Silicon Valley has a phrase for people who miss the AI wave: the "permanent underclass". It usually describes a future in which machines take the jobs. Writing in the Guardian, Arielle Pardes reports that for women in tech, it is starting to look like the present.
The popular story is that AI levels the field, because a technology this new cannot reward old advantages. We think the opposite is happening. The AI boom is rebuilding tech's gender gap at higher speed, because a gold rush (hiring fast, through referrals, with the old counterweights cut) is exactly the condition under which existing patterns copy themselves. Companies that treat this as a culture-war sideshow are misreading a talent problem they created.
The numbers come from LinkedIn. Its study of companies in 27 countries, reported by HR Brew, found that women held 26% of new AI roles in 2025, against half of new non-AI jobs. AI postings have doubled since 2023 and, LinkedIn says, pay on average more than twice as much as roles that don't involve AI. Across AI occupations, men's median pay is $45,000 higher than women's.
A field with no history hires through old networks
This is the best argument for optimism, and it gets the mechanism backwards. When no credential is reliable, hiring managers under pressure do not suddenly assess everyone on the merits. They reach for the one signal they still trust: someone they know vouching for the candidate. Brenda Darden Wilkerson of AnitaB.org puts it plainly in the Guardian piece: companies are hiring at breakneck speed but finding people through "the same networks, the same referrals, the same filters that they've always used." In a field where most of the existing network is male, that is a closed loop running at full speed.
Pace does further filtering. Jayeeta Putatunda, an AI engineering lead at Turing, describes a culture where 12-hour days are not unusual, and recalls returning from four months of maternity leave to find completely different frameworks and models. She managed to catch up with supportive colleagues and a husband who equitably split childcare. Her diagnosis of everyone else is the sharpest line in the reporting: "They don't lack the interest or ability to keep up. They may lack the infrastructure that makes keeping up possible."
The youth of the field, then, makes the gap worse. Newness strips out the formal signals that structured hiring once relied on, and speed rewards anyone who never has to step away. When no credential means anything, the referral becomes the credential, and referrals run through the people already in the room.
When no credential means anything, the referral becomes the credential, and referrals run through the people already in the room.
The gap is widest where the power sits
If this were a pipeline issue, the gap would narrow as women who got in moved up. LinkedIn's data shows the reverse. HR Brew describes what the report calls a triple penalty: women are 10 percentage points scarcer in AI roles, 5 points scarcer at AI firms, and roughly 15 points scarcer in the C-suites of AI companies than at comparable non-AI firms. Women are 13.9% of CEOs at AI companies, against 19.1% elsewhere. LinkedIn's Sarah Steinberg told the Guardian that women in AI are concentrated in lower-paid roles such as data annotators, and the Guardian attributes part of the $45,000 gap to the types of jobs men and women hold.
That pattern points at promotion, and the best retention evidence available agrees, though it covers all of tech rather than AI specifically. The UK's Lovelace Report, by Oliver Wyman and WeAreTechWomen, estimates that between 40,000 and 60,000 women leave digital roles each year, with a quarter citing lack of opportunity to advance. Computing's analysis of the same report runs its section on why women leave under a pointed heading: it's not all about childcare. Any leader tempted to answer this with a better parental leave policy and call it done should take note. Leave matters; so does who gets handed the high-visibility AI project when they come back.
Put the pieces together and the gap is manufactured at a handful of decision points that sit squarely inside management's control:
- Who gets referred in, and whether anyone looks beyond the referral list when roles are filled in a hurry.
- Whose long hours get read as commitment, turning a caregiving constraint into a performance signal.
- Who gets brought back up to speed after an absence, and who is left to sink or swim against a year of model releases.
- Who inside the company is picked to move into AI work, and who is left in the roles AI is about to absorb.
Dropping DEI did not drop the problem
The political backdrop is real. The Guardian notes that the Department of Justice has targeted corporate DEI programmes, and that Accenture, Deloitte, IBM and PayPal have agreed to multimillion-dollar settlements. Felicia Newhouse of AI Powered Women says advocates inside companies are struggling with having women-focused initiatives. Many firms have concluded the safest move is to stop talking about it.
Silence is a poor talent strategy. The programmes were scaffolding around decisions that still get made every week about who to hire, promote and retrain. Cutting the scaffolding leaves those decisions running on defaults, and the defaults are the networks described above. A company filling its scarcest, best-paid roles from a fraction of the available pool is overpaying for talent and calling it prudence.
The most acute case sits in the customer service and administrative teams that many companies are now automating. LinkedIn finds women overrepresented in roles highly exposed to AI disruption, and research from Brookings and the Centre for the Governance of AI, reported by CBS News, found that of roughly 6 million workers who would find it hardest to adapt to AI-related job loss, 86% are women. Brookings' Mark Muro stresses that this reflects the jobs these women hold, not their competence. These workers already know the company's customers and where its products fail. Moving them into AI roles through structured retraining is internal mobility, a practice no compliance team should struggle to defend, and it attacks the gap at the point where the boom is widening it.
Newhouse warns that the deepest risk is a participation gap becoming a power gap, and inside a company the mechanism is concrete. The people building and supervising AI now will decide how it is deployed across the organisation for the next decade, including which jobs it replaces. A firm that staffs that work through its existing referral networks has already taken a side in the debate it hoped to sit out: it has decided that the technology reshaping its workforce will be built by the same narrow group that built the last one, and it is paying a premium for the privilege.
