Episode Summary
Executive Summary: Alex Hanna of DARE argues that AI should be treated as a socially situated technology rather than an inevitable general-purpose solution. The conversation focuses on community-led research, data provenance and ethics, the politics embedded in datasets, and the dangers of AI hype—especially in high-stakes domains like medicine—while highlighting DARE projects that build useful, culturally grounded tools for underserved communities.
Main Topics: DARE’s origin and research model (Priority: 5/5): Hanna explains her path from sociology to machine learning fairness work at Google, then to DARE after Timnit Gebru’s departure from Google. DARE’s research agenda is not top-down; it is built by bringing in fellows and researchers with existing expertise and lived experience, then supporting them in producing research aligned with community needs. Community-centered AI and multiple ways of knowing (Priority: 5/5): A core DARE principle is that AI should serve communities, not extract from them. Hanna emphasizes that lived experience, activist knowledge, and academic expertise are all valid forms of knowing, and that DARE aims to turn researchers into advocates and advocates into researchers. Data politics, provenance, and dataset ethics (Priority: 5/5): Hanna discusses her paper 'Do Data Sets Have Politics?' and argues that datasets are value-laden, shaped by choices about sourcing, consent, representation, speed, and who does the labor. She stresses that better practice requires transparency, auditing, and regulation—not just awareness. Useful AI use cases versus extractive hype (Priority: 4/5): Hanna distinguishes narrow, beneficial applications like machine translation and speech recognition for under-supported languages from the broad, extractive deployment of LLMs in workplaces and consumer products. She argues that current hype often ignores whether a tool actually meets a real community need. Risks of AI in medicine and other high-stakes settings (Priority: 5/5): The conversation repeatedly returns to medical and mental-health applications as especially dangerous when deployed without rigorous evaluation. Hanna cites examples such as therapy chatbots, diagnostic tools, and mushroom-identification systems that could directly harm people. Historical continuity of AI hype (Priority: 4/5): Hanna frames today’s excitement around LLMs as part of a long history of AI hype, referencing Joseph Weizenbaum, ELIZA, MIT AI culture, and modern claims about sentience or job replacement. She argues hype is tied to power, funding incentives, and devaluation of human expertise.
Key Arguments: AI is not inevitable; it should be evaluated as a tool with narrow, context-specific benefits rather than assumed to be a universal solution. Community expertise and lived experience are legitimate sources of knowledge and should shape AI research agendas from the start. Datasets embody politics through choices about sourcing, labeling, consent, representation, and labor conditions; they are not neutral inputs. Current AI practice overvalues model-building and undervalues the slow, careful work of data collection, documentation, and stewardship. Transparency is essential because many harms in AI come from opaque datasets and un-auditable training pipelines. High-stakes applications, especially medicine and mental health, require robust evaluation, clinical validity, and close human oversight before deployment. AI hype is driven by institutional and financial incentives, not just technical capability, and can normalize harmful uses before evidence exists. Machine translation and speech recognition can be genuinely valuable when developed with and for underserved language communities and governed with data sovereignty in mind.
Data Points: DARE launch timing: December 2021 - Hanna says DARE will be almost two years old in December, describing the institute’s early history. Hanna joined DARE: February 2022 - She joined as director of research, employee number three, three months after DARE’s announcement. Google research milestone: First social scientist hired on that ladder - Hanna says she was the first research scientist social scientist ever hired at Google on that ladder. Conference participation start: 2018 - She began attending FACT, a fairness conference, regularly starting in 2018. Dataset study sample size: 100 datasets sampled from roughly 500–700 computer vision datasets - For 'Do Data Sets Have Politics?', the team built a population-level view and coded a sample plus the 14 most highly cited datasets. Dataset variables coded: 100 variables - The paper coded datasets across a large number of quantitative and qualitative dimensions. Language population example: 20–30 million - Hanna estimates Amharic is spoken by 20 or 30 million people in Ethiopia. Language population example: About 3 million - She estimates Tigrinya is spoken by about 3 million people. Investment in AI industry: $44 billion - Hanna cites Pitchbook data on industry investment at the time of the interview. Model valuation: Trillion dollars - She says the AI industry had a trillion-dollar valuation. Med-PaLM 2 benchmark result: 68% accuracy - Hanna references a Google evaluation that initially reported this score on the U.S. medical licensing exam. Med-PaLM 2 later benchmark result: In the 80s - She notes later accuracy reportedly improved into the 80% range, while arguing the benchmark still lacks clinical validity. AI sentience controversy: Lambda / sentient claims - She references Blake Lemoine’s firing and subsequent public commentary by Google leaders and OpenAI figures.
Pivotal Quotes: "AI is not inevitable" — Alex Hanna: Core framing for DARE’s mission and the broader critique of techno-determinism. "How do we turn researchers into fighters and fighters into researchers?" — Alex Hanna: Describing DARE’s approach to blending scholarship, activism, and community knowledge. "We need to develop technology that works for people based in our communities." — Alex Hanna: Summarizing the institute’s practical and ethical North Star.
Implications: Listeners should expect more scrutiny of AI claims, especially in medicine, labor, and language tech. The episode suggests the future of responsible AI depends on community governance, transparent data practices, and resisting one-size-fits-all deployment.