Can AI-Generated LCA Data Be Trusted? Progress, Challenges and Validation

‍

Two months ago we asked under which conditions machine-generated LCA data can be trusted, and said it has to be traceable, transparent and verifiable. Here is where that stands.

‍

How Does AI-Generated Life Cycle Assessment (LCA) Work?

‍

A user types one line, "1 kg primary nickel, US, 2026, cradle-to-gate", and the engine returns a result, in this run 13.0 kg CO2e per kg. That number alone says little, so the tool is built to show what sits behind it: the footprint breaks down by production step, any process can be opened to see what it consumes, and every figure carries its source or is labelled as an estimate.

Making the model visible turned out to be the most effective way of finding errors. The engine was at first adding up emissions instead of solving the supply chain, which understated the result by about a third. Fixing it brought the study's electricity consumption per kg of nickel to within 6% of the published industry figure, from 87% above it. No validation rule had caught it.

‍

Key Challenges in AI-Generated LCA Data Validation

Ensuring Accuracy Through Human Review

• Individual values still need human review. The structure is plausible, the quantities less so. In one study, the smelter's electricity use was six times too low, even though every step looked reasonable on its own.

Enforcing Consistent LCA Methodological Rules

• Methodological rules have to be maintained, not just written. The engine carries 26 rules that guide the model while it writes and 43 checks that test the result afterward. A rule that exists only in the prompt gets ignored; a check that exists only in code fires too late.

Addressing Reproducibility in AI-Generated LCA Results

• The same question does not give the same answer. Four runs for the same nickel product gave totals between 11.1 and 13.0 kg CO2e. The calculation is deterministic, so the variation comes from the AI's sampling.

‍

What's Next for AI-Driven LCA and Environmental Impact Assessment?

‍

The studies are starting to add up to something. Rather than letting the model reinvent diesel refining or ammonia synthesis for every run, we promote reviewed processes into a library the engine draws from. That is the beginning of our own LCI database, built the opposite way round to a conventional one: not compiled up front but accumulated from checked studies.

Whether the numbers hold up without human review is still open. And while we report on climate change here, eight further impact categories are already computed from the same inventory; bringing their validation to the same level is next.

And there's more to come! We'll soon share a closer look at how we're building our own LCI database and developing an LCA workbench to bring it all together. Stay tuned!