TrueSeeker · Verified claim report Case 8e202af46b · 2026-09-06

§ Claim under review · Mixed

"OpenAI CEO Sam Altman said that for every 38,000 ChatGPT queries, that is the same amount of water used in the production of a single almond in California" Accompanying post caption: "Sam Altman says producing one almond in California uses about as much water as 38,000 ChatGPT queries... The numbers are surprisingly close: an estimated 0.32 mL of water per ChatGPT query versus roughly 12.3 liters for one almond."

Circulating claim, as submitted.

Verdict

Source exists but framing is misleading

Confidence

High
§

Summary

Sam Altman did say this. On the September 1, 2026 Sources Podcast with Alex Heath, the OpenAI CEO said that 38,000 ChatGPT queries use about as much water as producing one California almond. The quote is accurate. The math behind it is not. Altman himself immediately added that he was unsure of the number and it might be wrong, a caveat the viral post leaves out. The 38,000 figure only works if you pair OpenAI's own unverified low-end estimate for water per query with the highest-end estimate for almonds, which counts rainfall and pollution offsets rather than actual irrigation. PolitiFact rated the claim Mostly False and found the real figure is closer to 1,000 to 10,000 queries per almond. What remains genuinely unknown is ChatGPT's actual water use, because OpenAI has never published the methodology behind its number, and independent estimates range from about 0.6 to 17 milliliters per prompt. The remarks were made while California was considering whether to require data centers to disclose their water use.

§

The readings

key figures from the evidence
38,000 queries

claimed ChatGPT queries equal to one almond's water use

1,000 to 10,000 queries

PolitiFact's estimated real ratio of queries per almond

§

Why this verdict

The attribution layer of this post is accurate. Altman said exactly this, and the video credit is correct. The substantive layer is not. The post's caption asserts that "the numbers are surprisingly close," which converts an admittedly uncertain, unverifiable, self-reported figure from an interested party into an apparently validated fact. PolitiyFact rated the underlying claim Mostly False, finding the real ratio is roughly 1,000 to 10,000 queries per almond rather than 38,000, and that Altman used a low-end AI estimate against a high-end almond estimate with mismatched water accounting boundaries. Confidence is High because the primary sources on both sides of the comparison were retrieved directly: Altman's own blog post and the Fulton et al. peer-reviewed study, plus two independent fact-checks with expert interviews. **A note on why the verdict is not "Accurate":** If the claim were only "Altman said X," the verdict would be Accurate. But the post does not merely report the quote. It endorses the arithmetic and states the numbers check out. That endorsement is the operative claim for a reasonable reader, and it does not hold. ---
§

Evidence

On attribution: Altman did say this. Multiple independent outlets, including PolitiFact and Tom's Hardware (which stated it reviewed a transcript), report the identical wording from the Sept 1, 2026 episode of the Sources Podcast with Alex Heath. The quote in the post is verbatim.

One critical omission: PolitiFact reports Altman prefaced the figure by saying he was not sure of the exact number, and that it "might be wrong, but it's close." The Instagram post does not include this hedge.

On the substance: The comparison does not survive scrutiny.

  • The ChatGPT side. Altman's 0.32 mL figure comes from his own 2025 blog post. PolitiFact reports that Altman's post is the only source it found for that number, and that the post gave no explanation of what the measurement included. OpenAI has not published data allowing verification.
  • Independent per-query estimates are far higher. A Sept 1, 2026 study estimating total water consumption including electricity generation found ChatGPT's 4o model used between 0.6 and 17 mL per prompt depending on input length. EcoLogits estimates a GPT-5.5 email at 6.11 mL. That is roughly 2x to 50x Altman's figure.
  • The almond side. The 12 L figure is real and peer-reviewed, from Fulton et al. (2019), which found California almonds averaged "10,240 liters per kilogram kernels (or, 12 liters per almond kernel)." But this is a full water footprint, including rainfall (green water) and pollution-dilution volume (grey water). PolitiFact puts the direct fresh water consumption closer to 6 L per almond. A separate USGS-affiliated estimate cited in coverage puts it near 3.56 L.
  • The resulting arithmetic. PolitiFact concludes you would need roughly 1,000 to 10,000 conversational AI prompts to equal the 6 L of fresh water used to grow one California almond. Even using Altman's own favorable 0.32 mL figure against the 3.56 L almond estimate yields about 11,000 queries, not 38,000.
  • Verifiability. CalMatters quotes UC Riverside's Shaolei Ren saying "The information we have is so limited," and reports that far too many variables (location, outside temperature, cooling system, prompt length, reasoning load, output length) prevent reduction to a single statistic.

§

Findings

What's accurate 6

  • Altman did make this statement, in these words, on the Sept 1, 2026 Sources Podcast with Alex Heath. The quote is verbatim and correctly attributed.
  • The video credit to @alexheath is correct.
  • The 0.32 mL per query figure is real in the sense that Altman published it himself in June 2025.
  • The ~12 L per almond figure is real and traces to a legitimate peer-reviewed study.
  • The internal arithmetic is consistent: 38,000 × 0.32 mL = 12.16 L, matching the ~12 L almond figure.
  • The post correctly notes that water use varies by model, data center, and cooling system, and that the almond figure is a broader water footprint.

What's misleading 8

  • **Omitted qualifier (the speaker's own hedge).** Altman said the number might be wrong. The post presents it as a confident assertion in a headline card. The most important caveat came from Altman himself and was dropped.
  • **Marketing as evidence.** The caption's line "The numbers are surprisingly close" presents an unverified, self-reported, self-serving corporate estimate as if it were independently validated. The two numbers are close only because one of them was reverse-engineered from the other.
  • **Circular reasoning presented as convergence.** 38,000 × 0.32 mL happens to equal ~12 L. That is not two independent measurements agreeing. It is one input producing one output.
  • **Apples-to-oranges water accounting.** Comparing a full agricultural water footprint (rainfall plus irrigation plus pollution dilution) against a narrow operational data center figure is not a like-for-like comparison. The post's caveat that the almond figure is "a broader water footprint" gestures at this but does not explain that it invalidates the comparison.
  • **Exaggeration by source selection.** Altman used a low-end AI estimate against a high-end almond estimate. PolitiFact's range of 1,000 to 10,000 queries per almond means the real ratio is roughly 4x to 38x smaller than claimed.
  • **Omitted context: unverifiable source.** OpenAI has published no methodology. The post presents the figure without noting that no independent party can check it.
  • **Omitted context: motive and timing.** The remarks were made amid an active California policy fight over mandatory data center water disclosure. The post presents them as a neutral observation.
  • **Water type omission.** Almond irrigation water is generally not treated municipal drinking water. Data center cooling water often is. The post's framing invites a comparison that the underlying resource types do not support.

? What's uncertain 6

  • The full podcast transcript was not retrieved directly. The quote is confirmed by multiple independent outlets including one that stated it reviewed a transcript, but the primary recording was not accessed in this investigation.
  • The true per-query water consumption of ChatGPT cannot be established. OpenAI has not published sufficient data, and independent estimates span roughly 0.6 mL to 17 mL depending on model, prompt length, and whether electricity generation is counted.
  • Whether Altman's 0.32 mL includes water used for electricity generation is unknown.
  • The "12.3 liters" figure in the Instagram caption is slightly different from the 12 L in Fulton et al. The precise source of the .3 decimal was not located.
  • Shaolei Ren's peak-demand research is a preprint under review and has not completed peer review.
  • Almond water use has reportedly declined since the 2004 to 2015 study period. The current per-almond figure may be lower than 12 L. ---
Distortion flags omitted qualifier marketing as evidence exaggeration
§

Sources

1 of 8 linked to records
[1]

**Sam Altman, "The Gentle Singularity" (blog.samaltman.com, June 2025)**

primary First-party statement by the claimant. Confirms Altman's own published per-query figure of 0.000085 gallons (~0.32 mL) of water.
This citation could not be independently verified.
[2]

**Fulton et al., "Water-indexed benefits and impacts of California almonds," *Ecological Indicators* (2019), ScienceDirect**

primary Peer-reviewed study. Source of the 12 L per almond figure.
Registry-verified authors: Fulton J, Norton M, Shilling F
https://doi.org/10.1016/j.ecolind.2017.12.063 ↗
[3]

**PolitiFact fact-check, Sept 4, 2026**

secondary Quality journalism citing primary sources and an expert interview with Shaolei Ren. Rating: Mostly False.
This citation could not be independently verified.
[4]

**CalMatters fact-check, Sept 3, 2026 (syndicated to Palo Alto Online, GV Wire, and others)**

secondary Nonprofit newsroom, direct expert interview with Shaolei Ren.
This citation could not be independently verified.
[5]

**Tom's Hardware, Sept 2026**

secondary Reports reviewing a transcript of the podcast episode, confirming the quote wording.
This citation could not be independently verified.
[6]

**Almond Board of California, industry commentary on the Fulton study**

unknown Notes the water footprint figure is model-based, not grower-reported.
This citation could not be independently verified.
[7]

**Tom's Guide, IBTimes UK, Storyboard18, HypeFresh**

secondary Coverage recomputing the arithmetic; all trace back to sources 1 through 3.
This citation could not be independently verified.
[8]

**The Instagram post under investigation**

tertiary Aggregating a viral video clip. Not evidence. ---
This citation could not be independently verified.
How links are chosen. A source is linked only when the address comes from the investigation's own retrieval or from a registry lookup (PubMed, Crossref) that matches the citation's title and year. Author lists shown as registry-verified come from the registry record, not from the report text. Citations that cannot be matched are labeled, never guessed.
This is one case on the record See the full case, browse the archive, and search every checked claim on TrueSeeker Open on trueseeker.com →