Red team

Attacking a claim is nearly free. Proving it is costly.

With frontier AI, campaign groups, journalists, investors, and lawyers can read a sustainability report the way only a funded research team could. This page looks at what that shift does to the risk in a published claim, and at where AI-assisted criticism gets things wrong.

What a critic's reading found
  • 39%

    of graded claims an outsider can verify from public sources

  • 0 of 12

    reports give units produced, unsold stock, or product life

  • 14–72%

    verifiable among the nine companies rated CDP A or A-

From our review of twelve apparel and luxury reports. See the evidence

Checking got cheap. Proving didn't.
Checking a report
3 weeks3 hours

Two or three people before AI. One person with AI, in our study.

Proving a claim
Unchanged

It still takes data, a published method, and external assurance.

The old asymmetry

For years, the advantage sat with the company. A sustainability report runs to hundreds of pages, and the evidence behind its claims is spread across annual reports, basis-of-reporting documents, supplier lists, and earlier years.

Checking one report against all of that took a team weeks. Few critics had the time or the money, so most claims were never tested by anyone outside the company and its assurance provider.

Assurance helps, but most of it is limited assurance, and it covers only part of each report.

What AI changed

Frontier AI has collapsed the cost of reading. In our own study, work that would have taken two or three people about three weeks took about three hours of human time.

The cost of substantiating a claim hasn't fallen the same way. Making a figure checkable still takes data, a published method, and external assurance.

The readers with this new capacity include campaign groups, journalists, plaintiff firms, short sellers, competitors, and regulators.

The burden has moved. A claim that an outsider can't check from public sources now reads as a gap, whether or not the company holds the evidence.

Case study

We red-teamed twelve apparel and luxury sustainability reports

In November 2026, we ran the review on the 2025 sustainability reports of twelve listed apparel and luxury companies. We logged 486 claims and graded the 365 that state a target or a result. We used public documents only, and no company commissioned the review.

For each graded claim, we asked whether an outsider could verify it from public sources. Is it fully disclosed? Can it be checked without taking the company's word for a certificate, an estimate, or its own definition? Does it sit inside external assurance?

See the twelve companies and how the grades were made

Finding 1

Most claims can't be verified by an outsider

Three in four graded claims are fully disclosed. Only two in five are also checkable from public sources and inside external assurance. No claim in the sample carries reasonable assurance; the strongest is limited assurance.

Three in four claims are documented; two in five can be verified
  1. Graded claims (targets and results)100% (365)
  2. ↓ 86 are not fully disclosed

    Fully disclosed76% (279)
  3. ↓ 97 rest on certificates (41), own definitions (27), estimates (25), or blurred totals (4)

    Also checkable by an outsider50% (182)
  4. ↓ 41 sit outside any external assurance

    Also inside external assurance39% (141)

Each claim is counted once, by the main reason it drops out.

Finding 2

The reports leave out what drives a product's footprint

Every report gives emissions and materials figures. None discloses units produced, unsold stock, or how long products last, so no report lets a reader compute a garment's footprint per wear.

Every report measures emissions; none counts units, unsold stock, or wearsReports with a figure for each topic
  • Emissions12 of 12
  • Materials12 of 12
  • Repair or resale4 of 12
  • Units produced or sold0 of 12
  • Quantity of unsold stock0 of 12
  • Product life or number of wears0 of 12

Repair or resale: at least four companies. A keyword search can miss a figure in an image or a financial note.

Finding 3

So ESG ratings are inconclusive about impact

Nine of the twelve companies hold a CDP climate score of A or A-, and their share of claims that pass all three tests ranges from 14% to 72%. CDP bases its score on the company's own questionnaire response, and a study of six major ESG raters found that their ratings agree at 0.54 on average, where 1 is full agreement.

A top CDP score sits alongside verification rates from 14% to 72%Share of each company's graded claims that pass all three tests
CDP climate score A or A-
14%72%
No public CDP score located

Each dot is one company. Part of the spread reflects how much of each report sits inside external assurance. A CDP score assesses disclosure and management, not the claims tested here.

Where the risk now sits

Claims that rest on the company's word

Figures that hold only through purchased certificates, estimates, or the company's own definitions. In our study, these explain most of the gap between claims that are documented and claims an outsider can verify.

Silence

What a careful reader expects and doesn't find. No report in our sample discloses units produced, unsold stock, or product life, and AI makes absences as easy to list as claims.

Contradictions across documents

Where the sustainability report, the annual report, marketing pages, and last year's figures disagree. Cross-checking them used to take days and is now routine.

Moving targets

Reset baselines, withdrawn targets, and paused goals. Each can be legitimate when explained, and each is now easy to track across years.

Incentives tied to easy metrics

Executive pay linked to certificate-based electricity, offset-based neutrality, or a rating score. Critics can now read remuneration reports alongside sustainability reports.

Consumer law

From September 27, 2026, EU consumer law bans generic environmental claims without proof of excellent performance, and claims that a product is neutral because of offsets. The proposed Green Claims Directive stalled, but these rules apply.

The asymmetry cuts both ways

AI-assisted criticism makes mistakes too. It misses evidence in images, tables, and linked documents, and it can't see what a company hasn't published. In our own study, every correction so far has raised a grade: the first pass overstated the gaps.

An allegation is now cheap to make, and a rebuttal still takes time and evidence. That puts a duty on both sides. Companies need their evidence public and easy to find. Critics need to publish their method, state their limits, and give the companies they name a chance to respond. The notes on our case study, below, say which of these it hasn't done yet.

Our part

We read sustainability reports the way a well-equipped critic now can, and we say where that reading falls short. For a company, it shows which claims an outsider can't verify, and what evidence would close each gap, before the report goes out.

Commissioned reviews stay confidential. Public studies, like the one on this page, show their method and limits.

The twelve companies

  • Company A: European sportswear brand
  • Company B: European luxury group spanning several product sectors
  • Company C: US premium lifestyle brand
  • Company D: European luxury apparel brand
  • Company E: European mass-market fashion retailer
  • Company F: European luxury menswear group
  • Company G: European luxury group built around fashion and leather goods houses
  • Company H: European sportswear group
  • Company I: US casualwear company
  • Company J: European luxury fashion house
  • Company K: European premium fashion brand
  • Company L: European luxury fashion group

If you commission a review

A written list of weaknesses can become evidence in a dispute that the company knew about them. A company can run the review through its legal function or outside counsel, under privilege.

Independence

Reviewing a strategy we designed isn't red teaming. Where we've advised on the underlying work, we say so, as the case study on this page does.