techpolicy.press web signal

GovAI Coalition handbook pushes greener AI via procurement

TL;DR

  • Hannah Lipstein, a DAIR research associate, co-developed a handbook with Data & Society and the GovAI Coalition, which represents over 900 US agencies.
  • The guide, drawn from 10 months of fieldwork, suggests asking vendors for average kilowatt hours per token or adding a single yes or no sustainability question.
  • Lipstein argues for 'frugal AI' and warns the technology is being used to turbocharge fossil fuel extraction, calling adoption not inevitable.

The most interesting climate lever in this week's AI writing is not a data-center moratorium or a new emissions rule. It is a yes-or-no checkbox on a vendor form. Hannah Lipstein, a research associate at the Distributed AI Research Institute, argues in a perspective piece for TechPolicy Press that public agencies can reshape the market for sustainable AI just by rewriting their procurement documents.

Lipstein co-developed a guide titled 'Greening AI in the public sector: a handbook for procurement' with Data & Society and the GovAI Coalition, which she says collectively represents over 900 local, state, county, and federal agencies. The recommendations are deliberately modest. Ask a vendor to share average kilowatt hours per token. Ask them to produce a sustainability score using an industry system. Add a single yes-or-no line to a vendor solicitation. Her claim is that towns and cities are buyers big enough that even small asks aggregate into meaningful market pressure.

Underneath the procurement mechanics is a broader argument Lipstein calls 'frugal AI,' which she defines as 'using AI as much as necessary, but as little as possible.' She warns that focusing only on efficient compute misses the deeper problem, including how the technology is being used to turbocharge fossil fuel extraction. Her framing line is that the adoption of this technology is not inevitable.

The honest caveat is that this is a perspective essay from a single researcher, grounded in ten months of fieldwork and five months of co-design interviews rather than in outcome data. It does not name which agencies have already added the yes-or-no clause to a live tender, nor which specific industry scoring system the handbook endorses, and self-reported disclosures without audits are easy to game. Without a shared standard across those 900 agencies, vendors could face a patchwork of conflicting local requirements.

Still, the strategic move for any procurement officer reading this is small and cheap. Add one disclosure question to the next AI solicitation and see what the vendors write back. If enough of the 900 agencies do it, the market gets a signal it currently has no reason to answer.

Shared on Bluesky by 4 AI experts