Truace
Subscribe
52/100

Moderate evidence

Cite
More
Export CSVExport JSONEmailLinkedInX

AI chatbots misjudge carbon labels for recycled building materials, study finds

Study title: Towards trustworthy LLM-assisted EPD comparability assessment for circular construction materials

Evidence type
Benchmark or lab evaluation
Sources
1 sourcePeer-reviewed
Published
2 min read
Truvace record
TRV-2026-2700cite or check it

Study title: Towards trustworthy LLM-assisted EPD comparability assessment for circular construction materials

A peer-reviewed study tested how three large language models, GPT-5.3, Qwen 3.6-Plus and Gemini 3, read and compare Environmental Product Declarations (EPDs) for construction materials. The case materials were virgin aggregates and recycled concrete aggregates, and the focus was A1–A3 carbon emissions.

Read the study abstract

Environmental Product Declarations (EPDs) are increasingly used to support construction material selection, but their comparability depends on how system boundaries and allocation rules are interpreted. Although large language models (LLMs) are increasingly explored for sustainability assessment, limited attention has been given to how methodological ambiguity affects LLM reasoning reliability and comparability judgement. This study investigates how LLMs interpret A1–A3 carbon emissions in EPDs for virgin and re…

The problem

A study testing GPT-5.3, Qwen 3.6-Plus and Gemini 3 on carbon declarations for virgin and recycled aggregates found zero-shot models often defaulted to calling products non-comparable, and rule-based prompts that raised accuracy also made answers less stable.

In brief

Zero-shot models often gave stable but conservative answers, frequently defaulting to non-comparability when the methodological conditions were hard to synthesise. Rule-grounded prompting improved judgement accuracy and reasoning quality for GPT-5.3 and Qwen 3.6-Plus, but made responses less stable and more variable.

The authors conclude that current LLMs should serve as supportive tools rather than standalone decision-makers. They also call for clearer, more harmonised methodological representation inside EPDs so AI-assisted interpretation becomes more reliable.

Main points

  1. Three models were tested: GPT-5.3, Qwen 3.6-Plus and Gemini 3.
  2. The tasks were individual EPD interpretation and pairwise comparability assessment, under zero-shot and rule-grounded prompting.
  3. Virgin aggregates and recycled concrete aggregates were the case materials.
  4. A1–A3 emissions for recycled materials may not share the process boundaries of virgin ones, because recycling burdens can be allocated to different EPD modules.
  5. Zero-shot models often defaulted to non-comparability on hard cases.
  6. Rule-grounded prompting improved accuracy and reasoning quality for GPT-5.3 and Qwen 3.6-Plus but reduced stability and increased variability.

The full story

EPDs are increasingly used to support construction material selection, but how comparable they are depends on how system boundaries and allocation rules are interpreted. The study asked how that methodological ambiguity affects the reliability of LLM reasoning.

Models handled two tasks: interpreting single EPDs and judging whether pairs of declarations could be compared. Each ran under zero-shot prompts and prompts grounded in explicit rules.

The results showed an accuracy–stability trade-off for GPT-5.3 and Qwen 3.6-Plus: rules improved judgements but made outputs vary more. The authors say LLM-assisted EPD interpretation needs evaluation beyond final-answer accuracy.

Sources

1 on file
  1. Environmental Impact Assessment Reviewdoi:10.1016/j.eiar.2026.108606
    Peer-reviewed2026-07-09

Talk it over

Four questions to start onWho carries this cost, and what would have to change to prevent it?

Start with a question

AI chatbots misjudge carbon labels for recycled building materials, study findshttps://truvace.com/item/towards-trustworthy-llm-assisted-epd-comparability-assessment-for-circular-const

  1. Who carries this cost, and what would have to change to prevent it?

    The problem A study testing GPT-5.3, Qwen 3.6-Plus and Gemini 3 on carbon declarations for virgin and recycled aggregates found zero-shot models often defaulted to calling products non-comparable, and rule-based prompts that raised accuracy also made answers less stable.

    Answer in the thread

  2. How do we know? What can this kind of evidence show, and what can’t it?

    The evidence This rests on a benchmark or lab test, from Environmental Impact Assessment Review.

    Answer in the thread

  3. What does it leave open? What else would you want to know before acting on it?

    The record’s own caveat The evaluation covered only two case materials, virgin aggregates and recycled concrete aggregates, and A1–A3 emissions, so results may not extend to other materials or life-cycle stages.

    Answer in the thread

  4. Have you seen AI used this way where you study, work or live? Did it go the same way?

    Answer in the thread

The questions are the same on every record. The details under them are taken from this record as written.

What readers say

More from Climate