The only thing faster than someone gaining ESG credentials is someone becoming a Generative AI expert. Luckily, I’m at some strange intersection of both topics, albeit not a Generative AI expert yet. Still, I’ve written about the topic before in my ChatESG series (Part I and Part II) around the intersection of Technology as an ESG risk and opportunity. This time, I thought I’d cover Generative AI and Large Language Models (LLMs) head-on, as financial services and companies may look to these tools to help understand data around ESG.
A quick Generative AI Primer
Somewhere inside AI (as a discipline), Machine Learning (allowing a computer to learn without specific instructions), and Deep Learning (a neural network that can understand complex patterns), lies Generative AI. While Machine Learning and Deep Learning are excellent at numbers and prediction, Generative AI reminds me of the Liberal Arts and Humanities. These models are great for understanding qualitative information and generating text, images, music, etc.
In other words, if you need forecasting, classification, and mathematical analysis, Machine or Deep Learning might be better suited. For example, based on a company's last 20 years of financial data, a Machine Learning model can attempt to predict what will happen next quarter. Generative AI is a better tool for summarizing a large amount of text, generating new content or code, searching datasets, and extracting information (including numbers). For example, a Generative AI model can summarize an industry report in five key bullet points.
This article covers Generative AI and LLMs, which means models trained on text data rather than images, music, or other content. As you probably know, ChatGPT is one of the more famous LLMs. LLMs can be fine-tuned with domain-specific datasets for different applications. For example, a company can integrate its enterprise data and allow employees to query a tool against this dataset using natural language to get company-specific responses or generate new content in the company’s voice.
Generative AI fits new qualitative ESG data use cases
There is a broad range of ESG data types and different ways to look at the data. For this article, let’s focus on two specific data perspectives, quantitative and qualitative. First, there are quantitative, measurable metrics around Environmental and Social information. These metrics are what companies report in their CSR tables, aligned with standards like GRI and regulations like CSRD. The metrics might be metric tons of CO2e, KwH of energy, gender/age/racial diversity headcounts, etc.
On the other hand, there is qualitative data around material and non-material ESG projects, intangible concepts (ethics, Governance), and hard-to-quantify storytelling that the company publishes and surrounds the company. This data may be less numbers-based and more content-heavy. For example, a company may describe how it engaged with a supplier on their upstream environmental transformation, assurance on human slavery issues, and creating local economic opportunity.
Generative AI does very well in dealing with text-based information like this. Of course, these reports may have numbers that Generative AI can extract. Specific to ESG, it may be able to summarize the critical points of a CSR report, search through and help complete ESG-related RFPs, analyze regulations and standards reporting (CDP, TCFD formats), help understand ESG information through another perspective, and more.
How can Generative AI understand ESG?
As mentioned above, I’ve covered many AI challenges in the ChatESG series (Part I and Part II). While there is a lot of opportunity to capture in ESG analysis, there is one big challenge specific to applying Generative AI to ESG, which financial services firms and corporates must be mindful of.
To get unique value out of tools like OpenAI and ChatGPT, a company would start with a Pre-trained dataset (that’s the P part of GPT), of which there are four in the OpenAI base models to choose from. They increase in complexity, cost, and the size of the training model used as the alphabet proceeds from Ada, Babbage, Curie, and Davinci. From there, the company would connect high-quality domain-specific data aligned to a specific task, like ESG information, for analysis.
With the data connected, the company can start to fine-tune the model. This may include instructing on the desired task in the input by providing sample prompts, how to solve the problem, and outcome examples (Prompt Engineering). It may also include Reinforcement Learning with Human Feedback (RLHF), where a human selects the appropriate outcomes to reward and train the model.
Regarding responding with relevant outcomes from the provided ESG data, the model’s success hinges on one unique ESG problem: its definition. It should be no surprise that language is at the crux of a Large Language Model.
There is no agreed-upon definition of ESG. Some, like me, tout that it is a way for companies to understand their material risks and opportunities to pursue long-term resilience. Others lead with not always material ESG issues around sustainability, carbon reductions, DEI, etc. Both have a universe of trade-offs to consider. ESG and AI are related in this way - if you polled a hundred people on these topics, you would likely get a hundred different answers. This is the crux of the anti-ESG movement and friction with financial services firms in the US.
The challenge of ESG and LLMs lies in the language we use, and Generative AI may force that definition through human interaction via RLHF. Before unleashing a model, the company needs to define its ESG approach so that the human can reinforce the desired outcomes from the model.
The term “ESG” matters less than the outcomes the company wants to achieve, both material and non-material, but it requires speaking and writing in the language of intent toward impact and outcomes. For example, if an employee or analyst asked a domain-specific trained model a question related to ESG, would it respond in the context of materiality and stakeholders or non-material efforts that attempt to save the world? The response may be misaligned if the company’s objective isn’t clear. If the company doesn’t take this approach, it offloads the ESG confusion from humans to machines.
Generative AI models may help mainstream ESG
The misalignment in ESG language and intent permeates entire industries, companies, and US politics. Still, applying Generative AI may correct this misunderstanding by doing something companies have struggled to do: mainstream ESG, or at least Technology Governance, across the company.
Let’s start with Technology Governance. A systematic risk has surfaced around sharing company data with public ChatGPT models. For example, some Samsung engineers shared code with the public models, risking the company’s IP. From here, the company is now drafting policies around ChatGPT’s usage. This type of exposure and subsequent policy can increase quality Governance and bring employee attention to Technology risk in an ever-growing digital world.
More broadly, if the board understands the nuance of ESG, the company can get ahead of the language issue with a perspective on both ESG and non-material ESG issues. To do that means companies and firms need to do something they’ve been holding back on - mainstreaming the idea of ESG by defining it and then educating the company on it to drive consistency. So far, purpose and mission seem to be the preferred way to lead the ESG charge here, with vague references bringing context to employees. Observationally, more structured communication and engagement about the company’s true intentions are needed to drive ESG success and create defensible positions on the topic as the pushback continues.
If the company doesn’t lead with a consistent ESG strategy internally, it won’t be able to execute an effective Generative AI model to deliver these new use cases. The company may even find that it will unproductively scale up the confusion and make indefensible decisions.
The challenge with language doesn’t only occur at the high level of its intersection with the company but also within each siloed pillar of ESG. For example, suppose your company trains a model on diversity data along the Social pillar. In that case, the outcomes from the prompts may range in usefulness, depending on the country context, but the model would need that context. With the right mix of additional datasets surrounding the Social, the model may help inform an HR professional with meaningful context.
Generative AI can also help reduce siloed thinking in other ways. For example, a Chief Sustainability Officer with an Environmental Science background may lead with the Environmental metrics alone to inform decisions. While deciding to switch an energy provider from brown to green might be innocuous, switching material suppliers or making product changes without considering the Social or Governance information may introduce new risks or a challenge that runs counter to the company’s operations, stakeholders, and policies. Generative AI can help pull that context together, helping the CSO to reason over a bigger universe of information. Again, that means the model must be trained on more than just one pillar of data. In other words, the power of ESG is when it is considered together, and the model needs Environmental, Social, and Governance datasets (and likely more) to find these connections to empower the outcomes.
In these examples, Generative AI drives towards nuance, perhaps the most incredible opportunity to capture. It can help people reason over complex and seemingly disconnected sets of information, provide a range of perspectives, and uncover new connections, all of which underpin ESG thinking. If employees are receptive, it can supplement efforts to mainstream ESG across the company.
As companies investigate Generative AI, they will quickly uncover the link to its application to non-financial, text-based ESG information. Mainstreaming ESG should be part of the strategy to make sure the implementation is successful and careful consideration of a range of datasets. If executed thoughtfully, Generative AI can reinforce those mainstreaming efforts with new insights for review. Lastly, just as Social is in the middle of the acronym, humans must be kept at the center of the models, especially for decision-making.
One last note! If you’re heading to GreenFin 23, I’m hosting a Lunch Roundtable on Tuesday on Uncovering ESG Insights with Generative AI and Large Language Models. Don’t miss it!