Women, Work, and the Future of AI
Abstract
The session opened with two scenes from the documentary Humans in the Loop that illustrated how most AI training data is produced by largely invisible human workers—many of them women. A moderated panel then examined the structural reasons why women’s labor remains hidden, how gender bias infiltrates AI models, and what governance, research, and policy measures can surface and empower women’s knowledge. The discussion culminated in a live demonstration of the “AlignBench” gender‑bias benchmarking tool built on data collected from 20 000 women across eight Indian states, showcasing a concrete step toward community‑centered AI.
Detailed Summary
- Aranya Sahay introduced the session, framing the central paradox: women dominate data‑collection and annotation work (the “gig” layer of the AI pipeline) yet remain invisible in AI discourse.
- Two short clips from Humans in the Loop depicted:
- Data‑generation pipeline – women labeling images, often in low‑visibility, low‑pay gigs.
- A failure case – an AI model mis‑identifying a tribal woman because the community’s contextual knowledge was omitted.
- Sahay highlighted that without the lived‑experience of these workers, AI systems become “technically sophisticated but socially shallow.”
2. Panel Introduction & Energizer
- Moderator (Sahay) introduced the five panelists and asked each to respond in one‑to‑two‑second bursts to the prompt “Women + AI → Power or Precarity?”
- Saachi Bhalla (Gates Foundation) – Precarity mixed with opportunity; warned that without intentional inclusion women may be relegated to informal, under‑paid AI work.
- Dr. Shikoh Gitau (Qhala) – Emphasised possibility; cited a scene where a data‑labeler corrected a crop‑prediction error, illustrating how local knowledge can improve model performance.
- Dr. Kalika Bali (Microsoft) – Pointed to a Mid‑Journey example where the model refused to generate a Black doctor treating a white child, showing systemic bias in generative AI.
- Urvashi Aneja (Digital Futures Lab) – Mentioned a community‑health‑worker app that failed because women could not hide phones (they kept phones in their bras), underscoring design that neglects women’s lived constraints.
- Safiya Husain (Karya) – Noted that anecdotal stories are insufficient; AI currently reproduces inequality, centralizes power, and creates epistemic injustice that disproportionately harms women.
3. Core Themes Explored
3.1 Invisible Labor & Structural Invisibility
- Aranya Sahay argued that AI’s “aesthetic of automation” deliberately hides the human labor behind the curtain, especially the work performed in the Global South.
- The panel agreed that the invisibility is structural (industry‑wide norms that celebrate “intelligence” while erasing the “human hand”).
3.2 Data‑Worker Knowledge as Strategic Asset
- Safiya Husain cited Mary L. Gray’s “Ghost Work” and explained how Karya’s case studies revealed that women data‑workers already embed nuanced, culturally‑specific knowledge in annotations.
- Kalika Bali added that when women evaluate outputs, they flag issues men may overlook—e.g., recipe‑collection projects that surface seasonal, health‑related ingredient knowledge.
3.3 Governance: Representation vs. Redistribution of Power
- Urvashi Aneja stressed that mere representation (having women in annotation pools) can become tokenism if it does not accompany a redistribution of decision‑making power.
- She warned that exposing marginalized groups in data sets can inadvertently empower surveillance or harmful applications (e.g., facial‑recognition for law‑enforcement).
- Safiya Husain echoed this, noting the need for safeguards so that community‑generated data are not commodified without benefit to the contributors.
3.4 Platform‑Work Precarity for Women
- Urvashi Aneja described how task‑based gig platforms create income volatility, which, coupled with care responsibilities, pushes women out of the labor market or forces them into insecure employment.
- Algorithmic management systems often use proxy metrics (e.g., daily travel logs) that penalize women who stay at home for caregiving, thereby reducing their credit scores and access to fintech services.
- She highlighted emerging Indian policy signals (e.g., inclusion of algorithmic‑management concerns in the Economic Survey) and the importance of visibility as a first regulatory step.
3.5 Intersectionality & “One‑Size‑Does‑Not‑Fit‑All”
- Kalika Bali reminded the group that gender bias manifests differently across caste, class, and language.
- In Indian languages, bias is frequently expressed through personality and emotional descriptors, contrasting with English’s focus on occupation.
- Aranya Sahay illustrated this with the example of Adivasi and Dalit women being profiled by policing AI trained on dominant‑group data, perpetuating systemic oppression.
4. Demonstration – AlignBench Gender‑Bias Benchmark
- Safiya Husain (Karya) presented the AlignBench tool, a product of a two‑year, 20 000‑woman collaborative effort across six Indian languages.
- Methodology Overview
- Definition Phase – 100 women co‑created culturally‑grounded definitions of “gender bias” in their own languages (many Indian languages lack a direct lexical equivalent).
- Data‑Collection Phase – 20 000 women annotated real‑world sentences, marking bias direction, linguistic markers, target groups, and harm type.
- Framework Construction – The annotations were distilled into 13 parameters, grouped into four high‑level categories:
| Category | What it captures |
|---|---|
| Bias Presence & Direction | Whether bias exists and its polarity (positive/negative) |
| Linguistic Markers | Words/structures that signal bias |
| Target & Explicitness | Who is affected and whether bias is overt |
| Functional Harm | Psychological, social, or representational impact |
-
Live Demo – A sample sentence in Hindi was input: “Amit enjoys cooking; his friends mock him for being un‑manly.” The tool flagged:
- Negative tone
- Domestic domain, role‑based assumption
- Harms: psychological, social, representational
-
The demo highlighted how the benchmark can automatically surface bias that would otherwise remain hidden, even when the bias stems from cultural expectations rather than overt slurs.
5. Closing Reflections & Call to Action
- Saachi Bhalla reiterated that recognizing women as economic actors (not just beneficiaries) is essential. She cited India’s National Rural Livelihoods Mission as a model for collective empowerment, financial inclusion, and market linkage.
- Dr. Shikoh Gitau warned that AI tools built on data produced by men‑identified gig‑workers will continue to marginalize women unless platforms redesign attribution mechanisms.
- Kalika Bali urged the community to stop infantilizing women and to listen to the concrete demands of women farmers and gig workers (e.g., desire for physical‑labor‑reducing robotics rather than more information apps).
- The moderator thanked the audience, highlighted the presence of Karishma (author of The Human Touch) and Lakshmi (AI Kiran), and invited further collaboration on community‑centred AI initiatives.
Key Takeaways
- Invisible labor matters – Women constitute the majority of data‑labelers and gig workers that train AI, yet their contributions are largely unseen and undervalued.
- Local knowledge is infrastructure – Community‑grounded insights (e.g., seasonal ingredients, cultural norms) are essential for building AI that works beyond code‑level pattern matching.
- Representation ≠ empowerment – Adding women to annotation pools without giving them decision‑making authority risks tokenism and can even exacerbate harms.
- Platform work amplifies gendered precarity – Task‑based gig economies and algorithmic management ignore care responsibilities, leading to income instability and biased credit assessments.
- Intersectionality is key – Bias manifests differently across language, caste, and region; tools must capture personality‑ and emotion‑based stereotypes, not just occupational ones.
- AlignBench shows a scalable, community‑driven solution – Co‑created definitions and large‑scale annotation enable a multilingual gender‑bias benchmark that flags nuanced harms.
- Policy momentum in India – New governmental recognitions of algorithmic management and platform‑worker visibility signal a fertile ground for regulation.
- Economic‑actor framing – Treating women as active participants in AI value chains (e.g., via self‑help groups, collective bargaining) can shift power from “invisible” to “visible.”
- Measurement matters – Developing shared metrics and language (e.g., the 13‑parameter framework) is vital for tracking progress and holding stakeholders accountable.
- Future AI should be “minimum viable intelligence” for women – Systems must first ask women what problems matter to them (health, childcare, safety) and build modest, context‑specific solutions rather than grand, one‑size‑fits‑all models.
See Also: