Do No Harm, Build AI Safely

Forbes

21-05-2025

Yellow warning sign symbol or alert safety danger caution illustration icon security message and ... More exclamation triangle information icon on attention traffic background with secure alarm. 3D render.
When it comes to being safe with AI, a lot of people would tell you: 'your guess is as good as mine.'
However, there are experts working on this behind the scenes. There's the general idea that we have to adopt the slogan 'do no harm' when it comes to employing these very powerful technology models.
I wanted to present some ideas that came out of a recent panel discussion at Imagination in Action where we talked about what's really at stake, and how to protect people in tomorrow's world.
In a general sense, panelists talked about how the context for AI is 'political.' Or, I should say, in the Greek sense, where as historians point out, 'the polis was the cornerstone of ancient Greek civilization, serving as the primary political, social, and economic unit.'
In other words, how we use AI has to do with people's politics, and with political outcomes. The ways that we use AI are informed by our worldviews, and geopolitical sentiment as well.
'When politics are going well, it's invisible, because business rolls on, art, culture, everything rolls on and you're not really paying attention to politics,' said panelist Jamie Metzl, author of Superconvergence. 'But I'm the son of a refugee. I've lived in Cambodia. I spent a lot of time in Afghanistan. When politics goes bad, politics is the only story. So everything that we're talking about, about technology, AI, exists within the context of politics, and politics needs to go well to create a space for everything else, and that's largely on a national level.'
In terms of business, too, we have to look at how information is siloed for different use cases. One of the objectives of this kind of thing is global governance – AI governance that sees the big picture, and applies its principles universally.
A lot of people, in talking about their AI fears, reference the Skynet technology from the Terminator films, where there's this vague doom attached to future systems that may rule when the robots will be in charge.
But some suggest it's not as blatant as all that: that the overwhelming force of AI can be more subtle, and that it's more how AI is already directing our social outcomes.
'It's the algorithms that already today are denying people access to housing, access to jobs, access to credit, that are putting them at risk of being falsely arrested because of how a biased algorithm misinterpreted who they were, and how our legal system compounded that technical error with legal injustice and systemic bias,' said panelist Albert Cahn.
Cahn pointed, as an example, to a system called Midas that was supposed to seek out fraud in insurance systems. Instead, he noted, the system went too broad, and started catching innocent people in its dragnet, submitting them to all kinds of hardship.
'When we are talking about the scales of getting it wrong with AI safety, this isn't about missing a box in some compliance checklist,' he said. 'This is truly a matter of people's livelihoods, people's liberty, and in some cases, sadly, even their lives.'
That's something that we have to look out for in terms of AI safety.
Noelle Russell had a different metaphor for AI safety, based on her work on Alexa and elsewhere in the industry, where she saw small models with the capacity to scale, and thought about the eventual outcomes.
'I came to call these little models 'baby tigers,'' she said. 'Because everyone, when you get a new model, you're like, 'oh my gosh, it's so cute and fluffy, and I love it, and (in the context of model work) I can't wait to be on that team, and it's going to be so fun'. But no one is asking, 'Hey, look at those paws. How big are you going to be? Or razor-sharp teeth at birth. What are you going to eat? How much are you going to eat? Where are you going to live, and what happens when I don't want you anymore?' 23andme, we are selling DNA on the open market … You know, my biggest concern is that we don't realize that in the sea of baby tigers and excited enthusiasm we have about technology, that it might not grow up one day and … hurt ourselves, hurt our children, but most importantly, that we actually have the ability to change that.'
Panelists also talked about measuring cyber security, and how that works.
'In carpentry, the maxim is 'measure twice, cut once',' said panelist Cam Kerry. 'When it comes to AI, it has to be 'measure, measure, measure and measure again'. It's got to be a continuous process, from the building of the system to the deployment of the system, so that you are looking at the outcomes, (and) you avoid the (bias) problems. There's good work going on. I think NIST, the National Institute of Standards and Technology, one of my former agencies at the Commerce Department, does terrific work on developing systems of measurement, and is doing that with AI, with the AI Safety Institute. That needs to scale up.'
Going back to the geopolitical situation, panelists referenced competition between the U.S. and China, where these two giants are trying really hard to dominate when it comes to new technology. Russell referenced a group called 'I love AI' that's helping to usher in the era of change, and provides a kind of wide-ranging focus group for AI.
'What I've uncovered is that there are anywhere from 12 years old to 85 year old (people,) farmers to metaphysicians, and they are all desperate to understand: 'What do you mean the world is changing, and how do I just keep my head above water?'' she said.
Then too, Russell mentioned, toward the end, the imperative for AI safety and how to get there.
it's not a checklist you sign off on. It's not like you said, it's not that framework that you adopt, it's like the way you end up thinking the way you are the way you the way you build software, the way you build companies, it will need to be responsible.
These are some of the thoughts that I thought were important in documenting progress toward AI safety in our times.