AI Safety

The risks associated with AI have been a prime motivator for launching the AI governance programme.  Immediate risks such as bias, surveillance, deep fakes and cybersecurity posing threats to human rights, democracy and the rule of law. But there are also potentially catastrophic and existential risks arising from AI.  Many experts consider that Advanced AI could pose an existential risk, whether through the loss of control of the AI, or accidents or malevolence (whether civilian or military) that are facilitated byAI.  

In 2023, the UK announced that it would host an AI Safety Summit. The One World Trust is in full support for this initiative and is seeking to promote and engage with this on-going AI Safety Summit process.

In the summer of 2023, there was concern that statements by the minister responsible indicated that the UK process was going to be limited to Governments and tech industry representatives, prompting a blog by the One World Trust: AI SAFETY AND THE UN SECURITY COUNCIL. It was good to see a counterbalancing presence of Academia and Civil Society however at the November 1^st /2^nd summit.  In the run up to the Summit, several authors proposed solutions including Suleyman and Bremmer with their paper Beyond the Paradox.  Their proposals provided a vehicle for the One World Trust to express its views on the way forward: BEYOND THE PARADOX: OUR TAKE ON AI GOVERNANCE.

The Summit proved to be a great success, as summarised in the One World Trust webinar: Bletchley Unpacked: Successes, Failures& the Path Forward, bringing the combined assessments of

• ‍Dr Robert Trager, International Governance Lead at Centre of Governance of AI, Co-Director of the Oxford Martin AI Governance Initiative
• Dr Florian Ostmann, Head of AI Governance and Regulatory Innovation, at the Alan Turing Institute‍
• Lee Sharkey, Co-founder & Chief Strategy Officer, Apollo Research

‍

Key outcomes of the UK AI Safety Summit

1.   Agreement that the US, UK and other agreed AI safety Institutes can assess the new set of Large LanguageModels BEFORE they are launched.  This is a significant achievement - BUT

a.   Already however the tech companies have ben flexible their muscles, pointing out that the agreement is voluntary and demanding a time table.

b.   The question is what will happen if the evals lead to a requirement for major change.

2.   The launch of AI equivalent of the IPCC – with Yoshua Bengio leading for the first report.

a.   Now called the InternationalScientific Report on Advanced AI Safety.

b.   Initial findings to be published ahead of the Korean virtual summit, with a major report to be reviewed at theNovember summit in France.

c.    Its drafting will be underpinned by comprehensiveness, objectivity, transparency, and scientific assessment – a framework which will ensure a thorough, robust, and balanced assessment of the risks of AI. 

3.   The Bletchley Park Declaration, with everyone on the same page, committing to work together on AI Safety.

4.   Starting a process not an event, with

a.   A South Korean event (virtual) inMay 2024

b.   In person event in France in approximately a year after Bletchley Park.

These are all firsts, and represent a major step forward towards addressing the AI related catastrophic and existential risks

However, as the Bletchley Unpacked panellists observed, the Summit will not have been a success if it turns out to be one event only. But if it transpires to be the beginning of a process, then it will be seen as a significant milestone as there is still a lot of risk around.

‍