Why is Everyone So Wrong About AI Water Use??
Credibility score: 43/100 — Mixed Credibility. Several questionable claims detected. Watch with healthy skepticism.
BSmeter analyzed "Why is Everyone So Wrong About AI Water Use??" and rated it 43/100 for credibility (a BS score of 57/100 — mixed credibility), on 2026-07-13. Its weakest claim — "Corn beats AI water use — false equivalence framing" — scored 20/100 and was flagged as false equivalence. 36 claims were checked against the video transcript. Scores are produced by BSmeter's AI analysis of the transcript, not independent human verification.
Of 36 claims analyzed: 9 scored under 40, 27 between 40 and 69, and 0 at 70 or above.
Claims analyzed
Corn beats AI water use — false equivalence framing — False Equivalence (20/100)
At 0:00
Compares total US corn irrigation to global AI water — apples to oranges by design.
Why this score: False equivalence 101: one is a single crop grown across a continent for food and ethanol, the other is a new, rapidly scaling industry. The numbers may be accurate, but the framing implies 'AI isn't a big deal because corn is worse' — which dodges whether AI's growth trajectory is sustainable on its own. Fair framing would compare AI to other data centers or tech uses, not to agriculture.
Original quote: “AI data centers are going to use a lot of water, but they are not going to use anything like as much water as the weirdest industrial use of water that we have. It is so weird. I hate it so much. Corn is one of the thirstiest major crops grown in the US.”
Opens by contrasting tiny per-query figure with trillion-liter projection — framing mismatch setup — Missing Context (45/100)
At 0:00
Sets up 'how can both be true?' tension before explaining why the numbers aren't actually contradictory.
Why this score: By placing the individual query stat right next to the aggregate projection without any bridging info, the speaker creates the impression of a paradox. This is classic 'missing context' framing — the viewer is meant to feel confused first so the later explanation lands harder. A neutral version would have introduced the scale difference immediately instead of letting the numbers clash.
Original quote: “Sam Alman says that the average chat GBD query uses around 0.000085 gallons of water or roughly 115eenth of a teaspoon. But then at the same time, somehow a Morgan Stanley projection predicted annual water use for cooling and electricity generation by AI data centers could reach around a,000…”
Two stats presented as contradiction — framing mismatch — Missing Context (45/100)
At 0:00
Per-query number vs total industry projection can't cancel each other — different scales entirely.
Why this score: This is a classic apples-to-oranges setup. One figure is per individual query, the other is cumulative across every data center's cooling and power generation. By placing them side-by-side without explaining the difference in scope, the speaker creates the impression that at least one number must be wrong, when both can easily be accurate at once. A neutral framing would label one as 'per query' and the other as 'industry-wide annual total' so viewers don't feel like they're being shown a paradox.
Original quote: “Sam Alman says that the average chat GBD query uses around 0.000085 gallons of water or roughly 115eenth of a teaspoon. But then at the same time, somehow a Morgan Stanley projection predicted annual water use for cooling and electricity generation by AI data centers could reach around a,000…”
Headline framing: "blocking" + "safety fears" loads the angle — Loaded Language (35/100)
At 2:10
Calls it "blocking" state rules "despite safety fears" — steers you before the facts.
Why this score: This is classic loaded language. The word "blocking" implies obstruction rather than creating a single national standard, and tacking on "despite safety fears" frames the move as reckless. A neutral version would say "creating one national AI rule to replace varying state laws" without injecting the safety judgment.
Original quote: “Trump says that he will sign executive order blocking state AI regulations despite safety fears.”
"Bad actor states" + Congress failure — partisan framing — Loaded Language (30/100)
At 2:10
Labels states as "bad actors" and pins blame on Congress — clear partisan spin.
Why this score: The right-leaning version uses "bad actor states" to paint state regulations as the problem and Congress as negligent. This is emotional button language that turns a regulatory conflict into a morality play. A neutral headline would just describe the executive order replacing inconsistent state rules without assigning villains.
Original quote: “Trump to issue one rulebook AI executive order after Congress fails to curb bad actor states.”
"Many" vs "most" — vague scale without numbers — Missing Context (50/100)
At 3:48
Says "many" recycle but "most" use fresh water — no percentages given.
Why this score: The speaker contrasts recycling with municipal water use but never quantifies either. Without actual shares or facility counts, listeners can't tell if recycling is rare or common. A clearer framing would cite the percentage of data centers using each method so the scale is visible.
Original quote: “Many facilities use evaporative cooling where they literally turn clean water into water vapor... most of them still rely on fresh municipal water.”
OpenAI silent on water data — pivots to 'guesses' instead of evidence — Missing Context (45/100)
At 5:00
Frames lack of data as something he can 'guess' at — skips that companies often withhold location-specific data for competitive reasons.
Why this score: By saying 'I can make some guesses' right after noting OpenAI's silence, the speaker implies the numbers are guessable from public info. Missing context: water use varies wildly by data center location, cooling method, and local climate — none of which public data reveals without company disclosure. A neutral framing would note the data gap and stop there instead of signaling that educated guesses can fill it.
Original quote: “OpenAI did not release information on where the water use figures come from, but I can make some guesses.”
Sam's 'per query' number hides reasoning chains — classic hidden multiplier framing — Missing Context (55/100)
At 5:00
Calls out hidden follow-up queries — correct point, but never shows how much they actually multiply the original number.
Why this score: The speaker correctly identifies that reasoning models run internal queries. However, he immediately guesses 'multiply by 15' without evidence, then says 'even then it's still a teaspoon.' This is missing context: without actual OpenAI telemetry, the multiplier claim is speculation presented as likely. A cleaner frame would say 'we don't know the multiplier' rather than offering a dramatic but unverified number.
Original quote: “First, Sam is talking about the water use per query. But importantly, different queries work different ways with AI, and many queries will actually result in multiple queries you never even see.”
Calls Sam Altman's number a 'lie' — framing it as deliberate omission — Loaded Language (35/100)
At 5:00
Labels disagreement a 'lie' before showing evidence — emotional button, not just disagreement.
Why this score: The speaker escalates a measurement disagreement into an accusation of lying. This framing makes the audience react to intent rather than the actual technical question of how to allocate training water across queries. A neutral version would say the numbers use different system boundaries instead of calling one side dishonest.
Original quote: “me not really the lie in Sam's lie here. And I am going to call it a lie because I think it is. The actual lie is that it is only counting the water used during the part of the life cycle when you're actually querying the chatbot, not all of the other parts of the process that are necessary for…”
Describes weeks/months of GPU clusters 'burning through' water — emotional button on scale — Emotional Button (40/100)
At 5:29
'Burning through' paints continuous destruction without context on total water volume or efficiency gains.
Why this score: The vivid language makes the audience visualize waste. Fact-checks show the real issue is local water stress and how cooling is managed, not raw 'burning.' The framing skips that many data centers use recycled or non-potable water and that efficiency has improved, steering the reaction toward outrage rather than the nuanced local-impact picture in the sources.
Original quote: “Even while we're using one model, like GPD5, which is actually a bunch of models all stuck together, OpenAI and its competitors are constantly training newer, bigger versions that no one can use yet. And to create these models like the system runs for weeks or months on enormous clusters of GPUs…”
Argues training must be included in every query — framing choice as honesty test — Missing Context (55/100)
At 5:55
Presents one allocation method as the only honest one, ignoring that different boundaries answer different questions.
Why this score: The speaker frames including training as the morally correct choice. But sources note all numbers are correct within their scope; a per-query amortized number versus an inference-only number serve different policy uses. Calling one 'honest' and the other a lie is itself a framing move that hides the legitimate debate over what question you're trying to answer.
Original quote: “It's not really fair to treat that training footprint as separate from every conversation you have with the model. The conversation could not happen without the training. So if you wanted to be honest, you got to make some choices.”
Sets up 'training is the biggest missing piece' — missing context on boundaries — Missing Context (50/100)
At 6:20
Teases the 'biggest' omission without naming how big training actually is compared to inference.
Why this score: The speaker promises a major revelation but doesn't yet quantify training versus inference. Web sources note that all numbers can be correct within their own boundaries; without showing the actual split, the audience can't judge whether excluding training is a lie or a different scope. This leaves the framing of 'they're hiding the real number' hanging.
Original quote: “This is the big difference between Morgan Stanley's analysis and Sam Altman's. There are a lot of other parts of this process that are not the final step of the consumer querying the chatbot. And you've probably thought of one of them, but you probably haven't thought of the biggest one.”
Sam's number is a 'lie' because it only counts inference — training and hardware omitted — Loaded Language (30/100)
At 6:20
Labels Altman's per-query figure a 'lie' for omitting training — ignores that per-query metrics are standard industry practice.
Why this score: Calling it a 'lie' is loaded. Most environmental analyses of consumer tech (search, streaming, email) report per-use impact separately from manufacturing or training. The speaker is right that training water use exists and is large, but the rhetorical move of calling a standard per-query metric a deliberate lie misleads viewers about what the number was ever intended to represent. A neutral version would say 'this figure deliberately excludes training-phase water use' instead of 'this is a lie.'
Original quote: “The actual lie is that it is only counting the water used during the part of the life cycle when you're actually querying the chatbot, not all of the other parts of the process that are necessary for this system to exist.”
Claims training 'never really stops' — volume game on ongoing cost — Volume Game (45/100)
At 6:49
Uses 'never really stops' to imply constant massive water draw, without showing the actual cadence or scale.
Why this score: The phrasing suggests perpetual high-intensity training. In reality training runs are episodic and get amortized over billions of queries. The speaker correctly notes training matters but inflates the sense of continuous resource burn by not clarifying frequency or duration, making the water impact feel larger than the data supports.
Original quote: “and I want to do a test right now and find out if I'm right. So, if you thought of one place where the chatbot is using water that isn't the query, you probably thought of the training. The thing about these systems is that the training never really stops.”
Blames OpenAI secrecy for conflicting numbers — anonymous authority on 'lying from either direction' — Anonymous Authority (50/100)
At 7:39
Says 'it's so easy to lie' without naming who is actually lying or showing examples.
Why this score: The speaker correctly notes lack of transparency, but then broadens it to 'easy to lie from either direction' without evidence. This lets the audience assume bad faith on all sides. A cleaner frame would stick to the transparency problem and skip the moral claim about lying, since the sources show honest analysts still disagree on boundaries even with full data.
Original quote: “problem here is no one knows how to do that accurately because OpenAI doesn't share this information, which is part of why it is so easy to get numbers that are both fairly correct and very different from each other. And part of why it's so easy to lie about this from either direction, by the way.”
Sam Altman 'decided' to exclude training water — loaded intent framing — Loaded Language (35/100)
At 8:29
Turns a boundary choice into 'Sam decided to hide it' — implies motive without evidence.
Why this score: The phrasing shifts from 'different analysts draw different boundaries' to 'Sam personally chose to exclude training.' A cleaner frame would say 'OpenAI's reported figure covers inference only' and let listeners decide if that's reasonable.
Original quote: “Sam seems to have decided to not include it when discussing the water use of a chat GPT query. Because look, it is true that any individual query doesn't use that water.”
Cites UC estimate of 50% training share — cherry-picked to contrast with Sam — Cherry-Picked (60/100)
At 8:29
Drops the 50% figure without source year or whether it includes only training vs full lifecycle.
Why this score: The number lands with authority but the speaker doesn't clarify the study details or whether it matches the exact models being discussed. Web sources stress that context changes everything; using a single round number to imply Sam is obviously wrong skips the step of showing whether the 50% applies to the current GPT generation or older models. The framing makes the omission look indefensible rather than debatable.
Original quote: “According to estimates from the University of California, the training of AI models can account for around 50% of their total resource use. The training is big. And weirdly, Sam seems to have decided to not include it”
UC estimate: training = ~50% of AI resource use — cites anonymous study — Anonymous Authority (45/100)
At 8:30
Drops 'University of California' like a mic — zero paper, year, or author named. Classic authority move.
Why this score: By invoking a prestigious institution without the actual source, the speaker borrows credibility while dodging verification. A neutral framing would name the specific study or at least the department so viewers could check the 50% figure themselves.
Original quote: “According to estimates from the University of California, the training of AI models can account for around 50% of their total resource use.”
Pivots to power-plant water use — volume game to reframe total footprint — Volume Game (50/100)
At 9:50
Suddenly widens the lens to thermoelectric plants — makes AI's share feel smaller by comparison.
Why this score: This is the classic 'look at the bigger picture' move. It correctly notes that electricity generation consumes water, but doesn't clarify how much of that water is attributable to AI versus everything else on the grid. The neutral alternative would quantify the marginal increase caused by AI demand rather than just gesturing at power plants.
Original quote: “We know the data centers use lots of water, but they also use a lot of electricity. And you know what else uses A LOT OF WATER? POWER PLANTS.”
$100B+/yr on data centers by three companies — big number, no source — Anonymous Authority (40/100)
At 9:50
Hundred billion is dropped like common knowledge — no breakdown, no year, no report cited.
Why this score: The scale is meant to stun, but without a traceable source the number floats. A transparent version would say 'according to X report from Y date' so the viewer knows whether it's capex, opex, or total spend across all three firms.
Original quote: “Microsoft and Google and Amazon alone are collectively spending more than hundred billion dollars a year on new data centers.”
Drops 40% stat like a mic — classic volume game framing — Volume Game (45/100)
At 10:16
Uses a huge-sounding percentage without clarifying it's mostly intake-and-return, not consumption.
Why this score: The speaker correctly cites USGS but then immediately flags the return-flow issue, so the framing is self-correcting. Still, dropping the raw 40% first makes the number feel scarier before the nuance arrives. A neutral version would lead with '40% of withdrawals, most of which is returned.'
Original quote: “electricity generation accounts for, get this, 40% of all freshwater withdrawals in the United States”
Accuses critics of inflating numbers by counting returned water — missing context — Missing Context (55/100)
At 10:44
Frames opponents as deliberately choosing the bigger number without noting some studies do track total withdrawals for a reason.
Why this score: The speaker is right that many analyses count intake rather than net consumption, but the implication that this is only done to make AI 'look really massive' skips the legitimate policy argument: even returned water still requires infrastructure, affects local ecosystems, and competes with other uses during shortages. A fairer frame would acknowledge both the return-flow reality and why withdrawals still matter to regulators.
Original quote: “this is the water that people are sometimes counting if they want AI water use to look really massive”
Uses lawn-watering example to argue industrial vs municipal water are fundamentally different — False Equivalence (40/100)
At 12:55
Compares two very different end-uses to imply data-center cooling is harmless — the analogy hides competition for the same limited watershed.
Why this score: By contrasting 'raw' river water with treated municipal water, the speaker creates a clean split that ignores the real constraint: total water available in that watershed. A data center pulling from the same lake still reduces supply for other downstream users, even if it never touches a treatment plant.
Original quote: “Taking water from a river or a lake is different from taking drinking water from a treatment plant. Putting drinking water on your lawn is actually cheaper per gallon for the water company than using it in your sink or shower because it doesn't need to go through the wastewater treatment process.”
Only experts need this detail — False Dilemma on public understanding — False Dilemma (40/100)
At 13:45
Frames knowledge as all-or-nothing: either you're an expert or you don't need it.
Why this score: Creates a binary where non-experts are either fully informed or should stay ignorant. A more accurate framing would note that basic context (energy intensity, volume share) can be made accessible without requiring full expertise, which is exactly what the speaker is attempting in the video.
Original quote: “And that's not something that like anybody needs to know unless they're experts working on this problem.”
Ultra-pure fab water is "very small" but harder — Missing Context on energy cost — Missing Context (45/100)
At 14:11
Says fab water is "very small" then immediately emphasizes difficulty — the contrast does the heavy lifting.
Why this score: Speaker sets up a quantity-versus-quality tension but never gives the energy numbers for distillation. A neutral version would state both the volume ratio and the kWh-per-liter difference so viewers can weigh the two factors themselves instead of being steered by the word "harder."
Original quote: “Now, when you compare that to the water used for cooling, it's very small, but it's a different kind of water. It's not just not lake water. This has to be ultra pure water that has like nothing in it. It's like extremely distilled to make sure that there's nothing in it that could possibly mess up…”
Chip manufacturing uses "huge" water — Missing Context on scale vs cooling — Missing Context (45/100)
At 14:11
Calls fab water "huge" without numbers — leaves readers guessing how it compares to cooling.
Why this score: Speaker contrasts manufacturing with cooling but gives no actual volumes. A fair framing would include the specific gallons-per-chip or percentage of total AI water use so the "huge" label can be judged. Without that anchor, the word does the work of making fab water feel dominant even though the speaker later says it's a small fraction.
Original quote: “A great example of this, a lot of water is actually used to make the chips that run AI data centers like those Nvidia GPUs. A huge amount of water is necessary for the creation of those and that is part of the life cycle.”
Fab water needs "way more energy" — Anonymous Authority on the comparison — Anonymous Authority (35/100)
At 14:36
Claims "way more energy" with zero source or number — classic anonymous authority move.
Why this score: Speaker presents an energy comparison as settled fact while admitting he knows "next to nothing." A cleaner frame would cite the actual energy intensity data or acknowledge the range of estimates. Without it, the phrase functions as an unsupported intensifier rather than evidence.
Original quote: “So, it's a very small fraction of the amount of water used in the process of AI existing. But that water is harder to produce than drinking water. It requires way more energy to produce than drinking water.”
Water stats easy to mislead with — Loaded Language on both sides — Loaded Language (55/100)
At 15:27
Uses "really easy to mislead" twice — paints all opposing framings as deliberate distortion.
Why this score: Speaker correctly flags cherry-picking risks but applies the loaded phrase symmetrically to both under- and over-counting. A neutral version would distinguish between honest framing differences and actual misrepresentation instead of implying bad faith on either side.
Original quote: “My point here overall is that this is complex and that means to me two things. One, it's really easy to mislead about it. It's easy to misrepresent by leaving out training and making the water use seem really small. It's easy to misrepresent by including water flowing through a power plant to make…”
Calls including power-plant water a deliberate choice to inflate figures — loaded framing — Loaded Language (40/100)
At 15:54
Labels the inclusion as a manipulative 'choice' while downplaying that thermoelectric water use is a real, measurable impact of electricity demand.
Why this score: By repeatedly saying 'if they want AI water use to look really massive,' the speaker implies bad faith rather than acknowledging that some analysts include the full energy-water chain precisely because data centers increase electricity demand. The alternative framing would be: 'Some studies attribute a share of power-plant water use to AI; others don't — here's why both approaches exist.'
Original quote: “including the power plant water in the calculation for AI data centers is a choice that can make the numbers look really big”
Only 1% of corn eaten by humans — straw man setup — Straw Man (30/100)
At 18:41
Sets up an easy "but we eat it" objection then knocks it down — classic straw man.
Why this score: The speaker invents a defense no one actually made (that corn is mainly for direct human eating) then dismantles it. Real counterarguments are about livestock feed and biofuels, which he later addresses — so the initial straw man feels like rhetorical setup rather than genuine engagement.
Original quote: “And I totally forgive you if you are thinking right now, "Okay, Hank, yes, but corn is food. We eat it. Food is very important for people." But that's the thing. We don't eat it. Maybe 1% of corn is eaten by humans. A lot of it is eaten BY LIVESTOCK. BUT 40% OF it is burned in our cars and trucks.”
Frames AI water use as moral panic vs. everyday habits — Emotional Button (45/100)
At 18:41
Sets up personal guilt (sink brushing) against industrial scale — classic emotional button to soften the industrial issue.
Why this score: By anchoring the moral feeling in something tiny and familiar (turning off the tap), the speaker makes the jump to AI feel disproportionate. A neutral framing would compare water use directly to other industrial or agricultural sectors without the personal-morality layer.
Original quote: “And I get why people jump on AI water use. Wasting water feels immoral. We are told our whole lives to turn off that sink while we brush.”
Agricultural water use reframed as 'normal' — Missing Context — Missing Context (45/100)
At 19:12
Calls farm irrigation 'normal' while skipping that it's the dominant water consumer — shifts focus off ag.
Why this score: By labeling agricultural use as something we've 'learned to see as normal,' the framing makes industrial and AI water use look relatively small. The missing piece is that agriculture already accounts for the vast majority of freshwater withdrawals in the US and globally, so calling it 'normal' hides how extreme the baseline actually is.
Original quote: “They can't have new uses. Others have room. Some uses like irrigating the entire cornbt involve staggering amounts of water that we've just learned to see as normal.”
Calls Tucson data center 'dumb' — emotional button on water scarcity — Emotional Button (45/100)
At 20:36
Uses 'dumb' twice to trigger moral reaction instead of showing water data.
Why this score: The speaker skips any comparison to existing Tucson water users and goes straight to moral language. A neutral framing would name the percentage of local supply a data center would actually take versus agriculture or golf courses.
Original quote: “Building a water cooled data center in Tucson is just very dumb. It's a bad idea. It's dumb as”
Calling a water-cooled data center in Tucson 'dumb' — using loaded language to dismiss a specific choice. — Loaded Language (45/100)
At 20:36
Uses strong, dismissive language ('dumb', 'bad idea') rather than objective analysis of the challenges.
Why this score: The speaker uses emotionally charged words like 'dumb' to frame the decision to build a water-cooled data center in Tucson as obviously foolish. While the location (Tucson, a desert climate) does present clear water scarcity challenges, framing it with such dismissive language rather than a more neutral 'ill-advised' or 'problematic due to water constraints' steers the audience's emotional response. A neutral framing would simply state the facts about water availability and the center's needs.
Original quote: “Building a water cooled data center in Tucson is just very dumb. It's a bad idea. It's dumb as”
Frames water as 'small' vs power as 'order of magnitude larger' — comparative scale framing — Missing Context (45/100)
At 21:04
Uses 'maximalist goals' to bound the comparison, then pivots to power without numbers.
Why this score: He's setting up a relative scale argument: water use stays small even in worst-case AI growth. The missing piece is what those 'maximalist' numbers actually are — without them the claim that water is 'small' rests on an unshown baseline. The power comparison then feels bigger by default.
Original quote: “Even under the maximalist goals of AI companies, the projected increase in water use is small compared to what cities and industries already use. What I am much more worried about is that the projected increase in power demand is not small.”
Corn ethanol history proves lying about AI water is easy — false equivalence — False Equivalence (25/100)
At 23:41
Links corn ethanol's hype cycle to AI water claims — assumes same incentive structure applies.
Why this score: This is a false equivalence move. The speaker uses the well-documented corn ethanol subsidy story as proof that AI water figures are similarly distorted, without showing parallel mechanisms (government mandates, farm lobbies, specific reporting failures) for data centers. The analogy lands emotionally but skips the actual data transparency differences between agricultural policy and private AI infrastructure.
Original quote: “But yes, in conclusion, resource analysis is complex. The incentives are weird and we have a very long history of underestimating how dumb corn ethanol is. And all of that combined means that it is very easy to lie about AI water use.”
See the full analysis with timestamps →