June 13, 2025, by Brigitte Nerlich

Public engagement with AI: Some obstacles and paradoxes

I recently listened to a webinar by social scientists who had studied what AI researchers say about public reception of AI. The most important words I heard were ‘evidence’ (about public attitudes to and inclusion in AI) and ‘voices’ (of communities underrepresented in or negatively impacted by AI). The main argument was, I think, that one should listen to more diverse voices outside the AI bubble and make AI people inside it aware of such voices. That sounded good.

But then I thought: what about voices inside the AI bubble? Should one listen to them as well, at the same time as inviting them to listen to other voices? Should there be a two-way effort at speaking, listening and understanding each other? But what would that entail?

While I was listening to the webinar, I was also scrolling on Bluesky, where I follow a few AI people and people studying AI, and I must confess that I mostly don’t get what they are talking about, as many speak a language I don’t understand. It turns out that they are quite deeply into vector maths and other sorts of maths, something that is completely alien to me. 

AI, alignment and vector maths

You have all heard that AI has an ‘alignment’ problem. A while back I even wrote a blog post about this. But over time I have come to realise that when AI researchers talk about alignment, they mean something rather technical and that they use a lot of maths when grappling with the issues.

When alignment researchers describe their actual daily work, they might talk, for example, about using ‘vector projections to decompose model activations’ or ‘computing gradients of the loss with respect to model parameters’ or ‘projecting activations onto interpretable directions’. To most people, including me, this might as well be ancient Greek. Even the basic building blocks of that type of talk – what a gradient is, what high-dimensional spaces mean, why you would want to project vectors – require significant mathematical background to grasp intuitively.

As Claude tells me, this goes beyond just the math. AI researchers develop this whole conceptual vocabulary around things like ’inner alignment’ vs ‘outer alignment’, or debates about whether models are ‘mesa-optimising’ or just doing ‘deceptive alignment’. These are not just jargon terms though – they point to genuinely important distinctions about how AI systems might behave – and what wider implications that may have. But to an outsider, it all sounds like esoteric philosophy. 

Lay people like me might hear ‘AI alignment’ and think it’s about making sure AI systems are good or safe in some general sense. But researchers are often thinking about very specific technical problems like ‘reward hacking’ or ‘distributional shift’. Neither side is wrong, but they are operating in completely different conceptual frameworks. There is quite a deep language issue here that needs to be acknowledged and that might not be easily overcome by just asking for better communication or more diverse voices.

Evidence and voices

I started this post by musing about a webinar that focused on evidence and voices. I hope that my meandering thoughts (inspired in places by my chats with Claude) have provided some evidence that we need to listen not only to more voices from outside AI but also from inside AI if we want to create a context for mutual understanding and possibly better AI policy. Why is that?

The mathematical abstractions, the vector spaces, the optimisation landscapes that AI researchers work with daily are often where the most important insights about AI behavior emerge. The real action is happening in these abstract spaces that are genuinely difficult to visualise or understand for lay people like me. 

There is therefore this paradox where the very tools that make AI research possible are also what make it nearly impossible to communicate beyond the actual experts. It’s not that lay people just don’t understand technical details – but they just can’t engage with AI issues as AI researchers experience them every day. They can’t really walk in their shoes.

And there is more: On the one hand we have AI researchers trying to solve highly complex alignment problems and on the other we have social scientists who try to solve highly complex policy and public understanding problems. What does it mean for them not to share a common language? How can one meaningfully listen to and respond to voices from inside and outside AI, when people don’t share the same basic conceptual toolkit? 

This might be one reason why so much public AI discourse ends up being either superficial (talking about AI safety in general terms) or focused on immediate, tangible applications (like chatbots or autonomous vehicles) rather than the deeper technical challenges that researchers grapple with day in day out.

One can argue that this is like other technical fields, but AI feels different because the stakes are so high and the timeline is so compressed. I wonder who the AI science communicators are out there who can speak across languages, listen to diverse AI voices, understand them and engage wider audiences with them. I’d love to know.

PS: I had just finished writing this post when Kai Kupferschmidt, a brilliant science communicator posted this intriguing article entitled “How laypeople evaluate scientific explanations containing jargon”. I wonder how that would work with AI jargon. (Here is a thread by the authors)

Image: An artist’s illustration of artificial intelligence (AI). This image depicts AI safety research to prevent its misuse and encourage beneficial uses. It was created by Khyati Trehan (Google DeepMind, Pexels)

• Share on Facebook
• Share on Twitter
• Share on Google+
• Share on LinkedIn