Insights form the London Data Week – Whose AI is it anyway? Data, Sovereignty and the Geopolitics of Intelligence

Parturiunt montes, nascetur ridiculus mus? The key question the panel set-out to answer is: who is in control when it comes to AI? Reflecting on the discussion I would dare to answer this question with a resounding: ‘Nobody, in fact. Big deal?’

The public panel held at Kings College, London on 8^th July offered interesting insight over and above the actual discussion. Interest in the topic was huge. The event held at Bush House was fully booked with organizers operating a waiting list. Attendees seemed to be of diverse backgrounds across students, academic staff, practitioners, a Stanford MBA student and a lot of generally older members of the public. It was the audience that in the end drove the really interesting points.

The key question the panel set-out to answer is: who is in control when it comes to AI? The question was asked against the backdrop of AI models becoming more powerful, the capability gap between Chinese and Americal models shrinking and state actors restricting access to models at will.

In short, the panel did not offer a clear answer to the question posed which is also testament to the dynamic nature of this topic. Reflecting on the discussion however, I would dare to answer the question of ‘who is in control’ with: nobody, in fact. It also left me wondering if AIs transformative impact is not slightly exaggerated.

There were four interesting reflection points that I drew out of the discussion:

• It is not clear where and how exactly AI should be regulated. The panel discussed at great length the comparison to nuclear non-proliferation treaties. However, what exactly constitutes AI – Infrastructure, Cyber, Models, Data – is not yet clearly and sharply defined raising the question on what is actually to be regulated.  There is a recognition that there is a conceptual difference between preventing actors from ‘building more of’ (as in the case of nuclear non-proliferation) versus restricting certain security relevant use cases as would be the case with AI regulation.

  
  

• Cyber security is a key concern - But what is new here? The potency of the latest models in not only generating code, but also in identifying and exploiting potential loopholes in existing (critical) software is at the forefront of the debate and has informed decisions of actors – such as the US Government or AI companies like Anthropic – to withhold or restrict access to such models. The question however is: what exactly is new here? The identification of security vulnerabilities in code has always been the bread and butter of software/cyber security companies, intelligence services and other malicious actors alike. AI may bring greater pace and analytical capability but this is largely available to the whole spectrum of actors: good, bad and ugly. Yes, restricting access to specific models may tilt the balance in the short run. Over time, however I would expect equilibrium.

  
  

• Building ‘capabilities’ at the receiving end of AI are more crucial than formulating restrictions – This to me was a critical point made in the panel discussion. What exactly is meant by ‘capabilities’ however was not fully defined. Take two dimensions: end users and interfacing applications. What are the avenues end-users can take to build capabilities to work with AI effectively, to formulate good inputs and to critically evaluate output? At a basic level, do people even take the time to fully read all the blurb that AI generates before passing it on? Secondly, at the level of application, any credible online banking app has in-build security features to prevent fraudsters logging into your account. Are there any technical capabilities that need to be put in place at the level of APIs to protect against unwanted agentic access?

  
  

• Does AI live up to its promise in practice? There appears to be a gulf between theoretical enthusiasts and sobered end-users of AI. This became evident by a member of the audience summarizing their actual experience in using AI: verbose answers, hallucinations, incomplete and faulty meeting minutes; and the corresponding ‘cognitive overhead’ they had to invest to manage the workload AI was generating with the human in the loop. ‘Is this really a significant change or will we remember this in twenty year’s time as ‘the summer of AI’?

On balance, whilst a lot is not yet clearly defined, it seems that bottom line AI offers nothing groundbreakingly new. It is just faster than previous ways of working. It may be more fruitful to think about what is needed to use AIs features and capabilities well as opposed to thinking about restrictions. Over time, AIs output will likely get better. In the short run and whilst the mountains are still labouring, conceptualizing and investing into long term strategic human and technical capabilities will be critical.