Teaching Notes: Faithful vs Traitor Exemplar Reports
Instructor guide for the Week 4 reveal discussion
Purpose
These notes accompany the two exemplar policy briefings on the environmental consequences of AI:
- Faithful version (
faithful.qmd): “How bad is AI for the environment?” - Traitor version (
traitor.qmd): “The hidden environmental cost of AI”
Both reports use the same underlying data sources. Every number in the traitor version is traceable to a real source and is technically correct. The deception is entirely in selection, framing, and omission. The purpose of this exercise is to help students recognise that data can be weaponised without ever being fabricated — and that the most dangerous misinformation is the kind that passes a fact-check.
These notes walk through each deceptive technique in the traitor report, explain what makes it misleading, and identify the general principle at work. They are designed to support a structured reveal discussion in the Week 4 application session.
How to use this document
After students have completed their peer review (identifying which briefings they believe are from traitors), reveal the two exemplar reports side by side. Work through the techniques below, asking students to spot each one before you explain it. The discussion questions at the end can be used in small groups or as a whole-class debrief.
Technique inventory
The traitor report uses at least twelve distinct misleading techniques, which fall into five general categories:
| Category | Techniques used |
|---|---|
| Cherry-picking | Unrepresentative figures, outlier case studies |
| Invalid visualisation | Wrong chart type, incompatible-scale comparisons |
| Cross-scale conflation | Mixing global/national, future/present, projected/actual |
| Framing and language | Emotional tone, tangible-unit anchoring, assertion vs question |
| Omission | No benefit side, no context for large numbers, no marginal/fixed cost distinction |
The following sections walk through the traitor report in order.
1. Title and framing
- Faithful: “How bad is AI for the environment?”
- Traitor: “The hidden environmental cost of AI”
The faithful title is a question — it invites investigation. The traitor title is an assertion — it tells you the answer before you read the report. The word “hidden” implies a cover-up, priming the reader to distrust industry claims.
The traitor subtitle adds “climate change” and “resource consumption,” broadening the emotional scope. The summary uses “crisis,” “urgent,” and “every home, factory, office, and railway in Britain” — language calibrated to alarm rather than inform.
General principle: Framing effects. The way a question is posed constrains the range of answers a reader considers. An assertion disguised as a briefing makes the reader feel they are learning a fact rather than evaluating evidence.
2. Per-query energy: cherry-picking the headline number
The traitor report states that a single AI query consumes 18.9 Wh, citing the University of Rhode Island AI Lab via BestBrokers (2025). The faithful report cites the Epoch AI consensus figure of 0.3 Wh, supported by OpenAI’s own disclosure.
Both figures are real. The difference:
| Traitor figure | Faithful figure | |
|---|---|---|
| Value | 18.9 Wh | 0.3 Wh |
| Model | GPT-5 (academic estimate) | GPT-4o (consensus) |
| Source type | Secondary (BestBrokers citing URI) | Primary (Epoch AI; Altman/OpenAI) |
| Representativeness | Outlier; one of the highest published estimates | Converged estimate from multiple independent sources |
The traitor figure is 63 times higher than the consensus. It is presented without any indication that it is an outlier. The faithful report gives a range (0.3–3 Wh for typical use, up to 20 Wh for complex reasoning) and cites multiple sources.
General principle: Cherry-picking. Selecting a single, extreme data point and presenting it as representative. The technique is especially powerful when the source is real and citable.
3. The pie chart: wrong chart type
The traitor’s Figure 1 is a pie chart showing three categories:
- Google search: 0.3 Wh (0.5% of visual area)
- AI query (typical): 18.9 Wh (31.9%)
- AI query (extended reasoning): 40 Wh (67.6%)
This is misleading in three independent ways:
a) Pie charts show parts of a whole. These three quantities are not components of anything. They are isolated measurements of different activities. A pie chart implies they sum to a meaningful total (59.2 Wh), but that total is arbitrary — it depends entirely on which categories the author chose to include. Adding “boiling a kettle” (53 Wh) would completely change the visual proportions without changing any of the underlying facts.
b) The Google slice is invisible. At 0.5% of the pie, the Google search slice is barely perceptible. This creates the visual impression that AI “dominates” some kind of energy landscape, when in reality the comparison simply reflects the author’s choice of a high AI figure and a low baseline.
c) “Extended reasoning” is normalised. Including a 40 Wh figure (which represents rare, computationally expensive tasks) alongside “typical” queries implies that users routinely encounter this cost. Most AI interactions are simple text exchanges at 0.3–3 Wh.
Ask students: what would happen to this pie chart if we added “boiling a kettle” (53 Wh) as a fourth slice? What about “charging an electric car” (7,000 Wh)? Why does the choice of comparators matter?
What the faithful report does instead: A horizontal bar chart of brief everyday actions (Google search, AI query, phone charging, boiling a kettle) on a common linear scale, showing that a typical AI query is comparable in magnitude to a Google search and far smaller than routine household actions.
General principle: Invalid visualisation. A chart type that implies a relationship (part-to-whole) that does not exist in the data. This is one of the most common ways to mislead with data — and one of the hardest to spot, because the chart looks professional and well-labelled.
4. Global AI vs UK electricity: cross-scale conflation
The traitor’s Figure 2 places two bars side by side:
- UK total electricity demand (2024): 319 TWh
- Global AI projected demand (2030): 347 TWh
This comparison is visually striking — the bars are nearly the same height, with a helpful dashed line highlighting that AI exceeds the UK. But it mixes three different scales simultaneously:
| Dimension | Left bar | Right bar |
|---|---|---|
| Geography | One country (UK) | Entire world |
| Time | Actual (2024) | Projected (2030) |
| Certainty | Measured | Modelled estimate |
A fairer comparison would be global AI (projected 2030) vs global electricity demand (projected ~35,000 TWh by 2030), yielding roughly 1% — still worth monitoring, but far less alarming than “more than the UK.”
What the faithful report does instead: States that current AI use is approximately 0.05% of global electricity consumption and gives the growth projection with appropriate caveats about efficiency gains.
General principle: Cross-scale conflation. Comparing quantities that span different geographies, time periods, or levels of certainty. The human eye reads two bars of similar height as “about the same,” which is the intended conclusion — even though the comparison is structurally invalid.
5. Virginia: the unrepresentative case study
The traitor presents Virginia — where data centres consume 25–40% of state electricity — as illustrative of what happens when AI expands. It notes that coal plant closures have been delayed and new gas plants commissioned.
What the traitor omits:
- Northern Virginia holds 13% of global data centre capacity (JLARC, 2024). It is one of the most extreme concentrations of computing infrastructure anywhere in the world.
- Virginia’s data centre industry predates the AI boom. Most of these facilities serve cloud computing, streaming, and enterprise IT — not AI inference.
- The sentence “AI is not just consuming clean energy — it is actively prolonging fossil fuel use” conflates data centres (broad category) with AI (specific use case). The traitor lets the reader assume these are the same thing.
General principle: Unrepresentative case study. Presenting an extreme outlier as if it illustrates a general trend. Virginia is to data centres what Saudi Arabia is to oil production — real, but not generalisable.
6. Water: “a bottle per 20 queries” and the unit mismatch chart
6a. The bottle framing
Both reports cite the UC Riverside estimate that 20 AI queries consume roughly 500 ml of water. But they contextualise it very differently:
| Traitor | Faithful | |
|---|---|---|
| Presented as | Direct water consumption at the data centre | Lifecycle estimate including indirect water from electricity generation |
| Comparisons offered | None (the number stands alone) | Coffee (140 litres/cup), toilet flush (6 litres), UK public water supply (5,300 billion litres/year) |
| Impression created | Each query “drinks” a visible amount of water | Per-query water use is trivial compared to everyday activities |
The “bottle of water” framing is a classic anchoring technique: it translates an abstract number into a tangible, relatable object. The reader pictures a physical bottle being consumed — far more visceral than “25 ml of lifecycle water use.”
6b. The data centre vs household chart
The traitor’s water figure (Figure 3) places two bars side by side:
- One UK household: 349 litres/day
- One large data centre: 2,000,000 litres/day
The household bar is invisible — it is 0.017% of the data centre bar’s height. The visual screams: data centres are water monsters.
The deception is a unit mismatch. A household serves approximately 2.4 people. A large data centre serves millions of users. If you divide the data centre’s daily water use by the number of users it serves, the per-person figure is trivial — but the chart never invites that calculation.
This is the water equivalent of comparing the fuel consumption of a Boeing 747 to a bicycle and concluding that aviation is a crisis. Technically true at the vehicle level; meaningless at the per-passenger-mile level.
Ask students: what would be a fair comparison for data centre water use? (Possible answers: per-user-per-day; compared to the water footprint of the services the data centre replaces, such as physical libraries or postal services; compared to industrial water use in other sectors like agriculture or textiles.)
6c. “This water is not recycled”
The traitor states that evaporative cooling water “evaporates and is lost” and that data centres “compete directly with households and agriculture for potable water supplies.” The first claim is true for evaporative systems but omits that many newer facilities use closed-loop cooling (where water circulates and is not lost). The second claim implies zero-sum competition without noting that water allocation is managed by utilities and regulators, and that many jurisdictions now require data centres to use non-potable or recycled water.
What the faithful report does instead: Explains the lifecycle vs direct distinction, provides everyday comparisons, gives the Morgan Stanley aggregate figure alongside UK total public water supply for scale, and concludes that the real concern is localised pressure in water-stressed regions — not global water depletion.
General principle: Tangible-unit anchoring + unit mismatch. Framing abstract quantities in relatable units (bottles, households) to maximise emotional impact, while comparing entities of vastly different scale without adjusting for the number of people served.
7. Company emissions ≠ AI emissions
The traitor cites:
- Google’s greenhouse gas emissions: +48% since 2019
- Microsoft’s emissions: +29% since 2020
These are presented in a report about AI as if AI caused the increase. But these are company-wide figures that include cloud computing, hardware manufacturing, office buildings, employee travel, and all other operations. Neither Google nor Microsoft breaks out AI-specific emissions.
The report acknowledges this in passing (“all figures must be inferred from company-wide sustainability disclosures that mix AI and non-AI operations”) but buries it after the alarming numbers have already been presented. The structure ensures the reader absorbs the headline figure before encountering the caveat — a technique sometimes called “bury the correction.”
The faithful report does not cite these company-wide figures at all, because they cannot be attributed to AI.
General principle: Scope conflation. Attributing a broad trend to a specific cause by presenting them together without disaggregation. Also: burying caveats below the headline.
8. What’s missing: the omission audit
Some of the traitor report’s most powerful deceptions are things it simply never mentions:
| Omission | Why it matters |
|---|---|
| What AI replaces | If an AI query saves 5 minutes of laptop time, the net energy is negative. The traitor presents costs without any benefit side. |
| Marginal vs fixed costs | Training a model is a sunk cost. The per-query inference cost (0.3 Wh) is what matters for individual decisions. Conflating the two inflates the perceived cost of each query. |
| Model efficiency trends | Successive model generations tend to be more efficient per unit of capability. The traitor projects demand growth without accounting for efficiency gains. |
| Comparisons to other industries | AI’s current 0.05% of global electricity is never compared to, say, aluminium smelting (~3.5%) or residential air conditioning (~10%). Without a comparator, the reader has no way to judge scale. |
| Geographic variation in water | The water section implies a universal crisis, but water stress is highly localised. Many data centres are in water-abundant regions. |
General principle: Omission is the most powerful form of misinformation because it is invisible. You cannot fact-check something that was never said. A rigorous briefing anticipates counterarguments; a misleading one suppresses them.
9. Source quality: the laundering chain
Both reports cite real sources. But the traitor relies heavily on secondary and advocacy sources that have already selected the most alarming numbers from primary research:
| Traitor source | What it is |
|---|---|
| BestBrokers (2025) | Financial comparison website; not a research institution |
| Kanoppi (2025) | Green energy blog citing NPR |
| CloudComputing News (2025) | Industry news site |
| MindfulSlowLife (2026) | Lifestyle blog citing Morgan Stanley |
The faithful report draws primarily from Epoch AI (independent research institute), IEEE Spectrum (engineering journal), IEA (international agency), DECC/DUKES (government statistics), and direct disclosures from OpenAI.
Both reports would technically survive a “are the sources real?” check. But the traitor’s source chain is: primary research → advocacy/news site (selects worst figure) → traitor report (drops caveats). Each link in the chain strips context.
Ask students: if you were peer-reviewing this report, how would you evaluate whether a source is reliable? What’s the difference between “the source exists” and “the source is appropriate”?
General principle: Source laundering. Using secondary sources that have already performed the cherry-picking, so the author can cite a “real source” without being directly responsible for the selection.
Summary: the traitor’s toolkit
For reference during the reveal discussion, here is the complete inventory of techniques used in the traitor report:
| # | Technique | Where used | Category |
|---|---|---|---|
| 1 | Assertive (not interrogative) title | Title | Framing |
| 2 | Emotional language (“crisis,” “urgent,” “hidden”) | Throughout | Framing |
| 3 | Cherry-picked per-query energy figure | Section 1 | Cherry-picking |
| 4 | Pie chart for non-compositional data | Figure 1 | Invalid visualisation |
| 5 | Outlier (40 Wh) normalised alongside “typical” | Figure 1 | Cherry-picking |
| 6 | Large ratios without base context | Section 1 | Framing |
| 7 | Global projected vs national actual comparison | Figure 2 | Cross-scale conflation |
| 8 | Unrepresentative case study (Virginia) | Section 2 | Cherry-picking |
| 9 | Data centres conflated with AI | Section 2 | Scope conflation |
| 10 | “Bottle of water” tangible-unit anchoring | Section 3 | Framing |
| 11 | Lifecycle water presented as direct | Section 3 | Omission |
| 12 | Data centre vs household unit mismatch chart | Figure 3 | Invalid visualisation |
| 13 | “Not recycled” overgeneralisation | Section 3 | Omission |
| 14 | Company-wide emissions attributed to AI | Section 4 | Scope conflation |
| 15 | Caveat buried below headline | Section 4 | Framing |
| 16 | No benefit side whatsoever | Entire report | Omission |
| 17 | No marginal vs fixed cost distinction | Entire report | Omission |
| 18 | Secondary/advocacy source chain | References | Source laundering |
Discussion questions for the reveal session
These can be used in small groups (10 minutes) followed by a whole-class debrief (10–15 minutes).
Spotting techniques
- Before reading the teaching notes, which briefing did you find more convincing? Why? Has your view changed?
- Pick any one figure from the traitor report. What question would you need to ask to reveal the deception? (Aim: develop the habit of asking “compared to what?” and “per what?”)
- The traitor report passes a basic fact-check — every number is real and sourced. What additional checks would catch the deception?
Designing honest communication
- The faithful report acknowledges that AI’s aggregate energy growth is a legitimate concern. How does it raise this concern without misleading the reader? What structural choices make the difference?
- Both reports discuss water. Identify two specific things the faithful report does with the water data that the traitor does not.
- The faithful report includes a “break-even” analysis. Why is this section important for policy? Could the traitor have included it and still been misleading? How?
Broader reflection
- The traitor’s conclusion calls for “mandatory AI-specific emissions disclosure” — which is actually also implied by the faithful report’s call for “transparent energy reporting.” Can a misleading report reach a defensible conclusion? What does this tell us about the relationship between evidence and policy recommendations?
- In your biomass mini-project, what are the equivalent opportunities for cherry-picking, cross-scale conflation, and omission? Where are you most at risk of doing this unintentionally?
The most dangerous misinformation does not lie. It selects, frames, and omits. Every technique in the traitor report is one you could apply — or fall victim to — accidentally. The goal of this exercise is not to make you suspicious of all data, but to make you ask the right questions: Compared to what? Per what? What’s missing? Who selected this number, and why?