David Barnard-Wills

We wanted to draw attention to this in one of our agent scenarios: that there are likely to be privacy impacts based on people being able to easily task agents to find information on other people. The harassment part is outside our remit, but very likely.

I tried to make a "you're having me on? / Avignon?" joke here, but couldn't get it to work.

Thank you. I think an emergent, socially-grounded answer makes more sense to me than something like "around 0.005 encyclopedias".

Sure. Agree you can't be an expert on everything. If you should deepen your domain of expertise, it'd be useful to know what reasonably counts as a domain.. so history, or history of France, or history of medieval France, or...

Cat, do you have any examples of how 'big' an attainable domain of individual expertise might be?

AI mockup of the cover of a report. Titled ICO tech futures report, agentic AI. It features they stereotypical humanoid robot face made out of circuits...

We were quite careful to avoid humanoid robots as illustrations for AI agents. Then our report is publishrd and people create their own AI mockups of a cover to illustrate their linkedin takes, and the robots return...

Please consider applying for the 10th edition of CNIL-Inria Privacy Award.

The CNIL-Inria Privacy Award is intended to promote research in Computer Science and Privacy and to raise awareness among citizens and decision-makers on privacy and data protection issues.

1/2

Given my job, I have many, many opportunities to use this joke.

I apologise profusely to everybody in advance.

The viral AI agent Moltbot is a security mess - 5 red flags you shouldn't ignore (before it's too late) Handing your computing tasks over to a cute AI crustacean might be tempting - but you should consider these security risks before getting started.

www.zdnet.com/article/molt...

Making regulatory sandboxes work for people and society How can the UK AI Growth Lab contribute to governing AI in the public interest?

The UK government consultation on a proposed AI Growth Lab closed recently.

How can this regulatory sandbox contribute to governing AI in the public interest?

Elsa Donnat highlights the key design choices for building UK regulatory capacity for AI.

www.adalovelaceinstitute.org/blog/regulat...

NPL to establish new Centre for AI Measurement - NPL Read about the leading-edge work we do to make a difference to some of the biggest challenges the world faces in the 21st century.

www.npl.co.uk/news/npl-to-...

Possible retirement hobby:

Claude's new constitution A new approach to a foundational document that expresses and shapes who Claude is

www.anthropic.com/news/claude-... little bit manifesto, little bit person spec, little bit binding spell

Just gave a presentation online Messed up the screen sharing, got flustered, now feeling awkward. Watching videos from Davos to recalibrate.

These two should do the research functions at data protection authorities for their next paper

If an AI system cannot show who authorized a decision, it should not be making it. Wrote about judgment routing, signing authority, and why invisible discretion is dangerous:
makingpublicwork.com/who-holds-th...

On Evaluating Cognitive Capabilities in Machines (and Other "Alien" Intelligences) (Apologies for the length of this post, which means it gets cut off in the email version.

My latest on Substack -- a write-up of the talk I gave at NeurIPS in December.

aiguide.substack.com/p/on-evaluat...

Come and work with me! I'm so excited about the way this project is developing 😊

#PhDSky #AcademicSky #XR #VR #AR

For much more detail and particularly the impacts for regulation, and data protection of agentic AI have a look at ico.org.uk/about-the-ic...

4) Ubiquitous agents. High adoption of high capability agentic AI (but not AGI) gets closer to current marketing hype. Agentic systems mediate vast amounts of personal data and interactions between agents become highly complex.

3) Agents in waiting: low adoption despite increased capabilities and mitigations of llm risks, due to sociotechnical factors (maybe inc compliance or liability concerns). Orgs that do deploy tend to be knowledgeable and deployments are bespoke. Public use of agents is uncommon.

2) Just good enough to be everywhere: High adoption of agentic despite limited capabilities. Embedded in lots of digital infrastructure. Aggressive marketing. Lots of inappropriate or unwise use, and impacts from failures and errors.

1) Scarce, simple agents. Low adoption because of limited capability, close to current baseline. Challenges of LLMs remain. Agentic doesn't live up to hype. It's a toy, an experiment. Heavy supervision needed to get value from agents. Some orgs get burned. We see use cases that aren't feasible.

How widely adopted is agentic AI going to be? How capable will it be? Honestly, I don't know... But we mapped out four potential scenarios: ico.org.uk/about-the-ic...

YouTube video by B Team Films Privacy People (full documentary)

For those who haven't seen it yet but have an interest in #privacy, I encourage you to check out my documentary, Privacy People, streaming for free on YouTube at youtu.be/EqZOzwVaZp8.

(released on the 10th anniversary of Bowie's Black Star, which is auspicious for the report, if perhaps not for me)

One for @cstross.bsky.social maybe?

ICO Tech Futures report on agentic AI ico.org.uk/about-the-ic...

Leading AI expert delays timeline for its possible destruction of humanity Former OpenAI employee Daniel Kokotajlo says progress to AGI is ‘somewhat slower’ than first predicted

Nice gig. Make outlandish predictions, then push out the timeline every year or so. Get press attention every time... www.theguardian.com/technology/2...

History of warfare is probably full of this (for a problematic understanding of 'benefit').