Ask a Data Ethicist: How Do Metaphors Structure Our Thoughts on AI Governance?

Metaphors can be helpful in allowing us to apply known concepts to new things. But, the metaphors we use also structure our thinking, sometimes in limiting ways. As conversations around AI governance abound, how we think about “what AI is like” starts to directly relate to issues of governance. This month’s question asks…

How do metaphors structure our thoughts on AI governance? 

Let’s take a look at a few common AI metaphors: AI as a tool, AI as electricity, AI as a toxic substance, and AI as an agent. But first, a brief word about metaphors and governance.

Governing by Metaphor

“The heart of metaphor is inference … Because we reason in terms of metaphor, the metaphors we use determine a great deal about how we live our lives.”

― George Lakoff, Metaphors We Live By 

In doing some research for a new course I’m developing, I stumbled upon a very expansive definition of governance from the Commission on Global Governance, which was established in 1992, to examine ways to move toward greater global cooperation (bold emphasis is mine):

“Governance is the sum of many ways individuals and institutions, public and private, manage their common affairs. It is a continuing process through which conflicting or diverse interests may be accommodated and co-operative action taken. It includes formal institutions and regimes empowered to enforce compliance, as well as informal arrangements that people and institutions either have agreed to or perceive to be in their interest.”

What I love about this definition is that it frames governance as a continuously negotiated process, not merely a set of static rules, and it invites participation from diverse interests. We tend to think about governance as “the rules” put in place by those in charge, be that government regulations or management policies within organizations. Yet, this historical definition makes room for “informal arrangements” and “cooperative action” in addition to ideas of compliance.

The relationship between governance and metaphors, as it pertains to AI, starts to play out in terms of how we compare it to what we already know, which shapes how we envision managing the risks and benefits. 

AI as a Tool

“AI is like having a knife. You can use it to cut bread or to hurt someone. It’s up to us to decide how to use it.” ― Oren Etzioni

“AI is a powerful tool, but it’s just a tool. It’s up to us to decide how to use it.” ― Yoshua Bengio

The idea that AI is a tool is a popular framing. What kind of tool? Some say a hammer, others say a knife, and some leave it up to you to determine.

Tools are used by people. They require humans to power them. This is what differentiates a tool from a machine. Historian and philosopher of technology, Lewis Mumford frames this in terms of “a technological continuum” involving human bodies and automation. It’s not a clean delineation. We can have a power tool, like a power screwdriver, that still requires a lot of human intervention to use. But, some machines can operate with greater independence, like KUKA the pallet nailing robot. It nails at scale. 

The element of individual responsibility is deeply embedded in the “just a tool” framing of AI. We can do good or bad things with a tool. It’s up to us. There is a level of personal control that is assumed. 

While there is definitely an element of personal responsibility at play in the use of AI systems, the “just a tool” framing puts a lot of onus on the individual and isn’t critical enough of the ways in which AI systems shape the boundaries of what is possible. For example, a recommender system literally structures the universe of choices to highlight particular options. An automated decision system may require active intervention to change or over-ride its outputs. All of this happens with an opacity that makes it difficult to “see” what’s taking place, unlike KUKA the pallet nailing robot or any other physical machine.

Tools also have a universal quality. Everyone can use a knife or a hammer.* If we think of AI as “just a tool,” our governance choices might lean toward ensuring everyone has access (hammers for all!), that the tool is reliable and that individuals know how to use the tool so they don’t hurt themselves or others. We may also punish people who commit harm with the tool, but we don’t blame the tool. Murder with a bomb, gun, knife, or hammer is still murder. And yet, the tool can enable new arrangements or possibilities. For example, guns enable killing at a distance in a way that knives or hammers don’t. Machines, like automated drones, enable even greater distancing. Bombs enable mass killing and can be set to detonate at a specific time. Those rearrangements of time and space matter because they impact our notions of personal responsibility.  

AI as Electricity

“Just as electricity transformed almost everything 100 years ago, today I actually have a hard time thinking of an industry that I don’t think AI will transform in the next several years.” — Andrew Ng 

Data has been called “the new oil.” Computer scientist Andrew Ng extends the energy metaphor for AI as being like electricity, noting how the modern world was transformed by electricity. If we take this metaphor into the governance world, we may think about regulating AI like a utility. We understand that electricity is useful but also very dangerous. That’s why we have widespread access but tightly centralized control over electricity. 

Most people interface with electricity via an electrical outlet or socket. Countries determine the voltage for these sockets. There are standards set for the things that can be safely plugged in and those products are also subject to government regulations. We ensure the safe distribution of electricity through inspections and through licensing electricians. As users, we respect the power of electrical power and teach our children not to stick their fingers in the socket!

“Don’t touch the socket, it’s not a toy,

Don’t touch the socket, it can make you cry,

Electricity is dangerous, it’s true,

So keep your fingers away, it’s the best thing to do.”

— Don’t Touch the Electric Socket, a song for kids 

People are not permitted to spin up a powerplant in their backyard. The companies that generate electricity are centralized and highly regulated, and typically there are rules not only surrounding their operations, but also their ownership, including reviews of foreign ownership. Countries understand that having sovereign control over power production is a national security issue. These layers of infrastructure enable us to deliver electricity safely and equitably. 

As a business, utilities tend to be monopolies or oligopolies. This is necessary to support the business model due to the massive infrastructure investments needed to support a utility. The type of AI we are focused on now – large language models – speaks to similar business model dynamics related to the massive amounts of computing infrastructure necessary to power theses models. 

In this governance scenario the focus is on regulating suppliers and controlling who can generate and distribute electricity. While everyone has access, the governance regime is almost the opposite perspective of a tool in that that there is little left up to personal responsibility other than not electrocuting oneself.

AI as a Toxic Substance

“We are living in an attention economy and social media is addictive by design.” – Tristan Harris

“Social media has become the crack cocaine of the digital world.” – Simon Mainwaring

Social media is powered by machine learning algorithms aka AI. We’re having a moment where regulators across the globe are reckoning with the consequences of social media that have played out over the last decade, especially how these systems have impacted young people. 

There are now an overwhelming number of metaphors that compare social media to a toxic substance. Some people are also including chatbots in this toxic camp, given the impacts they are already having with certain people, inducing psychological harms (AI Psychosis) and even suicides. While not all types of AI fall into this camp, generative AI and the chatbot interface are at the center of the conversation.

In this governance framing, the issues tend to focus on protecting vulnerable people, especially kids. This might include restricted access. The governance solutions are similar to how we might treat alcohol, gambling or adult entertainment. We know it’s “bad” or possibly “addictive,” but we allow people of a certain age to make their own choices because it is pleasurable.

We also regulate the supply of these products and services to varying degrees. Here in Canada, certain provinces will tightly control the distribution of alcohol through government liquor stores. Conversely, in the U.S., you can pick up a six-pack at your local Walgreens. The control of supply differs between and within countries. 

We are at the early stages of governing social media, and possibly some chatbots, this way. Australia led the charge with a ban for kids under 16 years of age. Canada is the midst of a proposed regulation aimed at curtailing access, the UK is examining something similar and certain states have also enacted age restrictions.

There are also efforts to try and remediate AI to be less toxic. We could think of this as eliminating certain chemicals from the ingredients list of a product in order to make it safer. This line of thinking puts the onus on the tech companies to improve their products with more guardrails and oversight mechanisms. Governance through this lens might include a vetting process, inspection, or audit to prove that the product has been made in a safe manner. Yet, given the nature of how generative AI is constructed and its probabilistic qualities, it is questionable whether or not it’s fixable or if enough guardrails can be put in place to make it safe for everyone. 

Finally, we also have highly toxic substances that are so lethal, they are essentially banned. This type of highly toxic AI is something that the EU AI act has aimed to address, putting bans on social scoring systems, the use of emotion detection systems in workplace or educational settings, and predictive policing systems. Banned systems are not safe for use not only because the system can be biased, but also because the context of the use case itself is so high stakes. 

AI as an Agent

“AI agents will become the primary way we interact with computers in the future. They will be able to understand our needs and preferences, and proactively help us with tasks and decision-making.” — Satya Nadella

“An agent is generally defined as a person, entity, or automated system authorized or designed to act on behalf of another to achieve a specific goal.” — Merriam-Webster 

AI as an agent metaphors take on many forms, from full blow personhood (which we discussed in last month’s column), to more limited types of agency that we might ascribe to animals, to even more narrow types of machines with specific goals, like a thermostat regulating the temperature of a room. The AI agent that Satya Nadella describes is less like the thermostat and more like a helpful personal assistant. Yet, when it comes to control, it tends to be inversely related to agency. Greater agency = less control.

Recently in my city, there was a tragic case involving two very large dogs that mauled and killed an 11-year-old child. The dogs belonged to the roommate of the child’s father, and the child had been staying at the home with their father during a visit. The court determined that the owner of the dogs was negligent. The media story states that three elements were necessary to prove beyond reasonable doubt:

“The accused failed to take the reasonable steps to ensure Kache [the boy] would not be harmed, that her [the accused] conduct showed a wanton or reckless disregard for Kache’s safety, and that her conduct resulted in the death.”  — CBC

The dogs exerted their agency outside of the owner’s direct knowledge, but the owner was deemed negligent per the above. This might be the governing principle for agentic AI. The AI agent operates at the behest of the person who deploys it, irrespective of the actions it takes. If that person has not done enough to ensure they were not acting recklessly and harm is caused, we have negligence. We are back to personal responsibility in this case. 

However, we might also ask what is reasonable to expect from the company that is selling agentic AI? This is a product liability question. Is the AI agent defective or flawed in some way and therefore performed in an unexpected manner that caused harm? Was there a failure to warn of its risks? In terms of who will be held responsible, the Air Canada chatbot case has already set a precedent (at least in Canada), that it is the company that deployed the AI system that is held responsible for it. It’s less clear that the manufacturer is legally liable. Governance in this framing tends to fall into issues of duties, responsibilities, and corresponding tort laws. 

No Perfect Fit

None of these metaphors are a perfect fit for AI. On the one hand, they all encompass aspects of how we might approach AI governance. Yet, they also all seem to fall short in addressing the socio-technical nature of AI and how algorithmic systems are entangled in our everyday lives. Circling back to the expansive definition of governance from the Commission on Global Governance perhaps we might find paths forward with new metaphors that support notions on continuing processes, cooperative actions and informal arrangements, even as we pursue various aspects of these metaphorical AI framings to guide our current regulatory efforts.

*its adaptive design that makes this possible

Send Me Your Questions!

I would love to hear about your data dilemmas or AI ethics questions and quandaries. You can send me a note at [email protected] or connect with me on LinkedIn. I will keep all inquiries confidential and remove any potentially sensitive information – so please feel free to keep things high level and anonymous as well. 

This column is not legal advice. The information provided is strictly for educational purposes. AI and data regulation is an evolving area and anyone with specific questions should seek advice from a legal professional.