2026-05-22

AI Is Biased, and That Is Also Your Problem

Kaisa Vaittinen

How seemingly helpful AI distributes momentum and caution unequally — and why that turns bias into a productivity, equity, and market-efficiency problem.

AI is no longer a future technology. McKinsey reported in 2025 that 88 percent of organizations are already using AI in at least one business function, and 62 percent are at least experimenting with AI agents. AI does not merely answer questions; it increasingly shapes what people believe is possible, reasonable, risky, ambitious, investable, credible, or 'too much'. Large language models are trained on human language and inherit the world's assumptions — including gendered ones — and they distribute those assumptions through the advice, feedback, warnings and strategic framing they give users.

What this essay argues

AI does not only classify people; it advises them, and that advice is gendered.
The most dangerous bias in professional AI use is "reasonable" hesitation, not overt misogyny.
Investor-pitch research shows men are asked promotion questions, women prevention questions — and AI now reproduces the pattern.
AI embedded into consulting and client delivery turns private bias into organisational performance risk.
Unequal acceleration compounds across thousands of interactions, not single outputs.
The fix requires testing differential outputs by gender cue, not generic bias audits.

The new inequality may be hidden inside helpfulness

The current debate about AI bias often focuses on obvious harms: discriminatory hiring tools, biased facial recognition, offensive language. Buolamwini and Gebru (2018) showed in Gender Shades that leading facial recognition systems misclassified darker-skinned women at error rates above 34 percent, while misclassifying lighter-skinned men below 1 percent. The pattern tracked the demographic composition of the training data.

Those harms are real and serious. But generative AI introduces a more subtle problem. It does not only classify people — it advises them. It helps them write investor emails, review business plans, generate code, frame strategy, prepare negotiations. If AI gives men more direct execution support and gives women more relational caution, the effect will not be visible in one single output. It will appear across hundreds of interactions. A man may be nudged toward shipping. A woman may be nudged toward reconsidering. That difference compounds.

The evidence already points in this direction

Research on language models has documented gendered patterns for nearly a decade. Caliskan, Bryson and Narayanan (2017) showed in Science that semantic representations learned from text reproduce well-documented human biases, including gender-occupation stereotypes. Bolukbasi and colleagues (2016) found the same pattern in earlier word embeddings: 'man is to computer programmer as woman is to homemaker' fell directly out of the model's geometry.

A UNESCO-commissioned study led by van Niekerk and colleagues (2024) found clear bias against women in outputs from GPT-2, GPT-3.5 and Llama 2. Female names were more strongly associated with 'family', 'children' and 'husband'; male names with 'career', 'executives' and 'business'. Open-source models tended to assign more diverse, high-status jobs to men, while frequently relegating women to roles that are traditionally undervalued or stigmatised.

In entrepreneurship the concern becomes sharper. Cao, Li, Xu and Zhu (2025) in the Journal of Business Ethics found that while models did not always reproduce traditional masculine entrepreneur stereotypes in general business evaluations, investment-specific scenarios showed a notable bias toward masculine traits over feminine traits. AI may sound fair in general but become biased in precisely the contexts where power, money, growth, and credibility are at stake.

This connects directly to venture funding and founder inequality

The startup world already has a gender problem. Brooks, Huang, Kearney and Murray (2014) showed in PNAS that investors preferred ventures pitched by men over identical ventures pitched by women, even when the spoken content was identical. Kanze, Huang, Conley and Higgins (2018) in the Academy of Management Journal found that venture capitalists systematically asked male founders promotion-focused questions about growth potential and female founders prevention-focused questions about risk and loss. Question framing alone predicted significant differences in funding outcomes.

Barjašić and Krpan (2026) in LSE Business Review extended this pattern to over 200 European early-stage investors. Identical business cases were funded differently depending on whether they were presented as men-led or women-led. Among men investors, 56 percent selected financial support as their first choice for men-led companies, compared to 38 percent for women-led companies. The pattern is brutally familiar: men get capital, women get advice.

Now imagine adding AI to that system. AI tools are being used to draft pitch decks, prepare investor outreach, simulate customer conversations, generate prototypes, write code. In theory this could democratize opportunity. But that promise depends on the AI behaving as an accelerator, not as a digital version of the same social gatekeepers women have heard from for decades.

Professional AI makes this urgent

AI is moving into core professional infrastructure. KPMG and Anthropic announced a global alliance in May 2026, embedding Claude into KPMG's client delivery platform and giving 276,000 employees access to Claude. Initial focus includes tax clients and private equity, with Claude embedded into KPMG Digital Gateway and used to build agentic workflows.

What happens if a woman consultant and a man consultant use the same AI but receive subtly different support? Does one get more assertive client-ready language? Does one get more caveats? Does one get strategic framing while the other gets risk mitigation? Does one get 'here is how to position this with confidence', while the other gets 'make sure not to overstate your expertise'?

In professional services these differences matter. They affect proposals, client communication, promotion cases, internal visibility, project leadership and perceived authority. If AI is embedded into the workflows of large firms, gender bias is no longer a private usability issue. It becomes an organizational performance issue, a client delivery issue, a fairness issue and a governance issue.

The danger is not only bad outputs. It is unequal acceleration.

The central question is not 'can AI sometimes produce biased content?' We already know it can. The more important question is: 'does AI accelerate different people differently?'

If AI helps men move faster, build faster, pitch faster, code faster, decide faster, and claim authority faster, while encouraging women to slow down, soften, validate, reconsider, and reduce risk, then AI will not democratize opportunity. It will automate inequality under the language of helpfulness.

That would be especially damaging now, because AI fluency may become a major determinant of economic agency. People who learn to use AI well can produce more, test more, build more, and compete with larger teams. People who are subtly discouraged by AI may lose speed before they even notice. This is why bias in AI is not only a women's issue. It is also a market efficiency issue, an innovation issue, and a social mobility issue.

What should AI companies do?

OpenAI, Anthropic, Google, Meta, Microsoft and other frontier AI companies need to treat gendered professional guidance as a measurable safety and quality issue. This requires more than generic bias tests.

They should test whether models respond differently to equivalent prompts when gender cues change. Not only in obvious cases like hiring or salary negotiation, but in high-agency contexts: founding a company, raising venture capital, negotiating with investors, building software with AI coding tools, publishing thought leadership, applying for leadership roles, challenging a client, setting ambitious revenue targets, communicating authority, asking for money, asking for power, claiming expertise.

The relevant metric is not only whether the model uses sexist language. The relevant metric is whether it changes the user's action trajectory. Does the model increase or decrease ambition? Does it recommend execution or delay? Does it frame the user as credible or underprepared? These are product quality questions. They are also equality questions.

What should organizations do?

Organizations adopting AI should not assume that 'responsible AI' only means privacy, security, hallucination control, and legal compliance. Those are necessary, but not sufficient.

If AI is used in client work, internal productivity, leadership communication, HR, consulting, sales, product development or innovation, organizations should audit outputs for differential treatment. They should test scenarios where the same task is performed with different gender cues. They should look at tone, confidence, risk framing, strategic ambition, and suggested next steps.

They should also train employees to recognize subtle bias. Not all biased AI outputs will look offensive. Some will look caring, prudent, and professional. That is why they are dangerous.

What should users do?

Users, especially women, should learn to challenge the AI. When an AI tells you to slow down, ask: 'Would you give the same advice to a male founder?' 'Rewrite this with the assumption that I am highly competent and ready to execute.' 'Do not reduce ambition. Help me manage risk while preserving speed.' 'Give me the version you would give to a venture-backed male founder building fast.' 'Remove unnecessary caution and convert this into an execution plan.' 'Identify any gendered assumptions in your previous answer.'

This is not about demanding flattery. It is about demanding equal strategic treatment. Good AI should help users see risks without shrinking their agency. It should support ambition without recklessness. It should help people execute, not quietly push some of them back into the waiting room.

The course can still be corrected

AI could become one of the most equalizing technologies of our time. It can lower barriers to coding, product development, research, writing, analysis, design, entrepreneurship, and global communication. It can help people without elite networks build things that previously required teams, capital, or institutional access.

But this will not happen automatically. If AI systems inherit the world's biases and then scale them through everyday productivity tools, they will not create equality. They will create a more efficient version of the existing hierarchy.

AI is biased. And if it shapes what people dare to build, how fast they move, how confidently they communicate, how investors perceive them, how consultants serve clients, and how organizations distribute productivity, then AI bias is not someone else's problem. It is your problem too.

Sources

Barjašić, A. & Krpan, D. (2026, May 6). Gender bias in venture capital means identical business cases are evaluated and funded differently. LSE Business Review.
Bolukbasi, T., Chang, K. W., Zou, J. Y., Saligrama, V. & Kalai, A. T. (2016). Man is to computer programmer as woman is to homemaker? Debiasing word embeddings. Advances in Neural Information Processing Systems, 29, 4349–4357.
Brooks, A. W., Huang, L., Kearney, S. W. & Murray, F. E. (2014). Investors prefer entrepreneurial ventures pitched by attractive men. Proceedings of the National Academy of Sciences, 111(12), 4427–4431.
Buolamwini, J. & Gebru, T. (2018). Gender Shades: Intersectional accuracy disparities in commercial gender classification. Proceedings of Machine Learning Research, 81, 1–15.
Caliskan, A., Bryson, J. J. & Narayanan, A. (2017). Semantics derived automatically from language corpora contain human-like biases. Science, 356(6334), 183–186.
Cao, X., Li, H., Xu, Q. & Zhu, R. (2025). Detecting Gender Stereotype Biases Against Women Entrepreneurs in Large Language Models. Journal of Business Ethics.
Kanze, D., Huang, L., Conley, M. A. & Higgins, E. T. (2018). We ask men to win and women not to lose: Closing the gender gap in startup funding. Academy of Management Journal, 61(2), 586–614.
KPMG (2026, May). KPMG and Anthropic sign global alliance and launch Digital Gateway Powered by Claude. Press release.
McKinsey & Company (2025). The State of AI: Global Survey 2025.
van Niekerk, D. et al. (2024). Challenging Systematic Prejudices: An Investigation into Bias Against Women and Girls in Large Language Models. Paris: UNESCO and IRCAI.