On 9 May, UK’s Michelle Donelan will answer your AI Safety Summit questions on LinkedIn

Send your questions on LinkedIn now:

https://www.linkedin.com/posts/michelledonelan_aiseoulsummit-activity-7191344603336220672-Oc-W

Michelle Donelan MP on LinkedIn: #aiseoulsummit

Do you want to know more about the upcoming #AISeoulSummit?

On 9 May I’ll be answering your questions on the Summit which the UK is co-hosting with the…

hurt-tomato•5/1/24, 9:19 AM

posted

We've seen evidence of a "race to the bottom" dynamic, where countries on an individual level do not want to slow down AI development by prioritising safety, although they would on a collective level. The only solution to this problem is international regulation. We can get that through treaties. Treaties happen on summits. Treaties require one of the people present to take the first step and initialise negotiations. The UK is well-placed to do so - your AISI has amazing people.

My question: Will you initialise treaty negotiations to prevent the creation of catastrophically dangerous AI models? If not, why not?

moderate-tomatoOP•5/1/24, 9:38 AM

My question to Michelle Donelan:

Thanks for answering questions from the public. This is great to see.

One of the outcomes of the first AI Safety Summit was voluntary commitments from all the big AI labs. They promised to submit their products for testing before they are released.

A Politico article called "Rishi Sunak promised to make AI safe. Big Tech’s not playing ball." says that only one AI lab (Google DeepMind) has allowed "anything approaching pre-deployment access". The other AI labs have not honoured their voluntary commitments. Could you confirm whether this is true?

If this is true, does this demonstrate that we should replace voluntary commitments with mandatory standards? In other words, nations should work on drafting and passing domestic AI safety laws and international AI safety treaties. Will this be the focus of the upcoming #AISeoulSummit?

Here's a link to the POLITICO Europe article by Vincent Manancourt, Gian M. Volpicelli and Mohar Chatterjee:
https://www.politico.eu/article/rishi-sunak-ai-testing-tech-ai-safety-institute/

POLITICO

Rishi Sunak promised to make AI safe. Big Tech’s not playing ball.

A ‘landmark’ AI testing deal hailed by the UK leader is being shunned by the emerging tech’s biggest players including OpenAI and Meta.

Mmoderate-tomato My question to Michelle Donelan: Thanks for answering questions from the public...

moderate-tomatoOP•5/14/24, 9:04 AM

Michelle Donelan’s response:

Hi Michael,

I’m afraid the article’s assertion that only one AI lab has allowed "anything approaching pre-deployment access" is false. The AI Safety Institute has been working closely with leading AI companies, and has published details of its approach to model evaluations: https://www.gov.uk/government/publications/ai-safety-institute-approach-to-evaluations/ai-safety-institute-approach-to-evaluations. As I’m sure you’ll understand, there are commercial sensitivities around publishing evaluations, and there is a limit on how much detail I can share on model evaluations.

Legislation risks becoming quickly outdated by rapid advances in a technology that is still not yet fully understood. In our AI White Paper, we instead laid out an approach that is more agile and adaptable, taking a sector-by-sector approach and strengthening regulators, with a £10 million package to boost regulators’ AI capabilities. In February we also set out the case for future targeted, binding requirements on developers of highly-capable general purpose AI systems. We are continuing to work closely with developers, academics, and civil society so we can continue to ensure that AI is developed safely and responsibly.

GOV.UK

AI Safety Institute approach to evaluations

moderate-tomatoOP•5/14/24, 9:05 AM

My response:

Thanks very much for your answer.

I understand the need for commercial sensitivities, but the public should know how many companies have allowed pre-deployment access. Could you provide a number? If only to keep track of how well the new AI Safety Institute is performing its duties.

AI is advancing and changing rapidly, but surely well-written laws can keep this in mind? For example, criminals are always looking for new ways to commit crimes, and the police force is given flexible powers to respond.

Lack of enforcement powers means the AI Safety Institute cannot compel powerful tech companies to comply with even basic safety principles like pre-deployment access. Look forward to seeing laws passed so that the safety institute will have some teeth!

On 9 May, UK’s Michelle Donelan will answer your AI Safety Summit questions on LinkedIn

Similar Threads

Similar Threads

Similar Threads