AI and deep learning can instantly spot corporate greenwashing
By Sarah Zeines
April 15, 2021
NEWS - INTERVIEW
Sustainability is the business world's new mantra. But how many companies actually commit to greener ways? Hard to say, right? But thanks to ClimateBert, a deep natural language-based system, greenwashers will now have a harder time hiding the difference between spin, PR and action. That’s good news. The bad news, though, is that when ClimateBert (BERT is a machine learning technique) was finally let loose by its creators, they uncovered a pretty sad truth: the financial world's support for climate action “is mostly cheap talk and that firms cherry-pick to report primarily non-material climate risk information.”
Julia Bingler, Mathias Kraus and Markus Leippold, the three Swiss and German researchers behind ClimateBert got to their sobering conclusion by using AI to analyse the disclosures of close to 800 companies that adhere to the Task Force For Climate Related Financial Disclosures (TCFD).
Will ClimateBert serve as a deterrent? We can only hope so. For its creators, though, “the only way out of this dilemma is to turn voluntary reporting into regulatory disclosures.” A move that is already happening. TCFD disclosures are mandatory in the UK and Switzerland.
Sustained talked recently with Julia Bingler, head economist of the new climate algorithm that companies should now watch out for.
What does ClimateBert do, exactly?
It extracts climate-related information published in annual company reports, differentiated by four categories: governance, strategy, risk management, and metrics and targets. These categories have been recommended as useful disclosure structures by the Task Force for Climate-Related Financial Disclosures (TCFD). We focus on climate risks, so investors can better assess whether their investments are prone to certain climate-related risks or not.
The project was designed to identify how corporate climate disclosures developed over time. We initially focused on the disclosures of those companies that voluntarily support the TCFD recommendations. We find that on average, supporting the TCFD is not associated with considerable increase in climate-related disclosures, and that companies tend to be reluctant to disclose material information.
How did the project come into being?
Two years ago, Mathias, the computer guy, and I were chatting over a beer about our research. I used to work on sustainable finance and more specifically on TCFD, Mathias on using AI to improve business decisions. He was interested in doing more on climate, I wanted to assess whether the TCFD recommendations really had an effect on disclosures. So we started an AI research project to extract and track climate-related disclosures by companies in 2019. At some point, we realised that researchers from the University of Zurich started to work on a similar project—so we joined forces.
How does your algorithm function?
Our ClimateBert algorithm builds on a recent version of a state of the art NLP (Natural Language Processing) model, the BERT (Bidirectional Encoder Representations from Transformers) model. We trained it on thousands of human-labeled sentences with climate risk content from company reports. ClimateBert is able to detect, extract and classify climate-related disclosures in company filings, differentiated by the four TCFD categories governance, strategy, risk management, and metrics and targets. As a first step, we used the algorithm to assess the amount and categories of climate-related information of more than 800 TCFD-supporting companies for a period of six years. In addition, we are currently working on an extended version of our ClimateBert to enable better tracking of companies’ greenwashing over time.
How many of the companies analysed by ClimateBert were red flagged?
ClimateBert was not designed to red flag individual companies. We would rather use it to assess the overall level of climate disclosures in the global market, in individual sectors or in different countries and regions. Regardless, we found that climate disclosures in annual reports did not considerably increase when comparing the pre-TCFD period to after 2017. We only assessed the annual reports because this is where companies have to disclose material information.
Switzerland, for instance, is often criticised for its lack of transparency. How do you control companies that keep their practices in the dark?
We cannot force companies to follow TCFD guidelines, as a financial authority or democratic vote could. However, we can highlight detrimental practices. In Switzerland, a mandatory approach is currently being discussed by the FINMA.
Who is the technology addressed to?
ClimateBert is not tied to Switzerland. It is something we created in order to understand how companies respond to the TCFD recommendations on a global scale. Governments can use it to see if companies comply with climate-related disclosure regulations. It can be used for any firm on the global level that publishes annual reports.
Could you control governmental climate policies with your software?
The algorithm has been trained on company disclosures. This wording is very different from climate policy wording. However, climate-related policies of governmental agencies like communal water and electricity providers could be assessed.
How will ClimateBert go global?
It can be applied at global scale today. In addition, ClimateBert will become more specific to enable users to better track companies’ climate commitments and to identify possible greenwashing. We will make the algorithm publicly available, most likely during the course of this year. It can then be used by whoever is interested in or responsible for disclosure regulations. This includes ministries of economy, financial regulatory institutions, financial analysists, researchers and others, depending on the context.