Dr Tibero Caetano Shares What Safe AI Looks Like

Dr Tiberio Caetano – co-founder and chief scientist of Gradient Institute, an independent not-for-profit working towards safe AI – offers some points to consider.

Traditional AI is designed to accomplish a particular task – to play chess, to detect cancer – while generative AI is designed to have cognitive capabilities, to be smart. Because it’s a general-purpose technology, the risk profile – what can go wrong – is highly uncertain. What can go right is also uncertain. There are hallucinations, yes, where generative AI makes things up, but think of the opposite: circumstances where telling the truth is a problem. If not appropriately designed, this is an immature technology where 100 million people could type in: “How do I synthesise a pandemic-grade virus?”

The first step for an organisation is education, from the board to the data scientists, about how to create guardrails, how to think about trade-offs and the upsides and downsides, because often people put this technology into a “bad” or “good” box. It’s a massive opportunity to create a whole suite of new products and reach customers you historically couldn’t, to personalise communication for different sectors and demographics. There are also new exposures: being sued for copyright infringement or for the product doing something it wasn’t designed to.

You need technical experts to participate in the process of risk management, because much of the risk comes from nuances of how the technology is built or refined. But this technology can impact people in many different ways so you need voices from a range of disciplines: legal, ethics, the community and sometimes marginalised groups for whom these tools may be more dangerous.

For every new technology you introduce, you’re also introducing new vulnerabilities in your customer base. If you’re talking about large language models, some people may be more easily manipulated. Marginalised groups’ susceptibility to AI is not new and the risk of bias persists with this new technology. But think about this big land where many communities are isolated geographically. This technology can be a great leveller, automating education or healthcare services for those with no service at all.