Responsible AI?

What the food movement already knows about responsible AI. Reflections on a provocation for the Food Ethics Council, who are building their AI policy in the open.

Title slide: Responsible AI for organisations working in food systems

I was asked by the Food Ethics Council to join a session they were running, bringing 30 odd organisations together as they discuss developing an AI policy. The intro I gave was only Ten to fifteen minutes on responsible AI, to a virtual room of people who spend their days thinking about food systems. The FEC are developing their AI policy in the open, inviting people in while it's still half-formed, which I love. It's how this stuff should be done, and it's rarer than it should be.

This is what I talked about and some reflections since.

Not consumers at the end of the chain

I started not with AI, but with Food Citizenship.

Food citizenship is the idea that people aren't just consumers at the end of a supply chain, passively receiving whatever the system produces but should be active participants in the whole thing. They can shape it, question it, grow some of it themselves.

Now imagine pointing those ideas at AI.

Most organisations are being positioned, very deliberately, as consumers at the end of an AI supply chain. Here's the product. Here's the (ever increasingly costly) subscription. Take what you're given, renew annually. The entire commercial framing wants you passive.

But I'd argue you don't have to only receive AI as a finished product. You can in some ways, shape it, question it, help build it. You can choose smaller models , open models , tools you can actually see inside. You can be a citizen of this system rather than a consumer of it. The instinct is the same one the food movement has been teaching for years.

The speed of trust

I did talk about possibilities of course, because there are possibilities with AI. I talked about Caddy, built by Citizens Advice, my own Open Recommendations , examples from Wildlife Trust and more.

But I also talked about moving with care.

The tech industry's default speed is move fast and break things. Fine, maybe, if what you're breaking is your own product. But organisations working in food systems (or any other systems that involve people and planet) hold relationships with communities who have every reason to be wary already. When we break things, it's trust that breaks, and the people most affected who feel it. So the question isn't just how fast can we adopt. It's how fast should we, if at all?

I dropped in some, to my mind, excellent quotes. Rachels "FOMO is not a strategy " and Richard Popes "efficiency is a trap "

Doing the wrong thing faster is still the wrong thing. An awful lot of AI adoption right now is organisations getting measurably quicker at things they probably shouldn't be doing at all.

The other thing I warned about is the shadows. People are using AI tools unofficially, off the radar, pasting things into free chatbots because the official answer was no, or worse, silence. A policy that pretends this isn't happening isn't a policy.

Who guides the thinking

AI is sold to us as opaque, and along with the opacity comes a message: this is expert territory. I've heard several people say you can't really have a view on AI unless you really understand it. But what does that mean? Should people need to understand transformers, weights, training runs and the benchmark suites?

I disagree. I think you should be applying your own expertise to AI and developing your principles from that.

The room I was speaking to has spent decades thinking hard about provenance, about power in supply chains, about who benefits and who carries the cost, about what honest labelling looks like. That is the relevant expertise. You don't need to understand a transformer to ask whether a tool respects the people you serve, any more than you need to be a food chemist to ask what's in the tin. Apply your expertise, whatever it is, to AI. Work out what matters to you and hold AI up to it. Not the other way round.

And it matters who's offering to do the guiding. As I was writing this, Anthropic announced Claude Corps: $150 million to embed a thousand fellows, trained in the use of Claude, into US nonprofits for a year. There's nothing on that scale in the UK, but the shape is familiar here too, Microsoft and others fund AI programmes for nonprofits, training the sector in how to think about the thing they sell.

I'm not saying no good comes of it. It's hard to turn down free help, and plenty of organisations will take it gladly and get value from it. But I think we need to pay more notice to the arrangements. It feels a little like Nestlé running programmes on child nutrition.

The food movement learned this lesson the long, hard way: the people with the most to gain from your choices are not the people to outsource your thinking to. Which is exactly why the principles need to be yours.

Principles outlive tools

And they need to be principles, not products. I shared an example in the session and it's not to pick on LOTI, but it illustrated my point. In 2023 LOTI published a one-pager of generative AI guidance for council officers. It named the tools of the day: Bard, "Windows 365 Copilot soon", Dall-E, Stable Diffusion.

Two years on, most of that list is renamed, superseded or gone. But underneath all that were their six rules: never upload residents' private data, reference your AI use, don't let it make decisions, stay accountable for what it produces. And those are still solid even though the tools have changed.

The lesson for anyone writing a policy: anchor it to principles, not products. A list of tools is out of date the moment you publish it. A list of principles is not.

And the principles don't need to be complicated.

Hold it up to the light

Whatever your approach, have some way of viewing AI that isn't just about capability. For the discussion itself I offered two sets of lenses, both things I've written about before. The seven bearings behind Bearing: quality, capability, speed, cost, sustainability, transparency, privacy, as a ready-made set of questions to hold any AI choice up to. And the TechFreedom lenses, which ask a different question: not whether a tool is good, but what depending on it costs you.

Seven bearings — quality, capability, speed, cost, sustainability, transparency, privacy

Read the label

I ended by bringing it back to Food.

The food movement taught us all to read labels. To ask what's actually in the thing. Where it came from. Who made it and under what conditions. What it costs that the price doesn't show. Whether the marketing on the front matches the ingredients on the back. It took decades, and it changed what the industry could get away with.

So: if your AI policy were a food label, would you buy it?

Closing slide: If your policy were a food label, would you buy it?

Would the ingredients list be honest? Would the provenance stand up? Is there anything ultra-processed hiding behind a wholesome front? Would the people you serve, reading it, trust what's inside?

The Food Ethics Council are exploring some of this in their own way, working it out in the open, with their community. If more AI policies were developed like that, more of them might pass the label test.

The full set of resources from the session - policies to borrow, tools to try, real examples from nature and conservation is here. Take what's useful.