My position on AI

I spent a few hours this week reading around AI and formulating my thoughts into coherent sentences. This is what I’ve got so far. 

I’d love to know what you think. Am I right? Am I being naive, or just sceptical enough? What am I missing? 

A bit of context and a disclaimer: my remit at the BFI spans the organisation’s own public-facing digital services, back office tech and digital ways of working. Others are leading our thinking on the role of AI in creative industries, from a lead body perspective. We all talk lots though. 

TL;DR – my position on AI

I’m sceptical about the tech’s current capability, while recognising it’s evolving at speed. I’m convinced it will change all knowledge work (like email in the late 90s) and supplant many professions, with enormous potential to do good and harm. I’m equal parts optimistic and pessimistic.

For my teams & organisation: I’m open to using it for service improvement and productivity gains, and conscious we can’t ignore it like other hype tech (NFTs/Metaverse) …but FOMO is not a reason to rush into it. We should start to include it in the work we’re doing, aligned to existing goals. Any use on my watch needs to be grounded in user needs and tempered with user centred design, with harmful impacts understood and kept within tolerable levels.

Mantras

You know me, I can’t resist a sticker-sized mantra (you can take the boy out of GDS but you can’t take the GDS out of the boy). Nothing beats them for creating heuristics that turn policy into behaviour.

• No AI instead of IA – c/o Matt Edgar (plus: No AI without UCD – me)
• FOMO is not a strategy – c/o Rachel Coldicutt
• A silver spanner not a silver bullet – c/o Dave Rogers
• There will be consequences – c/o Steve Messer

On what it is

• AI is a loose concept (“the appearance of a machine doing something clever” – c/o Dave Rogers)
• AI is a whole category of computer science (which includes natural language processing, machine learning, Gen-AI, Computer Vision, robotics…)
• AI predates the internet and we all use it daily, often without knowing
• Gen-AI is a novel, general purpose technology with widespread availability and emergent properties even its creators don’t fully understand, evolving at speed
• It’s a sophisticated guessing machine that generates text, sounds or pixels based on patterns it has learned (and is frequently wrong, incomplete, biased, uncanny, and generic/unoriginal)
• It requires a radical change in our trust in and relationship to computers as they begin to behave less predictably in response to instructions (c/o Harry Metcalfe)

Major players

AI helped me make this list. Probably riddled with errors and omissions. As at Feb 2025:

• OpenAI (ChatGPT, Sora, DALL-E 2 & 3, GPT-4)
• Google (Gemini, Imagen, Notebook LM)
• Microsoft (CoPilot, Azure AI)
• Amazon (AWS AI services)
• NVIDIA (GPUs and AI platforms like NeMo)
• Meta (Llama 2, various AI research in social media and VR)
• DeepSeek (DeepSeek Chat, DeepSeek Coder)
• Anthropic (Claude)

On the hype/trend

• Claims of its existential threat, utiility and economic impact are mostly exaggerated 
• Current products are mostly flawed prototypes, released too soon – use with caution
• It’s mostly tech-led, giving users tools they didn’t ask for and there’s evidence bundling it into products is a turn-off (most users distrust it)
• Expect an AI bust after the boom: shutdowns/price increases/enshittification

On the harms/impact

• Impact will be profound on society and economies, but hard to predict
• There are already and will be more significant negative consequences: job losses, climate impact, inequality, misinformation, privacy breaches and fraud, IP theft.
• (But also positive consequences: new/more fulfilling jobs, democratisation of access/lowering of barriers, breakthroughs in medicine and solutions to societal problems) 
• Jevons paradox means there will be more jobs/work but they will be different – things only humans can do. (Baldwin’s line “AI won’t take your job, but somebody using AI will” is probably true)
• It’s likely to change all knowledge-based work (akin to email) and disrupt jobs and business models in all sectors (akin to www)
• Slop is a concern: dilution of the world’s content quality and reliability, undermining of human creativity, and a cyber arms race. The downward spiral of AI-generated slop becoming training data is far more terrifying to me than any Skynet fears
• Doomsday event/singularity is a red herring, there is no evidence AGI is even possible outside of SciFi stories

On how we/I should approach it

Graphics from Dave Rogers’s talk

• Keep eye on it, be open to it; and we can’t ignore (like we could with NFTs/metaverse hype cycles)
• Make space for experimenting to feel our way (as Catherine says)
• Given the enormous potential for AI to be used for both good and harm, go carefully and thoughtfully to be sure which one/strike a net positive balance
• Use it to augment, accelerate and assist (possibly reduce) human activities but not replace them wholesale
• Beware false certainty about savings (the Jevons paradox again – spreadsheets led to more accounting not less, etc)
• Like all tech, start with user needs & outcomes not solutions/capabilities and assess the risks, and take a multidisciplinary test and learn approach
• Needs to be built on strong foundations of: org culture, dynamic operating model, good data & digital services. Aka no innovation until everything works or more pragmatically, innovation yes sure but on proviso we also continue to fix everything else.
• Mitigate risks of cloud providers re privacy and post-bubble price hikes by running local GPTs

More specifically, where to use it

Some high level themes/categories that feel like the right places to start:

• for a helping hand in daily work, like brainstorming, summarising, reducing the inertia of a blank page, recording and transcoding (captions, translations, transcripts, minutes)
• to make work less tedious and more productive, like automating repeatable tasks
• to do things quickly that humans would do slowly if at all, like analysing huge datasets

Where we should avoid using it:

• customer service bots that prioritise business need (avoid contact) over user need (get help)
• decision-making where fairness is an imperative (i.e. all evaluations/assessments)

On the government mandate/imperative to use it

• scan/pilot/scale approach in the plan is sensible; breathless announcements less so
• risk that the push for rapid adoption overrides voices for test and learn and UCD
• can only add value on top of solid foundations (what Dai said + what Dave said)

🤔 Thoughts?