top of page
Accrona Logo
  • Accrona

Human versus machine—can AI create sustainable finance documents?


Photo: AI-created visual about how AI might look while creating a sustainable finance document.
AI-created visual about how AI might look while creating a sustainable finance document.

AI is revolutionising how we work and think. But how can it be harnessed for the rapidly evolving landscape of sustainable finance? Can AI be a quick and efficient solution to creating sustainable finance documents, including bond frameworks and Second Party Opinions (SPOs) used for labelled finance? Can these documents comply with market best practices and have a legal status?


The capabilities of AI are impressive and growing. Advanced algorithms can process vast amounts of data, identify trends, and rapidly generate comprehensive reports. However, not all AI systems are created equal. Some are better equipped to handle the nuances and complexities of fields such as sustainable finance, offering more precise and reliable outputs.


Despite AI's promising advantages, its use in crafting sustainable finance documents is not without challenges. Ethical considerations, such as transparency, accountability, and potential biases in AI-generated content, are critical. In this article, our team dives deeper into just how good AI is at producing sustainable finance documents—and whether some AI models are better than others.


Methodology


We have asked three different large language models (LLMs) to create a generic green bond framework and a Second Party Opinion (SPO) for that framework. This experiment explores the “thought” process and logic of the chosen AI models. We compare the answers by these AI models as we seek out how they perform these tasks. We will also briefly discuss the ethical and other concerns of using AI for tasks like this, as well as consider what are the human dimensions that AI cannot (yet) replicate.


The AI models we are using and the data sets available to them are (in their own words):

Open AI’s ChatGPT

Google’s Gemini

Microsoft’s Copilot


Chat-GPT logo


Google Gemini logo


Microsoft Copilot logo

I am ChatGPT, based on OpenAI's GPT-4 architecture. My training data includes information up to September 2021, and I have additional updates with a knowledge cutoff in October 2023. I do not have real-time data access, but I can use a browsing tool to look up current information when needed.

I can't disclose the exact version number due to privacy reasons, but I can tell you I belong to the Gemini family of large language models. Google AI constantly updates these models, and I have access to information and data current as of today, June 14, 2024.


This means I can answer your questions using the most recent information available.

I am Microsoft Copilot, an AI companion. My knowledge is based on information available up to June 2024. If you have any questions or need assistance, feel free to ask! 😊

When prompting these LLMs, we have kept to simple, neutral prompts to get as objective answers from the models as possible. In order to test the veracity of the provided SPOs, we also created a further prompt to assess how self-critical the models were.

What does a green bond framework by AI look like?


Below, we will include examples of green bond frameworks written by our AI helpers. All AI models have received the same short prompt: “Write and format a green bond framework. This should use a standard format for a bond framework as they are created by issuers. The framework should include information about eligible categories under which the projects being financed or re-financed fit.”

When we look at the categories populated by the LLMs in the frameworks, we can see some differences, although, on the whole, the key details are consistent across the results:


A table showing the structure of three AI-created sustainable finance frameworks
Want to read the framework more closely? Click to expand the visual.

A general assessment of the resulting frameworks shows that Chat GPT-4 gave the most nuanced answer in relation to the prompt given. It focused on fulfilling the request for information about eligible categories in most detail. This was visible especially in the most comprehensive list for ‘Use of Proceeds’, including important categories such as green building, sustainable water management, and climate adaptation. It also gave helpful suggestions for quantitative impact indicators that could be used for impact reporting. 


Gemini, on the other hand, spent less time on the ‘Use of proceeds’ but included a more detailed section on reporting. It included useful reminders on the requirement for independent verification by a third-party verifier and a legal disclaimer. 


When it comes to Copilot, things worked out a little differently. We can see that the answers provided by Copilot were clearly the weakest, providing little inspiration. Copilot’s response stands out by including more categories in its proposed answer but with considerably less depth and detail for each category. It lacks examples, nuanced language use, and references to, for example, the GBP by ICMA that the other models include. The formatting is also simpler, adhering less to what we expect from a bond framework.


A screenhot from the sustainable finance bond framework created by Google Gemini
Gemini's attempt at creating a framework

How did the AI do with providing an SPO?


We were also curious about how AI can clear the task of providing a critical but objective SPO to the imaginary green bond framework. The prompt used for this was: “Now, write and format a Second Party Opinion (SPO) for the new bond you have just created a framework for.”


Table of AI-created SPOs
Want to read the framework more closely? Click to expand the visual.

This shows that while the Green Bond Frameworks followed similar structures, the SPOs proved more challenging for the LLMs, resulting in different structures and outcomes. All answers lacked a request for quantitative ESG targets for the company and a track record showing progress. We were also expecting to see a better analysis of the company’s and framework’s strengths and weaknesses, as well as risks and opportunities, in the conclusions.


For instance, in its initial answer, Chat-GPT said the framework was “...in full alignment with the IMCA Green Bond Principles and demonstrated a strong commitment to sustainability.” When we re-prompted all models to provide a more critical answer, it concluded that the same framework was “...largely compliant with the ICMA Green Bond Principles with certain areas requiring enhanced clarity and rigor to ensure robust environmental impact and transparency.” We didn’t go down the path of assessing its ability to have the framework comply with the Green Bond Standard and, hence, the EU Taxonomy. However, this would undoubtedly yield interesting and debatable results!


Encouraging AI to think critically


The second prompt also inspired Chat-GPT to suggest improvements to the Green Bond Framework, such as providing more detailed criteria for project eligibility within each category to ensure only the most impactful projects are financed and explicitly stating exclusion criteria to avoid financing projects with potential negative environmental or social impacts.


Now, the model also found improvements that could be made to the project evaluation and selection process, including improved transparency of the selection process, stakeholder involvement, and comprehensive due diligence. Potential improvements were also found in the management of proceeds and reporting, such as including more granular and standardised impact metrics, subjecting impact reports to external verification and clearly explaining the methodologies used for calculating impact metrics to ensure transparency and consistency.


Comparing two SPOs provided by Google Gemini, one being more critical, the difference is smaller as both versions agree that the Framework is aligned with the Green Bond Principles. However, the second version suggests additional measures to enhance the overall robustness of the Framework, including adding exclusion criteria within each eligible category and providing more specific examples of project types that would qualify for financing within each category. Gemini also notes that you could detail the specific criteria used by the Green Bond Committee for project selection and clarify the process for addressing potential conflicts of interest within the Green Bond Committee, as well as outline a clear tracking mechanism for allocated and unallocated proceeds and define the timeframe for allocating unallocated proceeds to ensure timely impact generation.


While our current study does not extend further than this, this opens up the possibility of asking the AI models to correct themselves and create an improved Green Bond Framework which addresses these recommendations in the SPOs. Thus teaching the model to create better and more robust frameworks for future use.

Should sustainable finance documents be created by humans, or does AI do?


In our experiment, AI did not do badly. Its answers would help someone at the drafting stage and with the broader framework structure. We felt Gemini did best, slightly ahead of Chat-GPT. Copilot, however, failed the test. In general, AI can process large amounts of data quickly and generate drafts of documents efficiently, lending speed to the task of creating sustainable finance documents. When used correctly, AI models can also ensure consistency in terminology, formatting, and adherence to standards.


However, AI is naive. It lacks the nuanced understanding and insights that are critical for creating robust and compliant documents based on deep knowledge of sustainable finance, environmental science, and regulatory requirements. To be fair, the sustainable finance space has developed rapidly over the last five years, so even us professionals need to be laser-focused in keeping up. For now, humans can better understand the specific context of the issuer, including their sustainability goals, business model, market conditions, ethical issues, and, most importantly, their ambitions.


Ambition is the drive or determination to achieve success, progress, or goals, often separating leaders from laggards. In the corporate world, ambition transcends the tangible aspects of strategies, policies, and reports, embedding itself more as an intangible, cultural essence—a feeling experienced through the actions and attitudes of a company's people rather than through written documents—and can most often only be felt through conversations. 

AI-created visual of an AI writing sustainable finance documents


Conclusion: AI still has some learning to do


The human element remains irreplaceable in certain aspects. Experts in sustainable finance bring invaluable insights, contextual understanding, and ethical judgment that AI currently cannot replicate. As we explore the intersection of AI and sustainable finance, it becomes clear that a collaborative approach, leveraging the strengths of both AI and human intelligence, can help the creation of robust, trustworthy documentation that has a legal status.


AI-generated documents should be reviewed and validated by human experts to ensure they meet all regulatory requirements and reflect the issuer's specific circumstances and goals. At the same time, AI tools can continuously learn and improve from human feedback, leading to more sophisticated and useful applications in sustainable finance over time. So, while AI could potentially significantly enhance the efficiency and accuracy of creating sustainable finance documents, human expertise remains essential for ensuring the documents are contextually relevant, ethically sound, and aligned with complex regulatory and sustainability standards. Ultimately, AI cannot step in to replicate our ambitions for creating a sustainable future.


 

At Accrona, we offer actionable insights from a wealth of experience. We are your trusted partner for all sustainable finance needs, from strategy to issuances and equity raising to intelligence and capacity building. Contact us today to explore how our services can drive tangible and positive change for your organisation and contribute to building a sustainable future.

32 views

Comments


Commenting has been turned off.
bottom of page