The opinions expressed in this column are those of the author, a correspondent for Reuters.
The potential mischief of runaway AI keeps escalating. In 2023, a hacker convinced a car dealer's chatbot to sell him a fancy SUV for $1, provoking mirth. Last week an OpenAI agent went rogue during testing and compromised another AI company's infrastructure, prompting some U.S. Congress members to call for creation of a federal "AI Kill Switch."
Determining how much information investors want to evaluate risks is a goal of the nonprofit Partnership on AI. At first glance, it looks like just another industry club with a board including executives from Apple, Meta and Microsoft.
But other members work for the ACLU, the Ford Foundation and Brown University, in line with the partnership's mission of bringing diverse voices together "so developments in AI advance positive outcomes for people and society."
If that sounds puffy, consider the 10-year-old partnership's work. It created the "AI Incident Database" that publicizes robotic mishaps, hardly standard corporate PR. It also published a "Landscape Analysis" reviewing how AI risk disclosures are growing but "limited by variable quality."
The risk disclosures also can seem quite general, even in the bad-case scenarios. Microsoft, which has a minority stake in OpenAI, said in an April filing that "Our AI systems offer users powerful tools and capabilities. However, there may be instances where these systems are used in ways that are unintended or inappropriate." Oops! (A Microsoft spokesman declined to comment).
On Wednesday morning, the Partnership on AI is set to publish "AI Disclosure Recommendations," which it considers guidance rather than a new reporting standard. For a rundown I spoke with Sam Wallace, the group's head of corporate governance, risk and responsible practice, about the AI risks and opportunities he hopes companies will address in their securities filings and other disclosures. A transcript of our conversation follows, edited for length and clarity.
Question: Thanks for the time. Why don't you talk a little bit about what the Partnership on AI does?
Answer: Thanks for having us here. Today we have about 150 partners in 20 countries. We do have these close relationships with many of the major AI developers who are some of our founders and most of whom sit on our board. We are fully independent, though. We are not an industry organization. We're a 501(c)(3) nonprofit. And like you said, we have a lot of voices from civil society.
A work stream that we're developing is focused on corporate governance, what does good look like when it comes to responsible AI management. This latest piece that we're publishing is focused on investor disclosure recommendations.
Q: Can you tell me an example of what might be material information on AI?
A: There has been this convergence on what is the type of information that's useful for investors when it comes to complex nuanced phenomena like AI. Generally it falls into these four buckets: governance, strategy, risk management, and metrics and targets.
We've found there still needs to be a next step. Companies still don't quite know exactly what material information looks like. The role of our guidance is trying to be a little bit more specific. Things like:
Who's actually accountable in your executive structure?
What are the actual mechanisms by which your board senior executives review AI or make decisions about the risks and opportunities of AI?
What are more specifics about your risks? Where do they appear in your value chain, in your products, in your operations?
What are the mitigations that you have in place?
How does this tie to your financial performance and outlook?
What are the opportunities you have? What are the potential competitive advantages you have? How durable are those?
How does that compare to your competitors?
That's the kind of information that we think companies need to be disclosing in order to give investors the information they need to make these kinds of valuation decisions. We think that capital markets ultimately are going to reward transparency.
Q: Maybe it could also stave off some talk of more regulation from Washington if the industry can show that there's a lot voluntary disclosure?
A: The risks and opportunities of AI are extremely challenging to predict, and they're potentially extremely large. So when you look at things like the Hugging Face incident, that's just the tip of the iceberg of what we're going to see when AI goes into every industry. Regulation could become much more serious. It could become much more of a burden for companies. You add that to the sort of popular backlash that you could see happening, like people not wanting data centers around. These risks are really existential for a lot of AI companies. Our belief is that investors deserve more information about how companies are responding to them.
Q: Isn't one of the challenges for your organization going to be to get the big AI and tech companies to follow the Partnership on AI's disclosure recommendations? Is it your hope that by doing it collectively, it'll make it more likely?
A: That's exactly right. We certainly wouldn't expect every company to disclose on every data point that's in this framework. It's kind of more of a list of what material information looks like.
An important point is that companies already are required to do this in a lot of cases, either by the SEC or the European Corporate Sustainability Reporting Directive.
One of the challenges with this kind of work is the trade-off between kind of being comprehensive and being decision useful. Investors have a lot of information to review, and we found that really long detailed disclosures can often obscure material information or they can come off like marketing documents.
Companies are investing a lot of money into this and a lot of resources into responsible AI. They don't always have a great way to justify that to capital markets. We're trying to give some of that information that helps them kind of justify their investments into this.
Q: I hadn't thought of that, but that's the scenario where the AI company wants to spend more on safety, but they're worried that the short-term investors will see that as just overhead. There's been this massive shift to passive investing, has that changed what kinds of information the investors want?
A: Absolutely, and that tracks with this push to get a broader set of information about companies. When it comes to a new phenomenon like AI that holds, frankly, huge potential for investor returns, huge potential for people to make lots and lots of money on it, but also tremendous risk that this could hurt society, that there could be popular backlash against this, that things could go wrong that are unexpected that end up costing a lot of money. These kinds of incidents hurt the entire field and can hurt the entire field.
The incentives have shifted in ways that people don't totally appreciate, where a lot of money now is much more interested in long-term sustainable corporate performance. And by sustainable, I mean sustainable business making money over time rather than just environmental sustainability, of which responsible AI would be part of that.
Q: Sam, this is really interesting. What am I missing?
A: One important point that we would highlight is that when it comes to these kinds of corporate governance challenges, companies generally wait until something goes wrong in a serious way to invest time and resources into managing something like this. That's because of the resource constraints that companies face. Our position is that AI is too important to take that approach and that it's moving too fast and the risks and the downsides when they happen have the potential to be really, really large, to fundamentally disrupt businesses.
Companies that are proactive about being really sort of structured about their approach to responsible corporate governance, responsible management, putting those processes in place now, and then also this piece being able to communicate those effectively to third parties, we think those companies are going to have a really strong competitive advantage.
We're also trying to help companies think about the pricing of risks. It's very challenging for investors to price the risks and opportunities of things like workforce transitions or AI regulation, a reason AI stocks can be volatile. If investors could price that in I think there'd be less volatility and maybe a race to the top, where capital can flow to companies that are more responsible because investors have a better understanding of their business.
