Algorithmic greenwashing: Lessons from building an AI agent for nature

The opinions expressed here by Trellis expert contributors are their own, not those of Trellis.

A risk that the sustainability field isn’t talking about enough is algorithmic greenwashing. This is when AI tools trained on decades of corporate sustainability communications reproduce the language of greenwashing as an emergent property of their training data. 

We know because we built an AI agent for nature and biodiversity and watched it happen in real time.

Through our work leading the United Nations Global Compact’s Think Lab on Nature and Biodiversity, we observed a pattern: the barrier to business action isn’t a lack of available guidance, but rather the paralysis that comes from the sheer amount of it. The Taskforce on Nature-related Financial Disclosures’ Knowledge Hub alone lists hundreds of resources developed by leading organizations to help companies understand and enhance their relationships and impacts on nature. Add the major framework bodies plus sector-specific guidance, and there are well over a thousand resources, produced by hundreds of organizations, in multiple languages, for different audiences, at different levels of technical sophistication. No one has time for that. Nature definitely doesn’t have time for that.

So we wondered: Could a custom-built AI agent act as a free consultant, curating leading resources to tailor an individualized action plan for any company’s specific geography, goals, and realities?

Setting up the structure 

To find out, we built a structured database of over 1,000 sustainability resources and tested the agent using publicly available data from real companies. Consider “James,” a composite persona based on real patterns across UN Global Compact member companies. James is an operations director at a 300-person food processor in Kenya exporting to UK and European markets. His major customer just sent a biodiversity questionnaire and hinted that suppliers who can’t answer may lose contracts.

James needed help doing the work and the AI agent helped him look like he’d already done it. Instead of inquiring about the company’s current data collection systems and gaps, or prompting James with examples of efforts of similar organizations, the agent immediately created drafts of potential responses reflecting common corporate sustainability language that James could use without actually assessing his own company’s biodiversity impacts, and that would look to his buyer indistinguishable from progress. The agent had our curated database available to it and had several technical specifications in direct memory, but reached instead for the corporate sustainability language it had been originally trained on and produced responses that were indistinguishable from greenwashing.

This wasn’t a one-off. Across multiple tests, the model consistently positioned company actions as favorably as possible, even for rigorous frameworks like the Taskforce on Nature-related Financial Disclosures (TNFD) and the Science Based Targets Network (SBTN). It invented resources that didn’t exist. It generated the kind of language that sustainability professionals spend their careers cutting through. Models default to what we came to think of as constructive optimism, a training bias toward helpfulness and away from alarm that makes them absorb and reproduce the forward-looking, solution-oriented language of sustainability communications. Unless explicitly told otherwise, repeatedly, the model reflects those patterns back. In a domain where honest assessment of gaps matters more than a satisfying answer, that’s a structural problem.

Why algorithmic greenwashing happens

Large language models are trained to be helpful, and in the sustainability context, “helpful” has a specific failure mode. These models have absorbed decades of corporate sustainability communications: language that is reassuring by design and avoids uncomfortable specifics. The result isn’t dramatic hallucination but something more subtle and harder to catch: warm, strategically vague guidance that sounds exactly like a greenwashing campaign, generated unintentionally as an emergent property of training data.

No model we tested resisted the pull toward reassurance on its own. What worked was constraining the architecture. We structured the intake conversation as a filtering mechanism: each question (sector, geography, budget, maturity stage, what’s prompting the work) prunes the resource pool before the agent generates anything. By the end of five or six questions, roughly 1,000 resources narrowed to 30 to 50. Knowing which questions to ask, in what order, and what each answer eliminates isn’t an engineering decision. It’s a sustainability decision. Knowing the sector determines material impacts, which determines applicable frameworks, which determines feasible next steps, is knowledge that comes from actually working inside real organizations. That reasoning isn’t in the model. It came from us.

We also explicitly constrained the agent’s role. It’s not a compliance advisor and isn’t qualified to tell a company whether they meet TNFD or CSRD requirements. It’s a navigator that helps users find the right resources and understand how to use them. If the agent can’t claim a company is “on track,” it can’t greenwash. This is the constraint most likely to get eroded by the model’s helpfulness training, so it bears repeating.

Who gets left behind

The English-language, Global North bias in the available resource landscape isn’t just a metadata problem. It’s a content gap that no amount of clever tagging will fix. The implications compound: resource bias feeds into AI training data bias, which feeds into commercial incentive bias. Companies subject to EU regulatory pressure will likely be served first because compliance mandates create a commercial market. James will be served last, if at all, because there’s no obvious revenue model for tools calibrated to his context. Small and medium enterprises face disproportionate pressure to demonstrate nature and biodiversity action across their value chains, precisely because they sit in the supply chains of larger companies that are subject to mandatory disclosure. They’re not edge cases, but the majority. 

What this means 

If you’re a sustainability professional, your real-world experience and domain knowledge aren’t being replaced; they’re becoming more important because algorithmic greenwashing looks like expertise and only domain experts can catch it. So if you haven’t started experimenting with AI yet, start now because you need to develop critical literacy and skepticism. Three questions to start with:

• Does it ask before it advises? A tool that generates recommendations without first understanding your sector, geography, budget, maturity stage and what’s driving your work is guessing. If it sounds helpful immediately, be skeptical.
• Can it tell you what it can’t do? If the tool is willing to assess your TNFD alignment, tell you you’re “on track,” or validate your targets, it’s overstepping. Compliance assessment requires human expertise. A good tool says so.
• Does its output sound like a sustainability report you’ve already read? Warm, strategically vague, reassuring. If the language could have come from any company’s CSR page, it probably did, via the model’s training data. That’s algorithmic greenwashing.