When Artificial Intelligence meets greenwashing in court

Since 2015, greenwashing complaints have been surging on both sides of the Atlantic with Volkswagen facing multiple class actions in Germany, Australia, USA and UK for its high-tech rigging of emissions tests.

Since 2015, greenwashing complaints have been surging on both sides of the Atlantic with Volkswagen facing multiple class actions in Germany, Australia, USA and UK for its high-tech rigging of emissions tests, and, as of March 2021, Chevron, the second biggest US oil company by market capitalisation, being accused of using misleading advertisements to create an image of a “climate-friendly” company despite its production plans. In another US lawsuit this time against Exxon Mobil, NGO Beyond Pesticides claims that the company purposefully markets its investments in renewable resources to appeal to increasingly climate-concerned consumers, while such investments remain minuscule compared to its fossil fuel production. The 2019 case brought by environmental NGO ClientEarth against BP under OECD guidelines voiced similar concerns about BP's marketing campaigns "Keep Advancing" and "Possibilities Everywhere" and ended up without trial after the oil major voluntarily withdrew the advertisements in February 2020.

What are the common trends in greenwashing cases?

A few distinguishing features can be derived from these greenwashing lawsuits:

1. They have been mostly filed by environmental NGOs or as class actions;
2. They have typically targeted public advertising or marketing campaigns; and
3. Evidence to support them was gathered by the combined forces of concerned environmental NGOs and citizens.

There is a logical explanation for the above trends – gathering evidence to substantiate greenwashing allegations may be difficult without access to a company’s internal documents. Reliance on public marketing and advertising combined with information about a company's investments, which is often available for a nominal fee via the relevant country's company register (with the exception of some corporate secrecy jurisdictions), makes the NGO-sponsored greenwashing lawsuits against advertising campaigns more likely to result in a successful outcome.  

However, these trends seem to be uniquely confined to greenwashing cases, which by definition, seek to protect consumers against being misled as to the corporate's environmental and climate change credentials. Large-scale public marketing campaigns and, with an increasing frequency, targeted advertisement on social media, are the primary means by which corporations communicate with their customers. Some of these customers include climate-concerned citizens as well as environmental lawyers, who may use such marketing campaigns as evidence in support of a greenwashing case. In other words, strategic assessment of viability of any greenwashing case is made much easier if supporting evidence is readily available in the public domain.

Nevertheless, there are other sources of evidence, which are arguably less readily available and much more opaque. Corporations regularly report on their corporate social responsibility either in dedicated publications or as part of their annual reports, which can run to hundreds of pages. The difficulty of identifying and analysing such evidence may be the reason why not many greenwashing cases have been brought based on evidence gathered from these publications. However, recent advancements in Artificial Intelligence (AI) may soon result in a surge of what some call robots-enabled greenwashing litigation.

The use of AI in uncovering a greenwash

Natural Language Processing (NLP) is an extremely powerful field of AI that effectively enables computers to understand and analyse human language in its written or spoken form. Just think of AI personal assistant Siri on your iPhone or virtual assistant Alexa developed by Amazon – both can understand simple questions and respond either with specific answers or references to online resources where such answers can be found.

Turning to climate change litigation, NLP can prove instrumental in uncovering incomplete or selective corporate reporting. ClimateBert, an AI-powered deep neural language model created by Swiss and German academics, is one example of an NLP tool developed specifically to target corporate reporting and disclosures. ClimateBert is said to have been trained on thousands of sentences related to climate risk disclosures and then used to review reports filed by 800 companies that have embraced Task Force on Climate-related Financial Disclosures (TCFD) reporting. According to the researchers it found that most have engaged in very selective reporting and disclosed mostly non-material climate risk information. Findings like this may be instrumental in advancing many nascent greenwashing cases to a successful conclusion. ClimateBert developers intend to make the tool publicly available, allowing any climate-concerned citizen or NGO to benchmark companies’ TCFD reporting. They are also working on an extended version of the model that would enable tracking of companies’ greenwashing over time. This and other AI-powered tools could make it much easier for stakeholders to identify under- or over-reporting of climate risk and readiness, and potentially unearth evidence of greenwashing.

Another example of the use of AI to review corporate reporting comes from The Financial Stability Board itself. The Board used an AI tool developed by PwC to examine companies’ financial disclosures and score them in its 2020 Status Report. It reviewed financial filings, annual reports, integrated reports, and sustainability reports of 1,701 large companies from 69 countries in eight industries. The AI technology used was trained on a set of labelled data comprising passages of text or excerpts identified as in line with the TCFD’s 11 recommended disclosures. The data was collected in a laborious process in 2018, 2019 and 2020 by a dedicated group who manually reviewed publicly available reports from a sample of 150 “high-disclosing” companies in a process that essentially narrowed down each recommended disclosure to a single yes-no question. For example, one such question in Governance recommendation was “Does the company describe the board’s or a board committee's oversight of climate-related risks or opportunities?”, while one question in Strategy recommendation asked “Does the company describe the climate-related risks or opportunities it has identified?”. The review process was followed by a peer review to ensure consistency of approach and reliability of the data collected.

The labelled data was then used to train the AI technology to understand those human judgments and then test its understanding on unseen and unreviewed company reports. In 2020, further training was undertaken in which the model learnt how to assign score threshold for each question ranging from high to low.  The resulting tool was still not perfect in its ability to assess the existing disclosures - it could only indicate whether or not a company reported in line with the TCFD recommendations (an approach pre-determined by training a model on a set of yes-no questions), as well as whether the reporting failure (as the case may be) was severe or not, but it could not provide analysis of the results or suggest improvements.

For its 2021 Status Report, TCFD used a different AI model provided by Moody’s. This AI model used computer vision techniques to identify relevant text passages and paragraph boundaries extracted from various documents. This approach identified thousands of paragraph passages, which needed to be narrowed down to only those relevant to climate-related disclosures. A language model–based information retrieval technique was used to score the relevance of each passage and only top ranked passages were then selected as relevant. At the end of this process, a language model fine-tuned for climate disclosure classification was used to determine if an entity’s report aligned with each of the TCFD’s 11 disclosures. The performance of the model was finally validated by a human annotator – if the model indicated that a disclosure of a particular company was aligned with the TCFD recommendations and the human annotator was in agreement, the assessment was marked as correctly classified.  

Despite their limitations and the need for human involvement at some stages of the process, both ClimateBert and the TCFD AI models have provided an invaluable insight into corporate TCFD-aligned reporting, which may otherwise have remained undetected on thousands of pages of annual and sustainability reports.

Challenges to widespread development of AI tools

Even though ClimateBert and the TCFD AI models signal a future of true AI-enabled corporate transparency and indeed a potential increase in robot-enabled greenwashing suits, there are certain barriers to a widespread adoption of AI in both reporting and litigation.  

Most significantly, lack of assessment accuracy due to insufficient training data is the obvious challenge to any AI-generated evidence. However, especially in greenwashing lawsuits, consistently structured data will arguably no longer be in scarcity as climate risk disclosures are on track to become mandatory in some countries, most notably the UK or New Zealand. With the relevant legislation is in place for a number of companies, starting with premium-listed companies in the UK, these publicly available disclosure reports will effectively provide an open source data for training AI models.

In addition, AI models like ClimateBert make it cost-effective to identify a shortlist of companies which might be engaging in greenwashing. This may result in significant savings for potential claimants, who could then focus their more traditional legal research and analysis on small group of selected targets.

In light of this, it is no overstatement to say that AI-enabled claims may be the future of greenwashing litigation.