Imitation of life in a world of AI

All the outcomes we cannot see

ESG focuses on long-term, durable results and a sustainable business model. By its nature, it focuses companies on outcomes and externalities. This is ultimately the point behind all those ESG commitments and goals companies set.

As with ESG and sustainability, the meaning of outcomes and externalities can be conflated, and one might be missed. In managing risks and creating opportunities, companies may only focus on externalities, mainly because these may have tangible or financial impacts across the Environment and Social that can be envisioned. While this is impact management, it is also one way to manage reputational risk. However, outcomes may be focused on Governance and require a much deeper look.

Let’s define each quickly, at least for this read:

• Outcomes - consequences for the company of action or inaction.

• Externalities - an impact from the company’s operations on an outside party

Unlike ESG and sustainability, the differences here are not often discussed. Still, they are just different enough that not recognizing them can put the company at a disadvantage and create harder-to-see harm for the company and its stakeholders.

To avoid certain externalities, companies will look into the future across their existing product or solution set and consider ways to prevent the negative impact. Outcomes may represent an unrecognized or ignored gamble for financial growth. An outcome is tied to decisions made in the past or the present that could have consequences for companies tomorrow. For example, the litigation attempting to hold fossil fuel companies accountable is an example of an outcome.

Companies focus on externalities when managing reputational risk, possibly overlooking outcomes for existing systems and processes. Sometimes, the focus on the positive and negative externalities is so intense that an unknown outcome is never or rarely discovered.

Take the case of Henrietta Lacks. Lacks was an African American tobacco farmer, mother, and subject of the book The Immortal Life of Henrietta Lacks. In early 1951, Lacks was treated at Johns Hopkins for cervical cancer. Unbeknownst to her, both healthy and cancerous cells were taken from her body. The cells turned out to be highly resilient, lasting longer and reproducing faster than other cells available. This made them perfect for scientific study since other cells would only last a few days.

Lacks passed in August 1951. Scientists used these cells, named HeLa, for Henrietta Lacks’s first and last name. Her cells lived on and drove critical scientific research. It wasn’t until the 1970s that the family discovered that these cells were taken and the outcomes.

HeLa cells have been used for decades and have had massive commercial implications, even in testing mRNA COVID-19 vaccines. In 2021, the Lacks family sued Thermo Fisher Scientific for profiting from HeLa cells, settling for an undisclosed amount in 2023.

There are many humanitarian considerations, privacy concerns, and potential racial issues in this case, but I want to highlight Lacks’ case for a different reason.

Her extraordinary cells have benefitted commercial industries like modern pharmaceuticals, biotech, and healthcare. Patients have been saved, and companies have reaped immense commercial rewards. She is well-recognized for her unwilling and nonconsensual contributions, but these ongoing extractive efforts had consequences for at least one company decades after her death.

In a similar case, Moore v. Regents of the University of California, the Supreme Court of California decided that patients had no rights to excised cells or any profits from them.

I don’t usually weigh in with emotions or judgment in these reads, but the Lacks and Moore stories gnaw at the edges of my mind and heart. While Lacks has been credited for her unwilling contribution to science, it doesn’t feel like enough.

We’re seeing this scenario again in another industry, but in a surprisingly similar way and on a frightening scale. Yes, I’m writing about large language models (LLMs). The more I learn about AI and read about Responsible AI trends, the more I’m convinced that Technology is another ESG pillar.

LLMs are AI models that have been trained on a similarly renewable resource: the information we all publish on the Internet. It is scraped and gathered without consent, and the sheer scale is staggering. Yet, any potential outcome is cast aside for the externalities it may cause.

The focus of LLMs and Generative AI, even Machine Learning, is now Responsible AI. Responsible AI is a thoughtful approach to avoiding the well-established negative externalities of these models—the potential harm they can cause at scale. With proper management, the commercial outcomes are ultimately preserved through brand reputation if a company follows its principles. Responsible AI is really good at managing externalities.

Still, the same fundamental questions are at stake with LLMs and data as we find in Lacks and her cells. Rather than physical cells, we need to consider intangible data. Just as the HeLa cells are immortally renewable and were presented as an opportunity, data on the internet is infinitely reusable and publicly available.

The temptation to commercialize the opportunities without considering the outcomes on the company is massive. It is more straightforward to deploy and roll back a poorly constructed model and deal with the fallout than to ask the hard questions about its creation in the first place. Again, we place externalities in front of outcomes.

Does data’s mere existence on the internet give its rights away to be commercialized? In other words, as I write this article, have I excised my thoughts to the digital realm, giving up my rights? Regulation around digital privacy and search engines could provide a clue as to the consequences of this pursuit.

Under the EU’s GDPR Article 17, citizens have a right to be forgotten. This means that an individual can request that a company remove infinitely reusable data about themselves once that data has been shared. Could I request that all AI companies remove my blog from their models?

After all, search engines are technically commercial models built from other’s information. Yet, there are ways to block a search engine from indexing your page.

Still, there is no equivalent regulation or methodology for disallowing public or copyrighted content for LLM training. Substack has a setting to disallow AI training models that honor that setting, but there’s nothing stopping a model from using my content regardless, nor would I be able to track down its use inside of an LLM to ask for its removal.

In other words, there is a right to be forgotten for personal information, but there is no equivalent to forgetting a person’s content. In fact, with how the courts translate ‘fair use,’ it may never be litigated in favor of the content creator. This is due to ‘transformative use.’ Per Justia, “This means that the new work has significantly changed the appearance or nature of the copyrighted work.”

For example, a parody of a song is transformative and could be considered fair use, but lifting content from a work to create another work may not be.

This is no revelation, and fair use remains a concern. Unite.AI wrote a piece in January 2024 that covers some of the recent IP-based litigation and the challenges in removing content from LLMs. It references a whitepaper that Microsoft (where I work) wrote in October 2023, ultimately concluding:

The imperative for ethical, legal, and responsible AI has never been clearer. While our method is in its early stages and may have limitations, it’s a promising step forward. Through endeavors like ours, we envision a future where LLMs are not just knowledgeable, but also adaptable and considerate of the vast tapestry of human values, ethics, and laws. 

This is a reasonable and proactive take on a potential outcome that LLM companies may need to address - removing copyrighted content or other intellectual property due to fair use concerns.

If a ‘right to be forgotten’ style of regulation is enacted for LLMs, what might happen to those models? As the above Microsoft Research shows, significant adjustments to the model’s performance could occur. If a piece of data is that critical to the LLM, is it transformative enough?

One could argue that public internet content is already processed by people and used in unexpected ways. Yet, any attempt to anthropomorphize current AI models is a poor imitation. However, once Artificial General Intelligence (AGI) is reached, the argument becomes WAY more complicated. An AGI model is on par or more intelligent than a human.

Can copyright laws survive when the model is an imitation of life?

If nothing else, these examples should provide your company with pause as it explores its ESG issues. ESG is only partly about looking at externalities and the related reputational risk that could be caused. Understanding ESG means considering outcomes as well while examining past and present decisions to determine what consequences you might be overlooking.

Anything less than a holistic view, especially as you integrate new technologies, is only a pale imitation of ESG.