Affiliation: Australian-born scholar; Research Professor, USC Annenberg School for Communication and Journalism; Senior Principal Researcher, Microsoft Research New York (FATE group) [TBD: confirm current title — has held this since ~2017]; Honorary Professor, University of Sydney; co-founder, AI Now Institute (NYU, 2017, with Meredith Whittaker); co-founder, Knowing Machines project; previously Visiting Chair in AI and Justice, École Normale Supérieure (Paris, 2021–2022 [TBD: confirm dates]) One-line position: AI is not artificial and is only weakly "intelligent" — it is an extractive industrial system built on labor, data, minerals, water, and energy, and treating it as anything else (a neutral tool, a coming superintelligence, or a purely software phenomenon) obscures the politics baked into every layer of the stack.
What she's reacting against
- Framings of AI as immaterial / software-only — argues the material substrate (data centers, mineral extraction, freshwater cooling, energy grid load, data labor) is constitutive of what AI is, not an externality
- "AI ethics" discourse focused on model-level fairness while ignoring upstream labor and downstream environmental cost
- X-risk discourse that centers future superintelligence while present, distributed, measurable harms accumulate — overlaps with Gebru on this but with a different (materiality-first) angle
- Classification systems presented as neutral or empirical — argues training-set design (ImageNet, IBM Diversity in Faces, etc.) carries normative commitments
- "AI as inevitable" framings from Andreessen / Altman — treats inevitability rhetoric as a political move that forecloses regulation
- Pure intra-industry self-regulation — argues the firms doing the extracting are not credible regulators of the extraction
Key claims
- Empirical: AI systems depend on a global supply chain — rare earth and lithium mining, freshwater for data center cooling, fossil-fuel-heavy electrical load, and underpaid annotation/moderation/RLHF labor (often in the Global South) — that is largely invisible in model-centric discourse (Atlas of AI, 2021)
- Empirical: training-set construction is an act of classification that encodes normative choices ("Excavating AI", with Trevor Paglen, 2019) — load-bearing example: ImageNet's "person" subtree carried offensive and politically loaded categories that had been treated as neutral ground truth
- Empirical: data center water consumption and energy demand for frontier-model training have grown substantially through 2023–2025 and are now material at watershed and grid scale [TBD: cite specific 2024–25 data; IEA Energy and AI report 2024 is the standard reference]
- Empirical: data labor (annotation, content moderation, safety review, RLHF rating) is concentrated in low-wage workforces — overlaps with Gebru, but Crawford emphasizes the supply-chain mapping ("Anatomy of an AI System", with Vladan Joler, 2018) rather than the representational-harm framing
- Normative: regulation must address the full stack, not just outputs — supply chain, labor conditions, environmental cost, classification politics, and deployment effects
- Empirical / political: AI-industry discourse is itself a site of power — what counts as "AI safety", who gets to define "intelligence", which harms are made visible and which are made invisible are political outcomes, not technical ones
- Normative: independent, non-industry research is necessary because the firms doing the building are structurally biased — AI Now Institute (2017) and Knowing Machines (ongoing) are the institutional bets
- Empirical: visual-classification systems (facial recognition, emotion recognition, demographic inference) often rest on weak or contested scientific foundations — emotion recognition in particular reproduces discredited 20th-century affect theories ("AI Now 2019 Report" co-authored with Whittaker; "Discriminating Systems", 2019) [TBD: confirm exact title/authorship of the affect-recognition piece]
Theories aligned with
- Critical AI / critical algorithmic studies (Crawford, Buolamwini, Noble, Benjamin, Whittaker, Gebru lineage) — currently no theory file [TBD: create
theories/critical-ai-studies.md] - Political ecology of computing (data center water/energy, mineral extraction) — connects to the broader water research thread in Rich's workspace
- Science and Technology Studies (STS) framings of "objectivity" and "neutrality" in classification
- Distinct from techno-optimism, AI alignment / x-risk, and post-labor economics — would argue all three under-theorize the material substrate
Where she overlaps / splits
- Overlaps with Gebru on rejecting x-risk centrality and on data-labor visibility; splits on emphasis — Crawford foregrounds the material/extractive supply chain (minerals, water, energy, infrastructure), Gebru foregrounds representational harm + the TESCREAL ideological-genealogy critique. The two are complementary critical-AI camps with different load-bearing arguments
- Overlaps with Zuboff on Big Tech as a power-concentrated political-economic phenomenon worth diagnostic seriousness; splits on the unit of analysis — Zuboff diagnoses the business model (behavioral surplus → prediction products → behavioral futures markets); Crawford diagnoses the production infrastructure (the material/labor/data supply chain that produces the systems in the first place)
- Overlaps with Varoufakis on Big Tech as a structural / political-economic problem; splits on framing — Varoufakis: techno-feudalism / cloud capital / mode of accumulation; Crawford: extractive industrial geography of AI itself
- Splits with Andreessen, Diamandis, Blundin, Ismail on the framing of AI as costless / abundance-producing — Crawford argues the costs are real, measurable, and unevenly distributed; abundance rhetoric is empirically thin on inputs
- Splits with Altman on the framing of "intelligence" — Crawford argues "intelligence" as deployed in industry discourse smuggles in assumptions about cognition that the systems don't satisfy and that the rhetoric obscures
- Splits with Yudkowsky, Bostrom on what the load-bearing risk is — Crawford treats x-risk discourse as a category error that displaces measurable present harm; would agree with the Gebru TESCREAL framing of the intellectual lineage running through Bostrom even while less programmatic about it
- Splits with Amodei, Hassabis, Bengio on what "AI safety" should center — Crawford: full-stack accountability for material and labor harms; their frame: alignment + governance of frontier capabilities
- Overlaps with Acemoglu on the institutional / power-concentration angle; splits on the lever — Acemoglu emphasizes institutional reform of capitalism, Crawford emphasizes mapping and constraining the extractive infrastructure
- Overlaps with Toner on the empirical failure of lab self-regulation; splits on the proposed lever — Toner: external accountability + compute governance via institutional design; Crawford: full-stack accountability extending to supply chain and labor
- Less engaged with Brynjolfsson, Cowen, Frey on macro labor productivity — would argue their data layer treats data labor itself as exogenous when it is the substrate
Notable predictions
- (2018, "Anatomy of an AI System") The material supply chain of AI (minerals, energy, labor, data) will become increasingly visible and politically contested — outcome: largely correct; by 2024–25, data center water consumption, grid load, and AI-driven mineral demand are mainstream policy concerns (IEA, IMF, FT, Bloomberg coverage 2024–25 [TBD: cite specific outlets])
- (2019, "Excavating AI") Training-set construction will be recognized as a politically loaded act, and major training sets will face scrutiny and revision — outcome: largely correct; ImageNet removed contested "person" categories shortly after the paper (2019-09 [TBD: confirm exact date]); subsequent dataset audits (LAION, etc.) follow the same template
- (2021, Atlas of AI) Environmental cost of frontier AI training will move from niche concern to mainstream regulatory question — outcome: largely correct; EU AI Act (2024), state-level US water-disclosure proposals, and major-outlet coverage of data center water use vindicate the framing in 2024–25
- (2019, AI Now report on affect recognition) Emotion / affect recognition systems rest on weak scientific foundations and will face increasing regulatory pushback — outcome: partially correct; EU AI Act (2024) restricts emotion recognition in workplaces and schools; deployment in adjacent areas (hiring, education) continues [TBD: confirm AI Act specifics]
- (2017, AI Now founding premise) Independent, non-industry AI research outside Big Tech is necessary and viable — outcome: durable institutional bet; AI Now Institute continued through multiple organizational transitions; the broader independent-research ecosystem (DAIR, Distributed AI, ARC, Apollo, etc.) suggests the premise has been validated
Track record
- Atlas of AI (2021, Yale University Press) is the load-bearing intellectual contribution — durable, widely assigned in graduate courses across STS, communication, and AI ethics; reviewed in major outlets; shifted the discourse on AI's material substrate
- "Anatomy of an AI System" (2018, with Vladan Joler) — the Amazon Echo as a labor/material/data extraction diagram — became a canonical teaching artifact; acquired by MoMA and V&A [TBD: confirm acquisition details]
- "Excavating AI" (2019, with Trevor Paglen) — the ImageNet "person" subtree exposé — drove Princeton to remove contested categories [TBD: confirm exact removal date and Princeton statement]; load-bearing case study in critical-AI pedagogy
- "Calculating Empires" (2023–2024, with Joler) — historical atlas of computational empires; less load-bearing than Atlas of AI but extends the methodology [TBD: confirm exhibition / publication details]
- AI Now Institute (2017–) — institutional vehicle; AI Now annual reports (2016–2019, then less regular) were widely cited in early AI governance debate
- Pattern: high-quality conceptual contributions paired with collaborative visual / curatorial work; less prolific on programmatic policy specifics than Toner or Bengio, but more empirically grounded on the supply-chain layer than any other persona in this project
- Reception: strong in critical-AI, STS, and political-ecology circles; engaged but contested in mainstream AI safety discourse (where the framing is sometimes received as a category challenge rather than a complementary critique)
Empirical vs normative
- Empirical: claims about mineral / water / energy / data labor as part of the AI supply chain; claims about classification politics in training sets; claims about commercial-facial-recognition / affect-recognition error rates and weak scientific foundations
- Normative: claims that the full stack should be the unit of regulation; that "intelligence" is a misleading frame for what current systems do; that independent research outside industry is necessary; that material and labor harms warrant priority over speculative future ones
- Microsoft Research affiliation is a meaningful conflict-of-interest to track (CONVENTIONS rule 25) — Crawford works inside one of the major firms while critiquing the industry; she has been transparent about this, but the structural tension is real and is itself an interesting data point (parallel to Sutskever's split-from-OpenAI but inverse — she has stayed inside)
- Less commercial dependency than frontier-lab CEOs; reputational stakes are academic + institutional rather than equity-linked
Sources
- Books: Atlas of AI: Power, Politics, and the Planetary Costs of Artificial Intelligence (Yale University Press, 2021) — load-bearing primary source
- Peer-reviewed / journal: "Discriminating Systems: Gender, Race, and Power in AI" (AI Now report, with Whittaker et al., 2019) [TBD: confirm publication venue]; multiple articles in New Media & Society, Information, Communication & Society, Science, Technology, & Human Values [TBD: specific citations]
- Reports: AI Now Institute annual reports (2016, 2017, 2018, 2019) — co-authored with Meredith Whittaker and others; widely cited in early AI governance debate
- Visual / curatorial: "Anatomy of an AI System" (with Vladan Joler, 2018) — diagram + essay; "Excavating AI" (with Trevor Paglen, 2019) — essay + image archive; "Training Humans" exhibition (Fondazione Prada, Milan, 2019, with Paglen); "Calculating Empires" (2023–2024, with Joler) [TBD: confirm venue]
- Project: Knowing Machines — ongoing research project on the dataset infrastructure of ML
- Interviews / podcasts: Ezra Klein Show [TBD: episode + date]; Tech Won't Save Us [TBD: episode]; Logic(s) Magazine interview [TBD: date]; multiple TED-format and conference talks
- Public position statements: regular op-eds in The Guardian, The New York Times, The Atlantic [TBD: specific pieces]
- Press: coverage of Atlas of AI (2021, multiple outlets); coverage of "Excavating AI" and ImageNet response (2019)
Weak spots / open questions
- Strong on the diagnostic side; less programmatic on what specific regulations / mechanisms would address the supply-chain layer — closer to Bostrom in being more conceptual than institutional-design, despite operating in a very different camp
- The "AI is not intelligent" framing is rhetorically effective but operationally hard to translate into specific policy — it cedes some ground in mainstream AI policy discussions where "intelligence" is the accepted vocabulary
- Material-extraction framing risks under-engaging with the productivity / capability side of the ledger — would benefit from direct engagement with Brynjolfsson-style augmentation evidence rather than treating it as out of scope
- Microsoft Research affiliation is a structural conflict-of-interest that critics on both sides will use (frontier-lab critics: "captured insider"; critical-AI peers: "compromised distance") — Crawford handles this transparently but it limits some of the moves she can make
- Less engaged with the alignment / x-risk technical literature than would be ideal for the inter-camp argument — the work mostly bypasses the Bostrom / Russell framings rather than engaging with them on their own terms, which makes the camps talk past each other rather than to each other
- Predictions on environmental cost have been largely vindicated, but specific quantitative claims about water / energy / mineral intensity vary across sources and would benefit from a primary-source pass [TBD: IEA Energy and AI 2024 report; specific freshwater consumption figures for major data center clusters]
- AI Now Institute has gone through organizational transitions [TBD: confirm 2024–25 institutional status and Crawford's current role there] — track whether the institutional vehicle remains active as a distinct critical-AI research center
Rich's take
- (your synthesis here)
converts-from: personas/kate-crawford.md · schema v1 · AI & Society domain