A Defensible Case on AI

Four short chapters for teachers and anyone deciding what to make of AI: what a prompt really drinks, what a randomized trial has already measured, who ends up holding a tool when careful people walk away, and what the early evidence on thinking does and does not show. The method is the briefing's own: lead with the fear you share, concede what is true, then measure what is contested. Every number spoken or shown traces to a source listed below.

The full cut · about 6½ minutes · 1920×1080 · captions burned in

01 Five drops

Water: the prompt, the comparisons, the one aquifer

A prompt is a few drops; the real water question is one aquifer at a time.

02 What it caught

The randomized screening trial, with its footnotes

The clearest benefit is already measured in a randomized trial: AI-supported screening missed no more cancers than the standard double read, with far less reading and a human still reading.

03 Who holds it

Dual use: the asymmetry of walking away

Walking away doesn't remove the tool; it removes the careful hands from it.

04 Augment or atrophy

Offloading judgment, and the evidence so far

Offloading is old; offloading judgment is new. The variable is how you use it.

Sources and the claims ledger

Each row is a number or named study the chapters use. "Contested" rows carry their concession here and are spoken with it in the video. References 25 and 26 were added at verification because two spoken figures trace to the paper and the analysis themselves rather than to the briefing's list. The full ledger with its external verification notes, the scripts and the storyboards are on the review page.

RowClaimConfidenceRefsChapter
W1Google: median Gemini text prompt ≈ 0.26 mL of water (about five drops)verified against the source[1]01
W3aUC Riverside, U.S. average per request: 2.2 mL on-site + 14.7 mL for electricity ≈ 16.9 mL total (range ≈ 7–48 mL by location)verified against the source[3] [25]01
W4UC Riverside: on-site cooling ≈ 2.2 mL per query (U.S. average data center)verified against the source[3]01
W6About 80 % of the water attributed to AI is off-site power-plant water ("a gallons headline can be ten times" the water consumed inside)verified against the source[1] [2]01
W6bThe off-site (power-plant) water still counts; it just is not the town's aquifer (the briefing's "much of it non-consumptive" is not restated)
Concession: Most of that water is used far away, at power plants, not drawn from the town's aquifer; it still counts, it just isn't local.		contested; stated with its concession[1] [2]01
W10One almond ≈ one gallon of irrigation water (estimates ≈ 0.75–1.7)verified against the source[6] [7]01
W13U.S. data centers ≈ 50 million gal/day on-site; AI ≈ 15–20 % of it ≈ 10 million gal/day
Concession: That's on-site cooling only. Count the power plants that run them and data centers use roughly thirteen times more.		contested; stated with its concession[9] [10]01
W14U.S. golf courses ≈ 2 billion gal/day of applied irrigation; against AI's on-site cooling that is ≈ 200 to 1, and counted alike (consumption on both sides) ≈ 8 to 1
Concession: Two hundred to one sets golf's applied irrigation against AI's on-site cooling; counted alike, consumption on both sides, golf's lead is about eight to one.		contested; stated with its concession[9] [26]01
W15Meta's Hyperion (Richland Parish, Louisiana): registered for > 23 million gal/day (8.4 billion gal/yr)verified against the source[10] [11]01
W16Meta projects 500–600 million gal/yr (≈ 1.5 million gal/day), closed loop, from the Mississippi River Alluvial Aquiferverified against the source[10] [12]01
W17LSU's 17-year simulation at the maximum registered withdrawal: groundwater drops more than 65 ft beneath parts of the siteverified against the source[10] [11]01
W18No regulatory body monitors Meta's actual water use or the aquifer's conditionverified against the source[10]01
B1MASAI: 105,934 women randomized; the first randomized trial of AI in breast-cancer screening (Sweden, Lund University, Kristina Lång)verified against the source[14] [15]02
B212 % fewer interval cancers with AI-supported screening (The Lancet, January 2026)
Concession: Twelve percent fewer, though the trial reports it as no worse than two radiologists, not as a proven reduction.		contested; stated with its concession[14] [16]02
B327 % fewer aggressive (non-luminal A) interval cancers
Concession: A small count, forty-three against fifty-nine, reported descriptively.		contested; stated with its concession[14] [16]02
B544 % less screen-reading workload for radiologists (Lancet Oncology 2023 interim safety analysis)verified against the source[16] [17]02
B6The trial measured detection and interval-cancer rates, not deathsverified against the source[14] [16]02
B7The result does not support replacing radiologists; AI-supported screening still needs at least one human readerverified against the source[18]02
B8Weight the peer-reviewed Lancet paper, not the vendor's press releasecarried from the briefing's source[16]02
D1Encryption: the 1990s fight ended with strong encryption everywhere, not a banverified against the sourcebriefing03
C1Plato's Phaedrus: writing "will produce forgetfulness" and offer "the appearance of wisdom"verified against the sourcebriefing04
C2The "Google effect" (Sparrow, Liu & Wegner, Science, 2011)
Concession: The original 2011 experiment did not replicate in 2018.		contested; stated with its concession[19]04
C3The photo-taking-impairment effect (Henkel, Psychological Science, 2014)verified against the source[20]04
C5Habitual GPS use and worse spatial memory (Dahmani & Bohbot, Scientific Reports, 2020)verified against the source[21]04
C7MIT Media Lab EEG study (Kosmyna et al., 2025): 54 people writing essays; weakest connectivity with ChatGPT; "cognitive debt"carried from the briefing's source[22] [23]04
C8Microsoft Research & Carnegie Mellon survey of 319 knowledge workers (CHI 2025): more trust in AI, less critical thinkingverified against the source[22] [23]04
C10Reviews and scaffolded-prompting studies: AI can improve critical thinking and problem-solving when used actively rather than passively; the form of integration matters more than the toolcarried from the briefing's source[22] [23]04
S1Three distinctions (water): consumptive vs non-consumptive, direct vs indirect, local vs aggregatethe briefing's own framingbriefing01
S2Three caveats (MASAI)the briefing's own framingbriefing02
S3Four distinctions (cognitive offload)the briefing's own framingbriefing04
S5Three waters (evaporated on site · returned to the river · used at the power plant): the chapter's own framing, derived from the briefing's consumptive/non-consumptive and direct/indirect distinctionsthe briefing's own framingbriefing01
C11Arithmetic tools (abacus, slide rule, calculator, spreadsheet): each drew the panic, each was fine once number sense was in place and harmful when it replaced building itestablished literaturebriefing04

References (numbering follows the briefing)

  1. [1] Data Center Dynamics, "Google: Median Gemini prompt uses 0.24 watt hours of power and consumes 0.26ml of water" (2025). source
  2. [2] Andy Masley, blog post on AI water use and the direct/indirect split (2025) source
  3. [3] The Register, "Google games numbers to make AI look less thirsty" (discussion of the UC Riverside 2023 study *Making AI Less Thirsty* and the ~2.2 ml on-site figure) (2025). source
  4. [6] Pacific Institute, "The California Drought: Almonds and the Bigger Picture" (520–560 gal/lb; 1.6–1.7 gal/nut). source
  5. [7] Farmonaut, "Amount of Water to Produce Almonds: California Almond Water Use" (~1.1 gal/nut; 1,900–2,100 gal/lb) (2025). source
  6. [9] Understanding Your AI, "Does AI Really Use 10 Gallons of Water Per Image?" (US data centers ~50M gal/day on-site; AI ~10M gal/day; golf ~2B gal/day) (2026). source
  7. [10] NOLA.com, "How much water will Meta's Louisiana data center use? New details unveiled for first time" (2025). source
  8. [11] Richland Today, "How much water will data center use?" (LSU 17-year simulation; 65-foot drawdown at maximum permit) (2026). — link dead at verification 2026-10-06; the same finding is in [10]
  9. [12] WBRZ, "Meta's Louisiana data center causing concern over water usage" (500–600M gal/year projected; closed-loop; Mississippi River Alluvial Aquifer) (2026). source
  10. [14] The ASCO Post, "Randomized Trial Shows AI-Supported Mammography Improves Sensitivity and Lowers Interval Cancer Rate" (105,934 randomized; 19 excluded) (2026). source
  11. [15] Lund University, "AI support in breast cancer screening: Fewer missed cancer cases" (2026). source
  12. [16] ecancer, "AI-supported mammography screening results in fewer aggressive and advanced breast cancers" (12% interval, 27% aggressive; published in *The Lancet*) (2026). source
  13. [17] EurekAlert / The Lancet, "AI-supported mammography screening results in fewer aggressive and advanced breast cancers" (44% workload reduction, 2023 Lancet Oncology; 29% detection increase, Lancet Digital Health) (2026). source
  14. [18] Gommers, J., Lång, K., et al., "Interval cancer, sensitivity, and specificity ... in the MASAI study" (first author's note that AI does not replace radiologists), ScienceDirect. source
  15. [19] Sparrow, B., Liu, J., & Wegner, D.M., "Google Effects on Memory: Cognitive Consequences of Having Information at Our Fingertips," *Science* (2011). source
  16. [20] Henkel, L.A., "Point-and-Shoot Memories: The Influence of Taking Photos on Memory for a Museum Tour," *Psychological Science* (2014). (confirmed 2026-10-06: PubMed 24311477)
  17. [21] Dahmani, L. & Bohbot, V.D., "Habitual use of GPS negatively impacts spatial memory during self-guided navigation," *Scientific Reports* (2020). source
  18. [22] EDUCAUSE Review, "The Paradox of AI Assistance: Better Results, Worse Thinking" (overview of the MIT Media Lab EEG study and the Microsoft/Carnegie Mellon survey) (2025). source
  19. [23] The Algorithmic Bridge, "What the Studies Say About How AI Affects Your Brain" (compilation including Kosmyna et al. 2025 MIT preprint, Lee et al. CHI 2025, and a preregistered retention RCT) (2026). source
  20. [25] Li, P., Yang, J., Islam, M.A., Ren, S., "Making AI Less 'Thirsty': Uncovering and Addressing the Secret Water Footprint of AI Models" (arXiv 2304.03271, v5; the per-request table: 2.2 mL on-site + 14.7 mL for electricity ≈ 16.9 mL U.S. average, 7–48 mL by location). (added at verification) source
  21. [26] Napkin Quest newsletter, "Golf uses more water than data centers, by 2x or…" (golf's consumed water ≈ 425 billion gal/yr vs data centers' total footprint; AI's share ≈ 52 billion gal/yr → ≈ 8 : 1 counted alike). (added at verification) source

The full ledger with its external verification notes, the scripts and the storyboard sheets: guides.letsharkness.com/ai-case/review/

Made with HyperFrames (open source, HeyGen). Narration synthesized locally with Kokoro. Figures as of mid-2026; several rest on contested or fast-moving estimates, and the ledger says which.