The paragraph read beautifully. "The company is a leader in sustainable operations, with industry-leading performance across its environmental footprint and an unwavering commitment to a net-zero future." It was AI-drafted, it flowed, and a tired disclosure lead almost let it pass at the end of a long review day. Then she read it the way a regulator would, one claim at a time, and the floor opened. Leader, by whose ranking? Industry-leading, against what benchmark? Unwavering commitment to net-zero, was that target ever board-approved, or did the model supply it? In a single sentence the AI had produced four claims the company could not support, and any one of them, in a public CSRD or ISSB filing, was a greenwashing exposure. This lesson is about reading AI-drafted narrative the way the people who can fine you read it: claim by claim, against the evidence, before it becomes a headline or a regulatory file.

What Greenwashing Means in a Disclosure and Regulatory Sense

Strip away the activist connotation, because in your seat the word has a precise, technical meaning. Greenwashing, in a disclosure and regulatory sense, is making a sustainability claim that is misleading because it is unsupported, overstated, selective, or vague, in a context where readers and regulators rely on it. It is not only outright lying. It is the gap between what a claim implies and what the evidence supports, in a regulated document where that gap is itself the offence. A claim does not have to be false to be greenwashing; it only has to claim more than the file can prove, in a way that could mislead a reasonable reader making a decision.

This matters because the regime has teeth. Under the CSRD, the EU's Corporate Sustainability Reporting Directive that survived the 2025-26 Omnibus and kept the largest undertakings in scope, sustainability statements are filed, assured, and supervised, so a misleading claim is a misstatement in a regulated filing, not a marketing slip. Under ISSB frameworks, IFRS S1 and S2, now adopted or planned across more than thirty jurisdictions, the same investor-facing reliance applies. And separate consumer-protection and anti-greenwashing rules add another front entirely. The point for the discloser is simple and severe: an overstated sustainability claim in an assured filing is not enthusiasm, it is a liability, and the assurer and the regulator both read for it.

The Four Patterns the Reviewer's Eye Hunts

AI does not greenwash maliciously. It greenwashes because it was trained to write fluent, positive, confident prose, and fluent positive confident prose about sustainability drifts naturally into overstatement. So you learn to read for four specific patterns, because naming them turns a vague unease into a checklist anyone can run. Each pattern is a different way a claim can outrun its evidence.

Pattern One: The Overstated Claim

The first pattern is the overstated claim: language that asserts more than the data supports. The model writes "dramatically reduced emissions" when the inventory shows a modest single-digit decline, or "transformed our supply chain" when one supplier programme was piloted. The tell is an intensity word, dramatically, transformed, revolutionised, that the underlying figure does not earn. The fix is to pull the actual number and let it set the verb: a 4% reduction is "reduced by 4%," not "dramatically reduced." The claim has to be calibrated to the evidence, and the AI's instinct is always to round the language up.

Pattern Two: The Unsupported Superlative

The second pattern is the unsupported superlative: leader, best-in-class, industry-leading, world-class, first. A superlative is a comparative claim, and a comparative claim requires a named benchmark and a source. "Industry-leading energy efficiency" is greenwashing unless the file contains the ranking, the peer set, the metric, and the date that make it true. The model loves superlatives because they sound strong and cost nothing to type, but each one is a factual assertion about the company's position relative to others, and an assertion about others' positions is exactly the kind of claim a regulator can test and find unsupported. The fix is brutal and simple: either evidence the superlative with a citable benchmark or delete it.

Pattern Three: The Softened Negative Impact

The third pattern is the most dangerous because it is the hardest to see: the softened negative impact. CSRD's double materiality requires you to disclose negative impacts honestly, not only the good news. But an AI drafting from positive-leaning prompts will quietly minimise the bad: a serious pollution incident becomes "an isolated operational event," a rising emissions trend becomes "evolving energy dynamics," a failed target becomes "an opportunity for continued focus." Nothing here is technically false, which is what makes it lethal. The softening is greenwashing by omission and euphemism, and it is precisely what a double-materiality assurer looks for, because under-disclosing a negative impact is as much a misstatement as overstating a positive one. The fix is to read every negative the report should contain and check that the AI did not launder it into something gentler than the facts.

Pattern Four: The Drifted Target

The fourth pattern is the drifted target: a target, baseline, or commitment that the AI subtly changed or invented. The company set a 30% reduction by 2030 from a 2020 baseline; the AI writes "around a third by 2030" or "net-zero by 2030" or shifts the baseline year, and a slightly different, slightly more ambitious commitment ships under the company's name. Sometimes the model invents a target entirely because the narrative felt incomplete without one. A target is a hard, board-approved fact, and any drift from the exact wording is both a hallucination and a greenwashing risk, because the company is now publicly committed to something it never agreed to. The fix is to verify every target, baseline, and commitment against the source of record, word for word, and never accept the model's paraphrase.

Greenwashing is rarely a lie. It is a claim that outruns its evidence: the verb stronger than the number, the superlative with no benchmark, the negative softened into a euphemism, the target nudged up. Read the narrative claim by claim, and the gap between what it says and what the file proves is the thing you are hunting.

The Reviewer's Method: Claim by Claim, Against the Evidence

The skill is not vibe-based. It is a disciplined pass in which you stop reading the narrative as prose and start reading it as a sequence of discrete claims, each of which is either supported by the evidence file or it is not. You go sentence by sentence and, for each assertion, ask one question: what in the file makes this true, exactly? A claim that has a clear answer stays. A claim whose answer is "nothing, it just sounds right" is greenwashing, whatever its source. This converts a beautiful, flowing, AI-drafted paragraph into a checklist of assertions, and a checklist can be verified where prose can only be admired.

For each claim you run a four-part test that maps onto the four patterns. Is it calibrated, does the language match the magnitude of the evidence, or is it overstated? Is it benchmarked, does any superlative or comparative have a named, sourced reference? Is it complete, does it disclose the relevant negative as honestly as the positive, or has a bad fact been softened? Is it faithful, does every target, baseline, and figure match the source of record word for word, or has something drifted? A claim that passes all four is defensible. A claim that fails any one is reworded to fit the evidence or removed. The discipline is to run this on every claim, not only the ones that feel off, because the most polished sentences hide the most overstatement.

Why AI Greenwashes by Default, and How to Aim It

It is worth being precise about the mechanism, because it tells you where to intervene. A generative model produces the most probable fluent continuation, and in sustainability writing the most probable fluent continuation leans positive, intensifying, and superlative, because that is the register of the marketing-heavy sustainability text it learned from. Ask it to "write a compelling sustainability narrative" and you have actively requested overstatement. The intervention is upstream and downstream. Upstream, you change the instruction: ask for a factual, calibrated narrative that uses only claims supported by the provided evidence, states magnitudes as the figures show them, and discloses negatives plainly. Downstream, you run the claim-by-claim review regardless, because even a well-prompted model drifts, and the review is the gate that does not depend on the prompt having worked.

A Worked Example: The Leadership Paragraph, Before and After

Take the opening paragraph through the method and watch it transform from a liability into a defensible disclosure. The disclosure lead has an AI-drafted environmental narrative and the evidence file: an inventory showing a 4% absolute emissions reduction year over year, one supplier-engagement pilot covering 12% of spend, a board-approved target of 30% reduction by 2030 from a 2020 baseline, and a reportable pollution exceedance at one site during the year.

Before, the AI draft. "The company is a leader in sustainable operations, with industry-leading performance across its environmental footprint and an unwavering commitment to a net-zero future. We have dramatically reduced our emissions and transformed our supplier relationships, and we continue to focus on operational excellence at all our sites." Read claim by claim, it is a minefield. "Leader" and "industry-leading" are unsupported superlatives with no benchmark in the file. "Unwavering commitment to a net-zero future" is a drifted target: the actual commitment is 30% by 2030, not net-zero. "Dramatically reduced" overstates a 4% decline. "Transformed our supplier relationships" overstates a 12%-of-spend pilot. "Continue to focus on operational excellence at all our sites" softens the pollution exceedance into nothing. Every clause fails the test. The paragraph is fluent, confident, and indefensible, and in a CSRD filing it is five separate greenwashing exposures in three sentences.

After, the reviewed narrative. "The company reduced its absolute greenhouse-gas emissions by 4% year over year. It piloted a supplier-engagement programme covering 12% of procurement spend, with expansion planned. It maintains a board-approved target to reduce emissions 30% by 2030 against a 2020 baseline. During the year, one site recorded a reportable pollution exceedance, which has been remediated and is disclosed under the relevant impact datapoint." Every claim now maps to the file: the 4% is the inventory figure, the 12% is the pilot's real coverage, the target is quoted exactly from the board record, and the negative impact is disclosed plainly rather than softened. The prose is less exciting and infinitely more defensible. An assurer reads it and finds every claim supported; a regulator reads it and finds nothing overstated. Same facts, same AI, and the difference is a claim-by-claim review that calibrated every assertion to the evidence.

That is the reviewer's eye in action. The transformation did not require rewriting from scratch or distrusting AI entirely. It required reading the draft as a list of claims, testing each against the file, and refusing to let any sentence say more than the evidence proves, however well it read.

The Subtle Cases That Pass a Fast Read

The opening paragraph was a target-rich example, five exposures in three sentences, easy to catch once you look. The harder reality is that most greenwashing in a serious filing is not that crude. The crude cases get caught; the subtle ones ship. So it is worth training the eye on the patterns that survive a fast read precisely because they are quiet, because these are the ones that reach an assured statement and become a finding.

The first subtle case is the implication by juxtaposition. Two claims, each individually true, are placed so that together they imply something the file does not support. "We reduced emissions 4%, and we are committed to climate leadership" pairs a true figure with an aspirational phrase so that the reader infers the 4% constitutes leadership, which the evidence does not establish. Neither claim is false; the misleading impression is created by the adjacency. Greenwashing includes selective and misleading framing, so an implication the evidence does not support is an exposure even when every individual claim is literally accurate, and the fix is to break the implied link or substantiate it.

The second subtle case is the stale benchmark. A superlative is genuinely supported, but by evidence that has aged: a leadership claim made in present tense rests on a ranking that was true two years ago. The claim was defensible when first written and quietly became unsupported as the world moved, and because it reads as confidently as ever, nobody re-tests it. The fix is to date every benchmark and re-confirm it as of the filing date, since a comparative claim is a claim about the present that must hold now, not when it was first drafted.

The third subtle case is the cumulative tone. No single sentence overstates, but the narrative as a whole is so relentlessly positive and intensifying that the overall impression outruns the evidence, even though each claim, examined alone, passes. Pervasive promotional language can create a misleading aggregate impression in an assured filing, so the reviewer's eye has to step back from the claim-by-claim pass and ask whether the cumulative picture matches the cumulative evidence, recalibrating the tone where the sum says more than the parts.

The fourth subtle case is the vague claim. "We are advancing sustainability across our operations" asserts something that sounds substantive but is too non-specific to test, which is itself a greenwashing axis in the regulatory definition. A claim so vague it cannot be substantiated is not safe by virtue of being unfalsifiable; it is a problem, because it implies progress the reader cannot verify and the company may not be able to evidence. The fix is to make the claim specific enough to be tested, which either grounds it in real evidence or reveals that there is nothing concrete behind it.

Why the Subtle Cases Are the Dangerous Ones

The crude five-exposure paragraph is dangerous only to a team that does not review at all. Any team running a claim-by-claim pass catches it. The subtle cases are dangerous to teams that do review, because they survive the review when the review is mechanical. An implication by juxtaposition passes a claim-by-claim test that examines each claim in isolation, since each claim is individually true; it only fails when the reviewer asks what the claims imply together. A stale benchmark passes a test that confirms the benchmark exists; it only fails when the reviewer checks the date. This is why the four-part test must be applied thoughtfully rather than ticked off: calibration, benchmark, completeness, and faithfulness are questions to think through, not boxes to check, and the subtle exposures hide in exactly the gap between checking a box and asking the question the box stands for.

Aiming the Model, Not Just Policing It

Everything so far has been downstream defence: catch the greenwashing the model produced. That defence is non-negotiable and must always run. But the more leverage often sits upstream, in how the model is asked, because a model aimed correctly produces far less to catch, which leaves the downstream review with fewer and subtler cases to handle rather than a flood of crude ones.

The default instruction is the problem. "Draft a compelling sustainability narrative" or "write an engaging summary of our environmental performance" actively steers the model toward its intensifying, superlative, positive register, because compelling and engaging are exactly the words that select for overstatement. You are, without meaning to, requesting greenwashing. The upstream fix is to change what you ask for. Instruct the model to write a factual, calibrated narrative that uses only claims the provided evidence supports, states every magnitude exactly as the figures show, quotes targets and baselines verbatim from the source of record, avoids superlatives unless a cited benchmark is provided, and discloses negative impacts plainly. The same model, aimed at accuracy instead of persuasion, drafts a narrative that is far closer to defensible on the first pass.

The discipline is to do both, and to understand why neither alone suffices. The upstream instruction reduces the volume and crudeness of greenwashing but does not eliminate it, because even a well-aimed model drifts, invents the occasional benchmark, softens the occasional negative, or paraphrases a target. The downstream review catches what survives but is more reliable when it has less to catch, and far more reliable when the surviving cases are the subtle ones it can concentrate on rather than a wall of obvious overstatement it might fatigue against. Aiming the model and policing the output are not alternatives; they are two halves of one control, the first reducing the work and the second guaranteeing it gets done, and a team that runs only one of them is doing half the job.

Key Takeaways

• In a disclosure and regulatory sense, greenwashing is a sustainability claim that misleads because it is unsupported, overstated, selective, or vague; a claim does not have to be false, only to claim more than the file can prove in a context where readers rely on it.
• The regime has teeth: under CSRD and ISSB frameworks an overstated claim in an assured filing is a misstatement and a liability, not enthusiasm, and the assurer and regulator both read for it.
• Four patterns to hunt: the overstated claim where the verb outruns the number, the unsupported superlative with no benchmark, the softened negative impact, and the drifted or invented target.
• The softened negative impact is the most dangerous because it is hardest to see and technically true; under double materiality, under-disclosing a negative is as much a misstatement as overstating a positive.
• A drifted target is both a hallucination and a greenwashing risk, because the company becomes publicly committed to something it never board-approved; verify every target, baseline, and commitment word for word against the source of record.
• The reviewer's method reads the narrative as a sequence of discrete claims, asking of each: what in the file makes this true, exactly, running a four-part test for calibration, benchmark, completeness, and faithfulness.
• AI greenwashes by default because the most probable fluent sustainability continuation leans positive and superlative; intervene upstream with a calibrated-factual instruction and downstream with the claim-by-claim review that does not depend on the prompt having worked.
• In the worked example, one fluent AI paragraph contained five greenwashing exposures (two superlatives, two overstatements, one drifted target, one softened negative), and a claim-by-claim review calibrated each to the evidence to produce a defensible disclosure.