Lens 1: Stochastic Parrots

Scale and Environment

Laura Wagner and Moritz Mähr

Learning question

Which environmental and data choices are hidden by a model’s apparent performance?

A benchmark score shows what a model can do. It does not show what the model costs, or where its data come from.

Slides: CC BY 4.0. Lesson page and worksheet: maehr.github.io/stochastic-parrots-and-beyond

Ask: “When you chose your last AI tool, what did you know about it?” Collect two or three answers. Most people know the interface and the price, not the data or the energy.

1. Scale is not a sufficient criterion

• Bender et al. ask: “How big is too big?” (2021, 610).
• Rising costs “doubly [punish] marginalized communities”: they benefit least and suffer the environmental consequences first (610).
• A language model stitches together linguistic forms “without any reference to meaning: a stochastic parrot” (617).

Read the definition slowly. Point out that the paper names costs and data before it names the parrot.

2. Models have a material footprint

• Energy: data centers used about 415 TWh in 2024, about 1.5% of global electricity. The IEA projects about 945 TWh by 2030, with AI as the main driver (IEA 2025).
• Water: training GPT-3 could directly evaporate about 700,000 liters of fresh water, according to an estimate by Li et al. (2025).
• Hardware: manufacturing servers and GPUs causes emissions before the first computation (Luccioni et al. 2023).

Estimates and projections. Use them as orders of magnitude.

Distinguish measurement, estimate, and projection. The IEA figures concern all data centers, not one model.

3. Use costs energy again and again

• Each request needs energy. Use is not a one-off cost.
• Multi-purpose generative models are “orders of magnitude more expensive” than task-specific systems for several tasks (Luccioni, Jernite, and Strubell 2024).
• The same computation causes different emissions on different grids.

Ask: “For your task, would a smaller or task-specific tool do the job?”

4. Data choices are value choices

• A larger web dataset is not a more diverse dataset (Bender et al. 2021, 613).
• One corpus discarded every page with a word from a list of about 400 “bad words”. This also suppresses spaces “built by and for LGBTQ people” (614).
• Undocumented data create “documentation debt” (615).

Transition to the case: “What can we know when a team decides to measure and publish?”

Case: BLOOM

Documented case

Data: Luccioni, Viguier, and Ligozat (2023). Original chart, CC BY 4.0.

BLOOM: 176 billion parameters, an open-science project with more than a thousand researchers, trained in France. Point out the three parts of the footprint. Idle infrastructure and hardware make up half.

What the BLOOM figures reveal

• The grid: 57 g CO₂eq per kWh in France. The authors explain the low emissions largely with this grid.
• Experiments: intermediate models emitted 35.8 t from dynamic power use, more than the 24.69 t of the final training on the same boundary.
• Use: 230,768 requests in about 18 days used 914 kWh. Roughly 75% of the energy kept the model in memory, ready for requests.
• Data: the 1.6 TB ROOTS corpus in 46 natural and 13 programming languages came with a data governance effort.

Sources: Luccioni, Viguier, and Ligozat (2023); Laurençon et al. (2022); Jernite et al. (2022).

Ask: “For the system you use, which of these four lines could you fill in?” Usually none. The empty cells are a finding.

Activity · 15 minutes

Trace one model or use case from its data sources to its use. Groups of three or four, on paper or on a laptop.

1. Choose a research-use card from your group. Use BLOOM only if no card fits. (3 min)
2. For each stage, note what is known, unknown, and avoidable, with a source. (8 min)
3. Mark the largest unknown and one avoidable cost. (2 min)
4. Prepare one sentence for the plenary. (2 min)

Worksheet: Lens 1, material-and-provenance map

Hand out the Lens 1 worksheet. Announce each step. At 13 minutes, call for the sentence: “The choice that the performance of this system hides most is …”

Expected output

A material-and-provenance map for each group:

• known, unknown, and avoidable costs for each stage,
• a source for each entry,
• the largest unknown and one avoidable cost.

Collect the sentences in the plenary. Keep the maps on the wall for the afternoon.

Sources

• Bender et al. (2021): On the Dangers of Stochastic Parrots.
• Luccioni, Viguier, and Ligozat (2023): Estimating the Carbon Footprint of BLOOM.
• International Energy Agency (2025): Energy and AI.
• Crawford (2021): Atlas of AI, chapter 1, “Earth”.

Takeaway

Performance is what a model shows. Provenance and footprint are what it hides. A missing number is an unknown, not a zero.

Transition to Lens 2: “We followed data as a cost. Next we follow data as a picture of the world.”