Tim spoke on a panel at the Green Party of England and Wales party conference alongside Rachel Coldicutt (Careful Industries) and Kierra Box (Friends of the Earth), and chaired by Rachel Millward (Deputy Leader, Greens) on “Who Owns the Machine? AI, Power, and the Green Alternative “.
Tim’s write up of the Panel We covered a lot of ground in the discussion - and there was much wisdom shared by my fellow panelists, but this post riffs on a few of the inputs I offered, as well as drawing on discussions we held earlier in the day in a participatory workshop on data centre and AI policy.
Who are we listening to? We started the panel with a question on recent headlines about how AI might kill us all within years. A show of hands in the room revealed many people are starting right now from a position of concern and fear. Yet, as Rumman Chowdhury highlights it in this piece for rolling stone we need to look behind the headline pronouncement and ask: how? Who is making these claims? How do they arrive there and on what evidence? And (how) do they respond to counter-arguments?
One question from the floor towards the end of the panel asked how we can take a more ‘matriarchal’ as opposed to ‘patriarchal’ approach to the future of AI. It’s a very good question, and I shared a list of some of the people whose writing and commentary I try to listen to, to put tech-industry pronouncements into perspective. People like:
- The aforementioned Rumman Chowdhury
- Emily Bender and Alex Hanna whose ‘Mystery AI Hype Theater 3000’ deconstructs recent papers and pronouncements
- Timnit Gebru
- Joy Boulumwini
- Ruha Benjamin
- Maya Indira Ganesh
- Karen Hao
- Olivia Guest
- Şerife (Sherry) Wong
- Alix Dunn and The Maybe and
- The (m)otherboard godmothers
Most of these are posting regularly on at least one social media channel, and as a corrective or challenge to mainstream dialogue, I can recommend adding at least a few to your media diet: or checking out their past books, talks, podcasts or writings.
I also find platforms like IT for Change’s Bot Populi and the broad coverage on Tech Policy Press useful to escape filter-bubbles of AI hype and AI doom. And We and AI’s work, led by Tania Duarte and an amazing team of volunteers and contributors provides some great mental models for helping to look behind the headlines.
Find agency in your community Of the open data movement, Tim Berners-Lee once said:
“It has to start at the top, it has to start in the middle, and it has to start at the bottom.”
I think we can apply that same sentiment to how we respond to AI. But, when it takes collective action and political will to start in the middle, and action from the top feels constrained by geopolitical forces, one of the best places for us to claim human agency around AI is at the bottom: through grassroots dialogue, engagement and action.
This is one of the key principles behind the Citizens’ Track on AI I’m currently working on - creating resources to help communities have informed conversations about AI, and to explore in parallel messages to hand up to policy makers - and actions that can be taken at an individual and collective local level.
Those actions can be simple: choosing a search engine like Duck Duck Go which doesn’t default to generating AI responses on every query; or committing to always add an AI Attribution to work where you have used AI. They can use existing collective rights and power: workers asking to dialogue over introduction of AI tools into the workplace, or parents asking schools to explain the benefits of AI Ed Tech tools before they are just turned on. Or they can use civic position: councillors putting questions to committees about procurement of AI tools, or citizens using the planning process to engage with data centre construction.
Fundamentally, although some of these approaches might use strategies found on the AI Resist List they are not about rejection of AI - but about taking an informed and evidence based approach to choosing when we want to use it, what for, and how we handle the distributional effects of that choice. For example, AI might make certain work more efficient, but have knock on effects for the workers whose role it changes - and that calls for conversation and adaptation - not imposition and alienation.
Manage demand: frugal AI AI is not new. The lineage of AI development traces back to the 1950s and even before. In the panel it was useful to draw a distinction between two kinds of computation:
- Deterministic code - where a computer follows defined rules step-by-step to come to an outcome.
- Probabilistic networks * - where a model is trained based on vast amounts of data to ‘predict’ an acceptable response for any given input.
The first is potentially very efficient. Sorting a list involves just comparing a defined feature of each item, and checking in a few computation operations which is bigger, which is smaller. However, if deterministic code was not written to do something… it can’t do it. And if you provide the wrong inputs, you get ‘Computer Says No’. For someone without programming skills then, the universe of computable things is small, and control over the everyday tasks a computer can complete rests with someone else.
By contrast, probabilistic networks (e.g. Large Language Models / Generative AI) are in many ways highly inefficient for any specific task. To sort a simple list of items, 10+ gigabytes of model weights might be kept in memory, and a ‘reasoning’ model might go through numerous steps to work out what the task is, how to approach it, and might iterate over all sorts of extraneous bits of data. However, as a ‘General Purpose Technology’, the Generative AI tool can provide a reasonable approximation of many of the tasks a user might throw at it. Suddenly, all sorts of tasks that would have had an up-front cost in time and effort to program a machine to complete become more easily computable.
This creates massive demand for computation (inference in AI terminology) with the consequent demand for energy, water, land, critical minerals and capital to build out more and more data centres. At the same time, speculative investment in data centres can create conditions for induced demand. And efficiency at the level of AI inference and code alone faces the Jevons Paradox.
Fights against supply side data centre construction, without addressing the demand side for computation risk simply offshoring the emissions and impacts from data centres to countries without the political freedoms for public campaigns on data centre construction.
Instead, a Frugal AI strategy can think about how we use the capability to generative AI to build better deterministic systems (which also have the benefit of clearer audit-ability when used in public sector contexts), how we build out capacity to make choices about where AI use brings enough value to offset the externalities it causes, and where we can build more resilient supply chains in our economy.
This too is not an anti-AI position, but rather one that invites the UK to develop it’s strength in careful and critical adoption of AI in ways that keep costs (economic and environmental) down, whilst maximising long-term gains.
PURRR Whilst I didn’t use it in the panel, in a few conversations afterwards I reached for my work-in-progress Mnemonic for thinking about AI: PURRR
- Perspective. We need to keep AI issues in perspective; when AI systems meet real-world contexts, there’s a lot more going on than just the technology. Arvind Narayanan and Sayash Kapoor’s work on AI as Normal Technology can be useful here, as can keeping in focus the evidenced climate crisis we face, and the global development goals we’ve internationally agreed to work towards.
- Use. There are many places where AI tools are genuinely useful, and experience of their use can also be really helpful to developing an individual or collective stance on what we do, and don’t, want AI to do in our communities, country and world.
- Refuse. We need to keep the rights to refuse use of AI. It should be ok to say ‘No’ if someone says ‘Can I turn my AI note-taker on in this meeting’, or to question introduction of AI into schools if there isn’t evidence of positive educational impacts.
- Respond. Whether or not we choose to use AI - it is being used in all sorts of ways around us. And that requires policy and practice that takes AI use into account. That might involve thinking about how we need to secure systems when adversarial uses of AI could bypass current protections. Or it might involve thinking about how public service processes can promote equitable access when there is a risk that those who can afford advanced AI agents might be able to better bypass bureaucratic barriers and gain more from the system.
- Reshape. Technology is shaped by human choices. As Gary Marcus argues in his 2024 book Taming Silicon Valley the version of AI offered to us right now is the product of particular investment choices, and is not inevitable. Actions at the bottom, middle and top can all reshape the research into, markets for, and adoption of AI.
A green vision on AI As Deputy Leader of the Green Party, Rachel Millward closed the panel pointing to the party’s five point plan on AI
- A strict regulatory system for AI, backed by a coordinating body with specialist expertise, alongside international cooperation to establish a worldwide body to oversee and regulate AI development.
- Legal bans on lethal autonomous weapons, AI deception used to manipulate behaviour or undermine informed decision-making or democratic processes, routine biometric identification in public places, and emotion-recognition systems in workplaces and education.
- A cross-sector plan for AI infrastructure that reduces dependence on overseas tech companies, prioritises renewable energy and protects water supplies, with strict environmental assessment required for new data-centre approvals.
- Public research funding and procurement directed towards AI that serves the public good. Where professionals use AI in their work, humans must retain control over all key decisions.
- Formal processes for workers, civil society and affected communities to participate in decisions about whether and how AI is introduced and where data centres are built, alongside protection for workers’ rights and support for people whose livelihoods are disrupted.
Point five in particular reflects input from Connected by Data into earlier AI policy development on the importance of democratic control.
To my knowledge, this was the first time AI has taken centre-stage at the conference of Green Party in England and Wales, so I’m looking forward to seeing where the conversation goes next.
Disclosures/statement of position:
- I took part in the panel both as Director of Research and Practice at Connected by Data, and as co-chair of the Science and Technology Policy Working Group of the Green Party of England and Wales.
- This post was entirely written by hand, without use of AI tools.
Image credit: Sue Griffiths