the set. The most reliable single signal is the "most influence" variant: 13.4% female on 1,772 mentions, a low share on the largest sample of any low-scoring variant.
The More AI Models Agree, the More Male the Answer Gets
| Named by | Unique people | Female | Male | Female % |
|---|---|---|---|---|
| All 4 models | 56 | 10 | 46 | 17.9% |
| 3 models | 64 | 15 | 49 | 23.4% |
| 2 models | 95 | 21 | 74 | 22.1% |
| 1 model only | 326 | 85 | 241 | 26.1% |
The names the four models agreed on were the least diverse names in the dataset.
Of the 56 individuals every model named, fewer than one in five were women: 10 women against 46 men, 10 percentage points under the workforce baseline. Drop to names returned by just one model, and the share rises to roughly one in four, close to the workforce figure.
The pattern is consistent across every tier. Where the four models converge on the same names, the share of women in the named pool is at its lowest in the study. The names these systems collectively treat as canonical authorities in ICT skew the most heavily male of any group in the dataset.
The AI That Searches the Live Web Was the Worst of the Four
All four models named fewer women than the 28% workforce baseline. The three trained models came in higher: OpenAI at 23.2%, Anthropic at 21.5%, Google Gemini at 19.3%, an average of 21.0% across 8,640 mentions.
Perplexity returned the lowest share of the four: 18.5% across 2,071 mentions. It is the only retrieval-augmented model in the panel, pulling from live web sources rather than training data alone, and access to the open web widened the gap rather than closing it. The body of authoritative ICT content available to retrieve is at least as male-dominated as the data the other models were trained on, and whoever holds citation equity inside that web shows up disproportionately in retrieval-augmented outputs.
AI Names Far Fewer Women as ICT Experts Than Actually Work in ICT
Step back from the individual cuts and the overall picture is one number against another. The ICT workforce is 28% female. Across 2,520 model responses, the four largest AI systems named women as experts 20.5% of the time. A gap of 7.5 percentage points, measured across 10,711 name mentions, control query excluded.
Including the control query, the female share across all 10 variants rises to 29.7% across 12,189 mentions. That single query lifts the figure by 9.2 percentage points on its own, which is why every headline number in this study leaves it out.
This is not the models failing to know the names. It is the models not returning them unless asked in a particular way. The next section is the proof of that.
Why The Default Question Returns So Few Women
One query settles what kind of gap this is. The control variant, "who are the most important women in ICT?", returned a female share of 96.5% across 1,478 mentions on the same four models, with no other change to the methodology.
The expertise is present in the sources the models draw on, and the models can return it when the prompt asks for it. What standard expertise prompts surface is a narrower subset of that pool, shaped by which names dominate authority-language content in the underlying data.
That puts the headline figure in context. With 20.5% of named experts female under neutral phrasing and 96.5% under direct phrasing, the 7.5-point gap from the workforce baseline reflects which names the default question retrieves rather than which experts work in the field.
“The biases that shape who gets seen in society at large are present in large language models too. Unless we make ourselves aware of that, the answers these systems return by default will keep reproducing it.”
Daniel Grainger, Ranking Atlas
The expertise is not in question. The control query settles that. What the data raises is a narrower and harder question: what does it take for a name to be surfaced by the phrasing people actually use, and who gets left out every time it is not.
Methodology
What we asked
Ten query variants, each applied to seven ICT categories (one top-level, six sub-categories):
- Who are the top experts in {category}?
- Who are the leading voices on {category}?
- Who should I follow for {category} analysis?
- Name the most-cited commentators on {category}.
- Who are the go-to analysts in {category}?
- Who writes the most influential content on {category}?
- Who are the most important women in {category}? (control, excluded from headline figures)
- Name some thought leaders in {category}.
- Who is shaping the future of {category}?
- Which people have the most influence in {category}?
Who we asked
Four AI systems, each queried independently with identical prompts: OpenAI ChatGPT-4o, Anthropic Claude Sonnet, Google Gemini 2.5 Flash, and Perplexity Sonar Pro.
How many times
2,520 total queries: 630 per model. The ICT top-level category received 15 runs per variant. Each of the six sub-categories received 8 runs per variant. No model refused any query across all 2,520 requests (0.0% refusal rate across all four systems).
Name extraction
Raw responses were processed by a second-pass extraction model (Claude Haiku) instructed to identify every named human expert in each response. The extraction prompt explicitly excluded company names, product names, and book titles. Near-duplicate name variants (for example "Andrew Ng" and "Andrew Y. Ng") were merged using fuzzy string matching with a similarity threshold of 88 or above. Ambiguous merge decisions were reviewed manually.
Gender classification
Classification used a three-tier cascade applied to 2,286 unique individuals:
| Tier | Method | Names classified | Share |
|---|---|---|---|
| 1 | Name-based inference (gender_guesser library) | 1,842 | 80.6% |
| 2 | LLM-assisted classification (Claude Opus, high-confidence only) | 284 | 12.4% |
| 3 | Manual human review with public sources | 164 | 7.2% |
Gender was classified from names and professional bios, not self-identification. Classification reflects binary gender assignment (M/F) based on publicly available information. Non-binary identities are not captured. A further 7.0% of individuals (160) could not be classified and are excluded from percentage calculations.
How mentions are counted
Headline figures count how often women are named, not how many unique women appear. If one woman is named 50 times and one man is named 50 times, that's 50/50.
Exclusions
Variant 7 ("who are the most important women") is excluded from all headline figures and reported separately as a control condition. 16 individuals were removed during manual review: entries that turned out to be channels, pseudonyms, combined entries, or names too ambiguous to classify.
Workforce baselines
| Sub-category | Female % | Source |
|---|---|---|
| ICT (overall) | 28% | Women in Tech Network / CompTIA State of the Tech Workforce 2024 |
| AI and machine learning | 22% | Stanford AI Index 2024 / Interface EU |
| Cybersecurity | 22% | ISC2 Cybersecurity Workforce Study 2024 |
| Software engineering | 22% | BLS 2025 / WomenHack |
| Data and analytics | 30% | World Economic Forum 2025 |
| Cloud and infrastructure | 14% | Statista 2025 / WomenHack |
| Networking and telecommunications | 19% | Eurostat / ITU composite |
These patterns reflect how AI systems prioritise sources and citations when generating answers.
Limitations
Query phrasing shapes results. The 10 variants in this study are illustrative, not exhaustive. A different set of framings would produce different numbers.
Gender classification used name-based inference and publicly available information, not self-identification. Non-binary individuals are not captured, and classification errors are possible, particularly for names uncommon in Western naming conventions.
LLM responses are probabilistic. Running the same queries on the same models at a different time would produce some variation. Results reflect the models as they operated in April 2026.
Asking "who are the top experts" favours people who are already well-known. That pool skews male for historical reasons. This study measures what LLMs return within that frame.
Workforce baselines vary by source and methodology. The figures used here are composites drawn from multiple studies and should be treated as approximate benchmarks, not precise measurements.
We ran the full study 15 times. The headline number landed between 15.6% and 22.9% every time, with a standard deviation of 1.8 points. The gap from the workforce baseline (7.5 points) is four times larger than the run-to-run variation. The finding holds.
Download the Data
The full dataset: 2,286 individuals with mention counts, gender classification, role type, sub-category breakdown, and cross-model consensus tier.
ai-ict-experts-dataset.csv
2,286 individuals · gender, role, category, mention counts, consensus tier
Citation: Ranking Atlas, "One Woman for Every Ten: How AI Names the Experts in Tech," April 2026. Research by Daniel Grainger.
For a different cut of this dataset, additional category or model breakouts, or methodology questions, contact contact@ranking-atlas.com.
Related Research
How Brands Get Cited By AI: The 2026 Guide To Citation Equity
This guide explains how AI answer engines choose which brands to cite, and what produces the editorial pattern that puts you in the answer.
Welcome to the SaaSpocalypse
This data report shows how vibe-codeable SaaS categories lost all pricing power while AI-resistant tools raised entry-tier prices three times faster than inflation.
The Hidden Economy Behind Google Rankings
This data report analyses 140,000 publisher listings to show how paid link markets expanded after repeated Google crackdowns.
Ranking Atlas is a specialist data campaigns firm. We turn proprietary data, commissioned research, and original analysis into stories and studies that earn authoritative coverage.
For methodology questions, additional data cuts, or research enquiries: contact@ranking-atlas.com.