ResourcesView as Markdown

References and Further Reading

UNC Water & Health 2026 · Side Session · Baobab Tech · October 2026

Numbers match the citations on the slides, in order of first appearance. Research notes link to the evidence summaries in baobab-tech/research-ai-energy.

References

#ReferenceSupports
1Luccioni, A.S., Gamazaychikov, B., Alves da Costa, T. & Strubell, E. (2025). Misinformation by Omission: The Need for More Environmental Transparency in AI. arXiv:2506.15572. https://arxiv.org/abs/2506.15572In May 2025, 84% of LLM token usage on OpenRouter's top 20 models went through models with no environmental disclosure
2Elsworth, C. et al. (Google) (2025). Measuring the environmental impact of delivering AI at Google scale. arXiv:2508.15734. https://arxiv.org/abs/2508.15734Median Gemini text prompt: 0.26 mL on-site cooling water (scope 1)
3Li, P., Yang, J., Islam, M.A. & Ren, S. (2023, rev. 2025). Making AI Less "Thirsty". arXiv:2304.03271. https://arxiv.org/abs/2304.032712.2 mL per request on site (scope 1), 16.9 mL including electricity (US average); GPT-3 training 5.44 million litres, 87% from electricity generation
4Akinade, Amanambu, Frame & Ren (2026). arXiv:2606.21760. https://arxiv.org/abs/2606.21760. Research noteMeta, Lebanon, Indiana: peak-day demand 1.34x the host utility's capacity. Caveat: that is the current municipal system; it falls to 0.25x against the dedicated supply being built, and the peaking factor is modelled, not metered.
5Congressional Research Service (31 Jul 2026). Data Centers and Water (R49057). https://www.congress.gov/crs-product/R49057. Research noteRichland Parish: permitted 8.4 bn gal/yr against 0.5 to 0.6 bn stated (about 14x); rates and service agreements often confidential
6Han, Y., Wu, Z., Li, P., Wierman, A. & Ren, S. (2026). arXiv:2412.06288 (v4). https://arxiv.org/abs/2412.06288. Research note$20.9bn and 1,262 deaths in 2028 under high growth; over 90% from generating plants
7Luccioni, A.S., Jernite, Y. & Strubell, E. (2024). "Power Hungry Processing." FAccT '24. https://doi.org/10.1145/3630106.3658542. Research noteImage generation 2.907 kWh per 1,000 inferences against 0.002 kWh for text classification (>1,450x); extractive QA: task-specific 0.3 gCO₂e against 10 g for a multi-purpose model
8Morrison, J., Smith, N.A. & Strubell, E. (2026). arXiv:2605.01158. https://arxiv.org/abs/2605.01158. Research noteReasoning post-training cost 17x the instruction variant (OLMo 3)
9Roussilhe, G., Dromard, B. & Mitra, S. (2026). arXiv:2604.26539 (preprint, not peer reviewed). https://arxiv.org/abs/2604.26539. Research noteMicrosoft–XTO Permian deal implies about 6.7 Mt CO₂ in 2025, 43% of Microsoft's 2024 market-based footprint; not in its Scope 3. Authors' own extrapolation from vendor claims.
10IEA (Apr 2025). Energy and AI, "Energy demand from AI." https://www.iea.org/reports/energy-and-ai/energy-demand-from-ai. Research note415 TWh in 2024, about 945 TWh by 2030 (base case)
11Rosa, L. et al. (Sep 2026). Environmental Science & Technology. https://doi.org/10.1021/acs.est.6c08088. Research note25% of 9,347 mapped data centres in cells where water demand exceeded supply at least one month a year; 33% at +3 °C
12IEA (14 Oct 2026). Powering Artificial Intelligence. https://www.iea.org/reports/powering-artificial-intelligence. Research note34% of AI data-centre capacity in high baseline water stress; hyperscaler WUE figures use two different water measures
13Alston, Love & Haider (21 Sep 2026). arXiv:2609.25437. https://arxiv.org/abs/2609.25437. Research noteMichigan grid model: air-cooled chillers cut a 200 MW site's total water 44%, with 96.6% of the remainder at power generators
14Gauld, K.D., Varon, D.J., Balasus, N. & Cusworth, D.H. (Aug 2026). arXiv:2608.22153. https://arxiv.org/abs/2608.22153. Research noteSatellite-derived NOx at Southaven about 19x the permitted rate
15Alpine, W., Alpine, H. et al. (2026). AI-driven productivity gains enable more CO₂ emissions than they avoid in a global energy–economy model. npj Climate Action. https://www.nature.com/articles/s44168-026-00411-0AI in oil and gas enables more emissions than AI avoids. Note: the journal is npj Climate Action, part of the Nature portfolio, not Nature. Axios names co-leads from Purdue University and an independent researcher.
16Martin, A. (13 Aug 2026). "AI Applications for Oil and Gas Companies Worsen Climate Pollution." Inside Climate News. https://insideclimatenews.org/news/13082026/ai-enables-oil-and-gas-production/0.47 to 1.8 Gt CO₂ a year; 1.2 to 4.8% of 2024 energy-related emissions; up to 13x the IEA's data-centre estimate; AI must lift renewables 4 to 5 times more than oil and gas to keep emissions flat
17Axios (11 Aug 2026). "AI could help unlock more oil — and emissions." https://www.axios.com/2026/08/11/ai-oil-gas-emissions-increaseRoughly 3 to 13x IEA data-centre emissions; 64 scenarios, emissions fell only when AI gave no fossil productivity gain
18Anthropic (28 Aug 2025). "Updates to Consumer Terms and Policy." https://www.anthropic.com/news/updates-to-our-consumer-termsFree, Pro and Max users asked to choose whether chats train models; five-year retention for those who allow it (30 days otherwise); commercial, education, government and API excluded
19Tom's Guide (Aug 2025). "Claude AI will start training on your data soon — here's how to opt out." https://www.tomsguide.com/ai/claude/claude-ai-will-start-training-on-your-data-soon-heres-how-to-opt-out-before-the-deadlineFor existing users the large "Accept" button opted them in, and opting out took a separate toggle
20OpenAI. "Memory FAQ." https://help.openai.com/en/articles/8590148ChatGPT "memory" keeps details across conversations
21Anthropic. "API and data retention." https://platform.claude.com/docs/en/manage-claude/api-and-data-retentionZero data retention is offered on commercial API terms
22heise online (2026). "OpenStreetMap is concerned: thousands of AI bots are collecting data." https://www.heise.de/en/news/OpenStreetMap-is-concerned-thousands-of-AI-bots-are-collecting-data-11157359.htmlAbout 100,000 IP addresses scraping OSM in coordination within one week, raising hosting costs
23Boettcher, H. (29 Aug 2026). "Consumed at Scale: AI-Driven Extraction and the Political Economy of OpenStreetMap." State of the Map 2026. https://pretalx.com/sotm2026-osm-science/talk/MWCJFD/Dates the 100,000-IP event to late January 2026; costs fall largely on volunteers
24Earthmover (Oct 2026). LinkedIn post, "the scientific community is rethinking the meaning and purpose of open data in the era of AI", with reply thread incl. Alex Logan (Cecil) and Olivier Mills. Text supplied by Olivier.Frontier labs harvest open data en masse; "Simply providing unbounded anonymous access may turn out to be unsustainable"; Earthmover lets providers see users, set licences, charge for access; Logan: "does this resolve the last mile delivery issue?"
25McKinsey & Company (25 Aug 2026). The state of AI in 2026: On the road to ROI. https://www.mckinsey.com/capabilities/quantumblack/our-insights/the-state-of-aiNearly 90% use AI in at least one function; 44% scaling across the enterprise (38% a year earlier); n = 1,719
26Bridgespan Group with NTEN (10 Sep 2026). Choosing Your AI Path. https://www.bridgespan.org/press-releases/turning-ai-opportunity-into-strategy70% of nonprofit leaders and staff think they're missing AI opportunities; 8% have a one- to two-year AI roadmap. Sample size not stated on the press page.
27Center for Effective Philanthropy (2025). AI With Purpose. https://cep.org/blog/ai-with-purpose-as-funders-and-nonprofits-adopt-ai-whats-missing90% of foundations offer grantees no AI adoption support; 400+ nonprofit leaders surveyed
28Fortune (21 Oct 2026). "AI helped discover the PHP vulnerability that could take down 70% of the internet." https://fortune.com/2026/10/21/ai-tools-discovered-the-php-vulnerability-that-could-take-down-70-of-the-internet-now-the-race-is-on-to-patch-it-before-ai-powered-hackers-attack/Two critical PHP flaws (CVE-2026-109618, -109619) found with AI models; disclosed 21 Oct 2026; Cloudflare saw tens of thousands of exploit attempts within 45 minutes
29Axios (22 Oct 2026). "Black October." https://www.axios.com/2026/10/22/php-vulnerability-security-update-websitesFlaws went undetected for more than a decade; PHP runs about 70% of websites whose server-side language is known; "Black October" label
30Wikipedia. "Jev (AI model)." https://en.wikipedia.org/wiki/Jev_(AI_model)TypeSafe AI; limited early access 15 Sep 2026; $40M seed led by DCVC; closed model
31Latent Space / AINews (19 Sep 2026). "Here are 6 Clones of Jev in 2 days." https://latent.space/p/ainews-here-are-6-clones-of-jev-in/Six clones within two days (Laya, DiffusionGemmaJev, Bespoke Nimble, SemIf/OpenJev, Jevlike, Kev-0.5B); several public on GitHub/Hugging Face
32TechCrunch (1 Oct 2026). "Amazon releases its own Jev clone as decision models flood the web." https://techcrunch.com/2026/10/01/amazon-releases-its-own-jev-clone-as-decision-models-flood-the-web/"Dozens of similar models"; AWS Strands Decider 2B, open source, 1 Oct 2026
33Wikipedia. "Shovelware." https://en.wikipedia.org/wiki/ShovelwareSoftware valued for quantity over quality; early-1990s CD-ROM bundles; the Wii
34Maciejewska, J. (29 Mar 2024). Post on X/Twitter, via WIST. https://wist.info/maciejewska-joanna/81891/"I want AI to do my laundry and dishes so that I can do art and writing, not for AI to do my art and writing so that I can do my laundry and dishes."
35The Register (8 Oct 2024). "Eric Schmidt speech." https://www.theregister.com/2024/10/08/eric_schmidt_speech/ ; also Windows Central, same weekSchmidt: "we're not organized" to hit climate goals, "I'd rather bet on AI solving the problem than constraining it and having the problem." Used as the third "promise" line.
36The Mary Sue (6 Oct 2026), on The Daily Show, "Jon Stewart Unpacks AI's Dubious Promise of a Better Future." https://www.themarysue.com/jon-stewart-calls-out-industry-giants-claiming-ai-will-save-humanity-and-his-breakdown-of-their-current-priorities-is-impossible-to-ignore/Quotes: "mostly murder, surveillance, and surge pricing for the McRib"; "If we let them do enough unfettered capitalism, we can apparently capitalism our way to socialism"; executives' "universal high income" promise
37Kobak, D. et al. (2025). "Delving into LLM-assisted writing in biomedical publications through excess vocabulary." Science Advances. arXiv:2406.07016. https://arxiv.org/abs/2406.07016At least 13.5% of 2024 biomedical abstracts show signs of LLM processing, across 15.1 million PubMed abstracts (2010-2024).
38EdArXiv preprint (posted 12 Aug 2026). https://osf.io/preprints/edarxiv/n29zd_v3, reported by AI Weekly (23 Oct 2026). https://aiweekly.co/alerts/edarxiv-audit-faults-all-14-ai-in-education-meta-analysesAudit of 14 AI-in-education meta-analyses: none valid; 2,000+ citations combined. The report does not say AI was used in the audit.
39Matter of Newby v Adelphi Univ., 2026 NY Slip Op 26021 (Sup. Ct. Nassau County, 28 Jan 2026, Marber, J.). https://law.justia.com/cases/new-york/other-courts/2026/2026-ny-slip-op-26021.htmlTurnitin gave a student's essay a 100% AI-generated score; the court annulled the violation and ordered the record expunged ("without valid basis and devoid of reason"). Note: the ruling turned on process and the professor's own emails, not on a finding about Turnitin's accuracy.
40Curtin University (4 Sep 2025). "Update on Turnitin AI detection tool." https://www.curtin.edu.au/news/oasis-news/update-on-turnitin-ai-detection-tool/Curtin disables Turnitin's AI-writing detection across all campuses from 1 January 2026; text-matching stays
41CSET, Georgetown (7 Dec 2022). "Profile highlight: Hugging Face." https://cset.georgetown.edu/publication/profile-highlight-huggingface/More than 80,000 models on Hugging Face in December 2022 (first bar of the model-count chart)
42Mind-Verse (19 Aug 2026), citing Hugging Face on X (18 Aug 2026). https://www.mind-verse.de/news/hugging-face-hub-meilenstein-3-millionen-ki-modelleHugging Face Hub passed 3 million public models in Aug 2026; 1 million in Sep 2024
43The Register (15 Oct 2025). "Schleswig-Holstein completes move to open source email." https://www.theregister.com/2025/10/15/schleswig_holstein_open_source/40,000+ accounts moved from Exchange/Outlook to Open-Xchange and Thunderbird
44The Register (27 Jan 2026). "France to replace US videoconferencing wares with sovereign alternative." https://www.theregister.com/2026/01/27/france_videoconferencing_visio/France's state-built Visio to replace Zoom, Teams, Webex and Meet across the public sector by 2027.
45Unsloth documentation. https://unsloth.ai/docsTrains models "2× faster with 70% less VRAM".
46Baobab Tech (Oct 2026). Fine-tuned tagger experiment, internal findings (unpublished).LLM tagger about $500–720 a month at list price. A fine-tuned EmbeddingGemma-2 with LoRA (280M parameters) scored 0.648 micro-F1 against human reviewers, against 0.468 for the LLM, and put the right tag first 88% of the time against 61%. Training cost $3.26; it tags 130–700 articles a second on one GPU, against about 2 a second via API. Prices are list-price estimates, not invoices.
47Baobab Tech (2026), internal project data (unpublished)Evidence-tagging model for evaluation reports: built in under two days, 10,000x faster than the frontier-model approach, no cloud GPU, $40 of experiments, under $2 final training run
48Baobab Tech. EvalExplorer classifier fine-tuning, README and experiments dataset. https://huggingface.co/datasets/baobabtech/evalexplorer-classify-experimentsFine-tuned Qwen3.5-2B reproduces pipeline labels at 0.847 mean field score; $0.89 per fine-tune (21 min, A100); 4B model as a 2.8 GB GGUF at 0.841. Labels are silver (unreviewed LLM output).
49Baobab Tech. Water Conflict Classifier collection, Hugging Face. https://huggingface.co/collections/baobabtech/water-conflict-classifierTwo models (33.4M and 22.7M parameters); training dataset of about 3,500 rows; evals published
50Baobab Tech. water-conflict-classifier README and PyPI package. https://pypi.org/project/water-conflict-classifier/v2.5 (Dec 2025): 33M params (bge-small, SetFit), about 2,000 training examples, 81.9% exact match, F1 micro 0.866, Weapon F1 0.634; 10.6 ms per headline in demo.
51Vivek, V. (7 Jul 2026). "Microsoft pushed Copilot everywhere, but barely anyone bought it, and even fewer use it." Digital Trends, citing Fortune (29 May 2026). https://www.digitaltrends.com/computing/microsoft-pushed-copilot-everywhere-but-barely-anyone-bought-it-and-even-fewer-use-it-report/20 to 30% of licensed Microsoft 365 Copilot seats used weekly, from enterprise surveys cited by Fortune
52Baobab Tech. decision-models-experiments, README and docs/landscape.md. https://github.com/baobab-tech/decision-models-experiments255 decision models catalogued in Han Xiao's All about Jev dataset by its 27 Sep 2026 sweep, 12 days after Jev's 15 Sep launch; Baobab Tech's own zero-shot and fine-tuning experiments on evaluation-report excerpts
53Holzwarth, L., González-Márquez, R. & Kobak, D. (11 Aug 2026). "Most biomedical publications show signs of LLM-assisted writing." arXiv:2608.10715 (preprint, not peer-reviewed). https://arxiv.org/abs/2608.10715Unbiased word-frequency estimate on 1,194,287 open-access PubMed Central papers (2017-2025). Abstracts: 9% (2023), 31% (2024), 53% (2025), 68% (Dec 2025). Full papers: 19%, 52%, 77%, 89%. Supersedes the 13.5% lower bound in [37], which used a different, conservative method; the two are not directly comparable.
54Sharma, Tiwari, Kaur, Park & Pinsky (1 Sep 2026). "The water cost of intelligence." Green (MDPI) 1(2):8. https://doi.org/10.3390/green1020008. Research noteReconstructs Google's 0.26 mL per median Gemini prompt (on-site cooling only) and adds 0.471 mL of electricity-generation water for 0.725 mL total, +179%. A long Gemini prompt reaches about 28 mL. All inputs are Google's self-reported figures; nothing measured independently.

Further Reading