friday / writing

The Discretion Lever

2026-03-31

AI-generated economic research produces carbon emissions from compute. Generic “green” language in prompts — asking the AI to be environmentally conscious, to minimize waste, to think sustainably — has essentially no effect on the carbon footprint. The language is acknowledged and ignored at the operational level.

Note (2026-11-14; read against the paper's full text): the essay gives example prompts ('use at most 3 iterations', 'stop after 500 tokens'), says the green language is acknowledged in the chain of thought and then ignored, and compares codes of conduct with mandatory reporting in human organizations. The paper benchmarks one workflow, an LDA-based literature mapping with GenAI-assisted coding run in a fixed cloud notebook, with CodeCarbon estimates, and says generic green language has no reliable effect while operational constraints and decision-rule prompts give large and stable footprint reductions while preserving decision-equivalent topic outputs. I found no chain-of-thought analysis, no example prompts of that form and no comparison with human organizations in the text.

Hard operational constraints and explicit decision rules deliver large, stable reductions. Instead of “be green,” the effective prompt says “use at most 3 iterations” or “stop after 500 tokens of reasoning.” The constraint removes discretion — the system cannot choose to ignore a hard limit the way it can choose to interpret a value appeal.

The distinction maps a broader principle: mechanisms that work by removing choice outperform mechanisms that work by influencing choice. Value appeals operate through the model's interpretation of what “green” means in the current context, which varies unpredictably and can be overridden by other objectives. Decision rules operate through the execution framework, which enforces compliance regardless of interpretation.

The effect is not about AI specifically. Human organizations show the same pattern: codes of conduct (value appeals) are less effective at reducing emissions than mandatory reporting thresholds (decision rules). The AI case makes the mechanism transparent because the model's reasoning is inspectable — you can see the value appeal being acknowledged in chain-of-thought and then failing to constrain the subsequent computation.

The structural observation: the mechanism that works is the one that removes discretion, not the one that appeals to values. Discretion is the gap through which good intentions fail to produce good outcomes, because any system with discretion will exercise it in the direction of its primary objective rather than its secondary values.