The AI Training Gap Is a Career Opportunity Gap
Washington Post, 7 November 2026: uneven support to learn AI is not proof of lower ability. A license is not paid practice time.
Christian Kosiek•Published •17 min read
A software license does not teach anyone. In a small team, three people can share the same login. Two can try a workflow while the inbox is quiet. The third spends the day on urgent requests and has no paid hour left for a bad first attempt. Months later, a manager may trust the person who has a folder of experiments. That choice can look like merit. Part of the folder is spare time.
That scene is an illustration, not a case. The public record behind it is narrower. It does not show that women are less able to use the tools. It shows that use, confidence, support and employment growth are different measures, and that a posting which asks for judgment is not a training record.
Support is not the same thing as ability
On 7 November 2026, The Washington Post published "Men see more career advancement in jobs with AI, as women lose ground," by Julia Carpenter. A fetch of the Post page returned the date, the byline and the opening, then stopped. A republication on DNYUZ carries the same opening and says the post appeared first on The Washington Post. The sentences below follow that republication where it matches the opening the Post page itself returned.1
The Post says female workers tend to use AI tools less often than men, according to a recent analysis from the Shorenstein Center on Media, Politics and Public Policy at Harvard University. Women also say they have comparatively less access to AI training, mentorship and other workplace support. Nancy Gibbs, director of the center, is quoted on competence. The published sentence uses a contraction, so this line is a paraphrase: it is not that women are less competent with the tools. She said they are less confident, often because of a lack of support at work, and also because they trust the tools less overall. She added that the skepticism is a valuable part of the debate over safety, guardrails, value and use.1
That is an interview sentence. It is not a test score, and it is not a finding that a named employer discriminated.
What the Harvard survey actually counted
The report is Are Women Being Left Out of AI?, posted on 14 October 2026. The authors are Laura Manley, Allison Wan, Brian Friedberg, Nancy Gibbs and Matthew A. Baum. The survey is the Civic Health of Institutions Project (CHIP50): 5,403 U.S. adults in July 2026. On the usage chart, the unweighted sample is 2,606 women and 2,431 men. The report prints a 95% margin of error of 0.7 to 1.9 percentage points for women and 0.9 to 1.9 for men on that chart. Unless it says otherwise, it treats reported differences between women and men as statistically significant. Most respondents identified as women or men. The report says the sample is too small for reliable findings about people who identify outside that binary.2
In the July survey, women were more likely than men to say they have never used AI (29% versus 22%) and more likely to say they have tried it but rarely use it (28% versus 21%). Men were more likely to say they use it occasionally (27% versus 24%), regularly (25% versus 16%), or that they create workflows or automations (5% versus 3%). These are self-descriptions. They are not an audit of employer logs.2
Desktop browsing tells a different sentence about "ever used". In National Internet Observatory data from July 2024 to December 2025, covering 5,513 U.S. adults, the report finds no statistically significant difference between women and men in having used AI at least once. Among people who did use it, men spent a larger share of browsing time on AI tools than women (4.6% versus 2.9% across the full period). The report says that difference is significant across the full period, but not in every month. A survey answer and a browsing trace are not the same measure. A headline that collapses them into "women use AI less" drops that distinction.2
| What the report states | What it does not state |
|---|---|
| Survey, July 2026: never used AI, 29% of women and 22% of men; regular use, 16% and 25% | That those shares are employer training records |
| Browsing, July 2024 to December 2025: no significant gap in having ever used AI; 4.6% versus 2.9% of browsing time among users | That the time share is a measure of skill or of promotion |
| Employed women are less likely to report AI training, mentorship and encouragement | A percentage-point gap for that access. The chart figures did not come through as readable text |
On the workplace, the prose is qualitative where a percentage would be easy to invent. Among employed respondents, women are less likely than men to report access to AI training, mentorship and encouragement to use AI. The report says those supports are associated with greater use, greater confidence and greater perceived value among both women and men, and that gender gaps often persist even among people who report more of that support. Figures A19 to A21 carry the shares. Those percentages did not come through as readable text in the PDF extract used for this check, so this article does not print a training-access percentage.2
The same report says privacy concerns and distrust of outputs are the top two stated barriers for both women and men. Fear of judgement ranks near the bottom: 5% of women and 6% of men. The page spells a cited surname both as Craney and as Cranney. This article uses the 5% and the 6% as the report's own barrier shares. It does not treat them as a finding from a paper this pass did not open. The report says the work was funded by Pivotal, founded by Melinda French Gates. That is a funding line, not a method.2
Gibbs also told the Post that when women's voices are missing from spaces where decisions are made, there is less cause to trust those decisions. That is her argument. The survey does not test whether a particular promotion round was unfair.1
Time pressure is an interview, not a time diary
Solène Delecourt, assistant professor of management of organizations at the University of California, Berkeley, told the Post that the adoption gap is closing, but slowly. She said women still report feeling more apprehensive or cautious about bringing AI into their work, and that for the newest tools the vast majority of early adopters are men. As an anecdote, she said she hears women say they are tired of AI and that it removes some of the joy of the work. She also said women report feeling more time-constrained than men: a larger share of housework and caregiving at home, and invisible labor on top of the job. Her point about tinkering is specific. The first experiments waste time. Someone whose day is already full may not have a spare hour for a bad draft.1
Those sentences are an interview. This article did not open a time-diary study under her name. One remark about spare time is her inference, and the published line uses a contraction, so it is paraphrased here. Do not promote it into a measured hour count.
The Post then says it appears that men's employment has grown faster than women's in occupations with significant AI exposure. It names software development and financial management as examples and cites Pew Research Center. That sentence does not carry Pew's estimates, the education split, or Pew's warning that the link is not necessarily direct. Those figures and limits are in why the AI jobs gender gap is widest without a degree. A newspaper sentence is not a second study.1
A different survey asks who reports any training, by seniority rather than by gender. See why AI training reaches bosses more often than beginners. The two gaps can show up in the same workplace. They are not the same table.
A job advertisement is not a training record
The draft this article started from cited a PwC Cyprus page for the June 2026 AI Jobs Barometer. The global release is the cleaner source. On 15 June 2026, in London, PwC said the barometer analysed more than one billion job advertisements in 27 countries and territories. At entry level, based on 2.4 million US jobs, the roles most exposed to AI are described as seven times more likely to require traditionally senior-level human-intensive skills. The bullet says judgement and leadership. The paragraph says leadership, creativity or face-to-face interactions. Those roles grew 35% since 2019, while other entry-level roles shrank 10%.4
Pete Brown, global workforce leader at PwC, said: "The traditional relationship between experience and expertise is changing. AI is removing some of the routine work that once acted as an apprenticeship, while increasing demand for judgement, leadership and adaptability much earlier in careers." That is the firm's reading. It is not a count of who was given a paid hour to practice.4
A posting that asks for judgment does not show that the applicant was trained, promoted, or given comparable time. The limits on "seven times", including the report's own definition of a traditionally senior skill, are already set out in why junior advertisements now ask for senior skills. This article does not reuse the wage or productivity figures. The Cyprus page, which the source file linked, prints 53% versus 36% in one bullet and 52% versus 36% in a later paragraph for headcount, and 9% versus 8.6% for the wider jobs market. The global release's bullet uses 52% versus 36% and 9%. Those figures are not findings of this article. The mismatch is why the global release is the citation.5
Who is supposed to teach the judgment a posting now asks for is a separate question. See who teaches judgment when a tool writes the draft.
What the skills report asks employers to provide
Changing landscape of skills in the age of AI is an inter-agency synthesis. Ummuhan Bardak of the European Training Foundation wrote it, with contributions from a working group that included Cedefop, Eurofound, the European Commission, the ILO and UNESCO. The manuscript was completed in May 2026. The suggested citation is European Training Foundation (2026), Publications Office of the European Union, Luxembourg, ISBN 978-92-9157-764-4, doi 10.2816/7083675. An ILO catalogue page lists 13 August 2026. The ETF file sits on a September 2026 path. A date of 24 September, printed in the draft this article started from, was not on the pages opened here, so it is not used. The report says its contents are the authors' responsibility and do not necessarily reflect the views of the ETF or the EU institutions.3
The policy chapter treats AI literacy as a foundational skill: the ability to understand, use, monitor and critically reflect on AI systems. It describes that skill as an enabler of human agency and inclusion, not only as a technical trick. The report's sentence uses a dash, so this is a paraphrase. It also says that acquiring that literacy is not only an individual's responsibility. Businesses and public authorities, in the report's words, have a shared obligation to provide adequate training, time and support. The report argues for broader human capabilities rather than a narrow skill set. None of that is an experiment on promotion rates. It does not estimate how many paid hours change who gets the next project.3
What a small team can still test
The steps below are an editorial suggestion. Harvard, the Post, the ETF and PwC did not test them on an eight-person organization.
Start with opportunity, not enthusiasm. For each person, ask whether they have an approved tool, paid time, a task that actually needs the tool, somewhere to go when the output is wrong, and a manager who will look at the reasoning. Track that by role and by working pattern. Do not collect sensitive demographic data for a pilot unless there is already a lawful purpose and a governance arrangement. The official GDPR text did not open in this pass, so this article does not cite an article number for that caution. It is not legal advice.
A small amount of paid time for everyone, including people on fixed shifts, is more even than an optional evening webinar. One structured hour every two weeks is a suggestion, not a researched optimum. Pair it with one real task: what a good check looks like, what must not be pasted into a tool, and what a person must verify. Rotate who gets a mentor. An informal invitation tends to reach the people already in the room.
Judge the check, not the number of prompts. A grant officer can compare a generated summary with the funding criteria, name what is missing, and say why the decision stands. A customer-service employee can test a draft against policy and against accessibility. The valued skill is the judgment. Speed is not a substitute.
Before naming an AI champion, ask three questions. Who had paid time? Who received feedback on a mistake? Who could show the work under comparable conditions? If the answers differ sharply by role, the organization is not yet looking at the same opportunity. If someone declines an optional experiment, do not read that as resistance. The constraint may be the schedule.
A counterargument is fair. Some people will use evenings and move faster. A tool can also open practice to someone who never had a specialist colleague. The outcome is not fixed by the software. It depends on whether practice is part of the paid job.
The legal duty is narrower than a fair schedule
This article is not legal advice. A paid learning hour is not, by itself, compliance with the EU AI Act.
Article 4, in the text the Commission service desk shows as amended, says providers and deployers shall take measures to support the AI literacy of their staff and of other people who operate AI systems on their behalf. They take account of technical knowledge, experience, education and training, the context of use, and the people or groups on whom the systems are used. The same paragraph says this obligation does not require anyone to guarantee a specific level of AI literacy for any individual. The service desk says the wording reflects the EUR-Lex consolidated text as at 27 July 2026, including Digital Omnibus amendments. The summary on that page is not legally binding. A shared hour may support literacy. Completing it is not proof the obligation is met, and the article does not require equal career outcomes.6
Annex III, point 4(b), lists AI systems intended to make decisions affecting the terms of a work relationship, promotion or termination, to allocate tasks based on individual behaviour or personal traits, or to monitor and evaluate performance and behaviour. A record of who attended a training hour is not, by that description alone, such a system. An AI tool intended to score people for promotion, or to assign work from personal traits, is a different question. Classification depends on the intended purpose. The service desk says Annex III rules apply from 2 December 2027.7
Article 26 is conditional on a system that is high-risk. Paragraph 7 says that before putting a high-risk AI system into service or using it at the workplace, employers shall inform workers' representatives and the affected workers. Paragraph 11 says that, where Annex III high-risk systems make or assist decisions related to natural persons, deployers shall inform those persons. These duties are not a finding about any employer in the Post article. They are the questions to ask if a later tool is intended to decide who advances.8
How hilighter can help
hilighter does not run the Harvard survey, and it does not score who deserves the next project. It can make the practice under a role visible. Teams can map skills by role, compare people with those requirements in skills and talent insights, record a check or a correction in skills evidence and recognition, and set personalized growth paths that do not depend on whoever had a quiet afternoon. A manager's review of that evidence sits in manager assessments. The AI job impact view is indicative. It is not a Pew readout and it is not a prediction that a role will disappear. The Free plan covers up to three job roles. The number of employees is not capped.
Sources
- Julia Carpenter, The Washington Post, Men see more career advancement in jobs with AI, as women lose ground, 7 November 2026. The Post page returned the date, the byline and the opening, then stopped. The body used here follows the DNYUZ republication, which repeats that opening and says the post appeared first on The Washington Post. Nancy Gibbs, Shorenstein Center director, on confidence, support and trust (the competence sentence is paraphrased because the published line uses a contraction). Solène Delecourt, UC Berkeley, on a slowly closing adoption gap, early adopters, time pressure and tinkering (contractions paraphrased). Pew is cited in one sentence on employment growth, without Pew's estimates. Gibbs on whose voices are in the room.
- Shorenstein Center, Are Women Being Left Out of AI?, 14 October 2026, and the October 2026 PDF. Authors: Laura Manley, Allison Wan, Brian Friedberg, Nancy Gibbs, Matthew A. Baum. CHIP50 survey of 5,403 U.S. adults, July 2026. Usage chart: 2,606 women and 2,431 men; never used 29% versus 22%; tried but rarely 28% versus 21%; occasionally 27% of men versus 24% of women; regularly 25% versus 16%; workflows 5% versus 3%. Margin of error on that chart: 0.7 to 1.9 percentage points for women, 0.9 to 1.9 for men. Browsing, July 2024 to December 2025, 5,513 adults: no significant gap in having ever used AI; 4.6% versus 2.9% of browsing time among users, significant across the period but not every month. Employed women less likely to report training, mentorship and encouragement; chart percentages in figures A19 to A21 were not readable as text, so they are not printed. Fear of judgement: 5% of women and 6% of men. Surname spelled Craney and Cranney on the same page. Funded by Pivotal.
- European Training Foundation, Changing landscape of skills in the age of AI, and the PDF. Manuscript completed May 2026. ISBN 978-92-9157-764-4. doi 10.2816/7083675. Ummuhan Bardak, ETF, with the inter-agency working group (Cedefop, Eurofound, European Commission, ILO, UNESCO). Authors' views, not necessarily those of the ETF or the EU institutions. AI literacy as the ability to understand, use, monitor and critically reflect; a shared obligation, in the report's words, to provide training, time and support. Not an estimate of promotion effects. The ILO catalogue page lists 13 August 2026. A 24 September date in the source draft was not on these pages.
- PwC, AI reshapes global labour market into two distinct paths, press release, London, 15 June 2026. More than one billion job advertisements in 27 countries and territories. Entry-level analysis: 2.4 million US jobs. Seven times, with judgement and leadership in the bullet and leadership, creativity or face-to-face interactions in the paragraph. Senior-style roles +35% since 2019; other entry-level roles -10%. Pete Brown on apprenticeship and earlier demand for judgement. Headcount 52% versus 36% and the wider market at 9% appear in the bullet and are not used as findings here.
- PwC Cyprus, 2026 AI Jobs Barometer press page. The URL in the source draft. The same page prints 53% versus 36% in a bullet and 52% versus 36% in a later paragraph for headcount, and 9% versus 8.6% for the wider jobs market. Local restatement, not the methods report. Those figures are not findings of this article.
- European Commission, AI Act Service Desk, Article 4: AI literacy. Wording reflects the EUR-Lex consolidated version as at 27 July 2026, with Digital Omnibus amendments. The summary is not legally binding. Paragraph 1 does not require a guaranteed level of literacy for any individual.
- European Commission, AI Act Service Desk, Annex III, point 4(b): decisions on terms of work, promotion or termination; task allocation based on individual behaviour or personal traits; monitoring and evaluating performance and behaviour. When does enforcement start?: Annex III rules from 2 December 2027.
- European Commission, AI Act Service Desk, Article 26, paragraphs 7 and 11. Information duties for employers and for people subject to Annex III high-risk systems that make or assist decisions. Conditional on a high-risk system.
Sources last reviewed: 13 November 2026. The Post URL did not return the full article in this pass; the body follows the DNYUZ republication where it matches the opening the Post page did return. Training-access percentages in Shorenstein figures A19 to A21 were not readable as text and are not printed. The 24 September date for the ETF report was not on the pages opened. Delecourt's remarks are an interview, not a time-diary study. Pew estimates are not restated here; they live in the 27 October hilighter article. The Cyprus page disagrees with itself on figures this article does not adopt. EUR-Lex did not open for the GDPR text (403), so no GDPR article is quoted. EU dates follow the Commission service desk and are not legal advice. Neither an AI-detector score nor the absence of Google penalties can be guaranteed.