People Define Your AI Equation

“Responsible AI can’t exist without ethical principles. It operationalizes ethics and is highly strategic when it comes to aligning AI systems with human-centric values. Whenever companies build or implement AI, they must review their design practices, including how they set up modeling parameters and whether this process aligns with the principles they have elevated into their business strategy. In a sense, responsible AI is about compliance with a business's fundamental strategic values."

Olivia Gambelin, Founder, Ethical Intelligence, shared her key leadership lessons in my "Leading Through Disruption" series.

Lauterbach: How did you become an AI Ethicist?

Gambelin: I studied philosophy and ethics and was looking for fields to apply my knowledge in practice. One day, I attended a presentation on data ethics at a conference, which altered my course. I created Ethical Intelligence, an initially academic think tank that evolved into a consultancy examining the crucial intersection of artificial intelligence and human values. As AI permeated various industries, my focus shifted to understanding how organizations were integrating and developing this transformative technology.

Lauterbach: The deployment of AI is still in its infancy, but there is discussion of potential harms associated with AI. Yuval Harari paints a bleak future.

Gambelin: From my perspective, the whole reason for innovation is to drive human impact. When we pursue innovation for its own sake, adverse effects emerge as we lose sight of the broader context. The question of impact gets blurred.

Lauterbach: Ethics evolve with time. In medieval Europe, both of us might have been declared witches and burned at the stake – it was considered ethical as the goal was to save the soul.

Gambelin: There is a concept of moral maturity: humanity matures in its understanding of what constitutes good and evil. Societies evolve, and so do values.

When it comes to scaling AI, people want a checklist, a list of pros and cons. When we embed these values at a hard-coded foundation, we risk something called moral lock-in. We make moral decisions and then lock ourselves into them.

Imagine we had AI back in the 1400s. These AIs would have been built with assumptions that certain women were witches to be burned. The decision would have been locked into the entire concept and scaled accordingly. Today, we must consider the idea of responsibility as we develop and adapt transformative technologies. The challenge with AI is that we can't operate only on the 'ones' and 'zeros' level. We review some traditionally vague and abstract concepts and apply these to concrete outcomes. We must be able to review outputs and shift if needed.

Lauterbach: How should governments and policymakers cope when looking into a field that is in flux?

Gambelin: Regulators implement different values and principles, shaping industrial standards. When it comes to healthcare, we encounter a long-standing tradition of applying such standards. We can review them against companies’ mission statements, and these, in their turn, influence what kind of customers these businesses serve and how they do it. Outcomes emerge from the triangulation of these different inputs.

Lauterbach: In AI, we discuss biases that have emerged due to historical prejudices encoded in data. The gender gap in medical research has entered the development of AI applications, for example. Even if AI makers' intent is noble, there is a tail of biases to work with. Addressing these is challenging. We can spot historical nonsense, such as the 'emergence' of black founding fathers in the Gemini model, but what about cases that aren't so obvious?

Gambelin: With Gemini, AI developing teams over-prioritized the value of inclusivity. 200 or 100 years ago, inclusivity wasn’t as highly regarded as it is today. It is a balancing act of what values to highlight. At Gemini, historical accuracy should have prevailed. The incident highlights the importance of being aware of tradeoffs, emphasizing the need for human judgment.

Lauterbach: Today, many influencers writing about AI assume things that aren’t valid, if we look into the mathematical foundation of these models. For example, Yuval Harari is concerned that AI has compromised language. He writes in his book, “Nexus,” that by hacking language, computers could make it extremely difficult for large numbers of humans to conduct a meaningful public conversation. But machines don't understand language or are even ‘aware’ that they work with it. They process a statistical token stream. And token is not even a word. I think it's like two-thirds of an English word. Our LLMs today lack reasoning and causation, and even if we describe them with human terms, such as ‘response’ or ‘reasoning,’ they don’t grasp any meaning in the sense that humans do. What are your views on the discrepancy between what these models do today and how we assess them?

Gambelin: Current AI architectures are not designed for reasoning, but rather for predictability. Predicting calls requires a very different structure than understanding and causation. I heard once a description of LLMs as “word calculator,” and I liked the comparison! We might have hacked counting words; we did not hack reasoning.

Lauterbach: So, how can we be practical about AI ethics and data ethics in a

product design and life cycle?

Gambelin: Responsible AI can’t exist without ethical principles. It operationalizes ethics and is highly strategic when it comes to aligning AI systems with human-centric values. Whenever companies build or implement AI, they must review their design practices, including how they set up modeling parameters and whether this process aligns with the principles they have elevated into their business strategy. In a sense, responsible AI is about compliance with a business's fundamental strategic values.

Lauterbach: Are you an AI optimist?

Gambelin: I am an AI realist. Responsible AI is about people, process, and technology. Sometimes, there is a process and technology, but team culture doesn’t bring the desired results. Deficits in motivation, a lack of education on human-AI alignment, and shortcomings in how teams communicate about the 'what' and 'how' of their work can jeopardize the best vision. People define your AI equation.