🦺 𝗔𝗜 𝗦𝗮𝗳𝗲𝘁𝘆:

• The 2026 edition of the𝗜𝗻𝘁𝗲𝗿𝗻𝗮𝘁𝗶𝗼𝗻𝗮𝗹 𝗔𝗜 𝗦𝗮𝗳𝗲𝘁𝘆 𝗥𝗲𝗽𝗼𝗿𝘁has been published. Under the lead of Yoshua Bengio, 100+ experts elaborate on emerging capabilities and risks of General Purpose AI. I found the chapter on 𝗥𝗶𝘀𝗸 𝗠𝗮𝗻𝗮𝗴𝗲𝗺𝗲𝗻𝘁 𝗼𝗳 𝗢𝗽𝗲𝗻-𝘄𝗲𝗶𝗴𝗵𝘁 𝗺𝗼𝗱𝗲𝗹𝘀 (𝟯.𝟰) particularly interesting.

• China specifically aims to passnew rules to protect childrenfrom unsafe and harmful use of AI.

📜𝗘𝗨:

• GPAI again: The Code of Practice on GPAI has been signed by a great variety of organisations and while it is still more interesting to me how those who did NOT sign will act, within the group of signatories a “𝗦𝗶𝗴𝗻𝗮𝘁𝗼𝗿𝘆 𝗧𝗮𝘀𝗸𝗳𝗼𝗿𝗰𝗲” has formed, promising joint efforts in achieving alignment with the CoP.

• The French CNIL (National Commission on Informatics and Liberty) publishedguidanceregarding the applicability of the GDPR-obligation to 𝗶𝗻𝗳𝗼𝗿𝗺 𝗱𝗮𝘁𝗮 𝘀𝘂𝗯𝗷𝗲𝗰𝘁𝘀 𝗮𝗯𝗼𝘂𝘁 𝘁𝗵𝗲 𝘂𝘀𝗲 𝗼𝗳 𝘁𝗵𝗲𝗶𝗿 𝗱𝗮𝘁𝗮 𝗳𝗼𝗿 𝘁𝗿𝗮𝗶𝗻𝗶𝗻𝗴 𝗔𝗜 𝗺𝗼𝗱𝗲𝗹𝘀.

• Germany-specific news: We now have anational law, translating the EU AI Act into national regulation.

🎛️ 𝗔𝗴𝗲𝗻𝘁𝗶𝗰 𝗔𝗜:

• UC Berkeley released the “Agentic AI Risk-Management Standards Profile – CLTC” which analyses in regards to which risks Agentic AI exceed general AI risks – and how to address them; starting, suitably, with the lever of “human control and accountability”.

• ATech Policy Press Perspectiveexplains 𝘄𝗵𝘆 𝗰𝘂𝗿𝗿𝗲𝗻𝘁 𝗘𝗨 𝗔𝗜 𝗿𝗲𝗴𝘂𝗹𝗮𝘁𝗶𝗼𝗻 𝗶𝘀 𝗻𝗼𝘁 𝗿𝗲𝗮𝗱𝘆 𝗳𝗼𝗿 𝘁𝗵𝗲 “𝗺𝘂𝗹𝘁𝗶”-𝗮𝘀𝗽𝗲𝗰𝘁 𝗼𝗳 𝗮𝗴𝗲𝗻𝘁𝗶𝗰 𝗔𝗜 𝘂𝘀𝗲 – multi-actors, multi-systems and also the possible incidents first being reported by third parties. It also proposes article 73 of the EU AI Act (reporting of serious incidents) as a “stepping stone” to fill this gap.

• Related, Anthropic, released a report on“measuring AI agent autonomy in practice”.

📸𝗚𝗿𝗼𝗸:

No need to warm this topic up, but what I did find interesting was how in different parts of the world different, highly culturally shaped reactions came out in a fairly short period of time:

•𝗠𝗮𝗹𝗮𝘆𝘀𝗶𝗮 𝗮𝗻𝗱 𝗜𝗻𝗱𝗼𝗻𝗲𝘀𝗶𝗮banned Grok altogether in light of the deep fake debate.

•𝗕𝗿𝗮𝘀𝗶𝗹, generally known for comparably strict tech regulation, gave xAI 30 days to create technical procedures to identify, review, and remove Grok‑generated fake sexualized content and to suspend related accounts. Grok has since responded by deleting thousands of posts and accounts – but not to the regulators’ satisfaction as further requests suggest.

• The𝗘𝘂𝗿𝗼𝗽𝗲𝗮𝗻 𝗖𝗼𝗺𝗺𝗶𝘀𝘀𝗶𝗼𝗻launched investigations into Grok and particularly X’s recommender system under the DSA.

5️⃣  𝗖𝗹𝗮𝘂𝗱𝗲:

• Thenew Claude constitutionand different vibes heard from other sources within the company, such as their Head of the personality alignment team, Amanda Askell, sound much different than previous versions and worrying to me in terms of suggestion of personhood and AI consciousness as well as directing the debate away from more immediately important things to (to me) still abstract about AI rights and welfare.

• Anthropic’s deal with the Pentagon hashit a standoff because the Department of Defense wants “𝗮𝗻𝘆 𝗹𝗮𝘄𝗳𝘂𝗹 𝘂𝘀𝗲” of Claude, while Anthropic refuses to allow uses like mass domestic surveillance and fully autonomous weapons, even 𝘂𝗻𝗱𝗲𝗿 𝘁𝗵𝗿𝗲𝗮𝘁 𝗼𝗳 𝗯𝗲𝗶𝗻𝗴 𝗹𝗮𝗯𝗲𝗹𝗲𝗱 𝗮 𝘀𝘂𝗽𝗽𝗹𝘆‐𝗰𝗵𝗮𝗶𝗻 𝗿𝗶𝘀𝗸 𝗮𝗻𝗱 𝗰𝘂𝘁 𝗼𝗳𝗳 𝗳𝗿𝗼𝗺 𝗳𝘂𝘁𝘂𝗿𝗲 𝗰𝗼𝗻𝘁𝗿𝗮𝗰𝘁𝘀. In his public statement, CEO Dario Amodei said these threats do not change Anthropic’s stance and that the company is prepared to lose hundreds of millions in federal work rather than drop those two safeguards. At the same time, Pentagon officials acknowledge that Claude is already deeply integrated into classified networks. As a response, Trump requested all federal agencies to cease use of Antrophic’s technologyon Friday, 27th of Feb.

More on the topic, which probably deserves an article by itself can be found here and here.

🔖𝗣𝘂𝗯𝗹𝗶𝗰𝗮𝘁𝗶𝗼𝗻𝘀 𝘄𝗼𝗿𝘁𝗵 𝘁𝗵𝗲 𝗻𝗼𝗻-𝗔𝗜-𝘀𝘂𝗽𝗽𝗼𝗿𝘁𝗲𝗱 𝗰𝗼𝗻𝘀𝘂𝗺𝗽𝘁𝗶𝗼𝗻 (𝗿𝗲𝗮𝗱)

• 𝗧𝗵𝗲 𝗦𝗶𝗹𝗶𝗰𝗼𝗻 𝗚𝗮𝘇𝗲: Thepaper “The silicon gaze: A typology of biases and inequality in LLMs through the lens of place” explores how the technology we currently use (mainly) is not only far from representative but inherently and deeply coined by its geographical origin by developing a five-part typology of bias (availability, pattern, averaging, trope, and proxy) that accounts for the complex ways in which LLMs privilege certain places while rendering others invisible. Everyone can live-test how this plays out in practice via this insightful website.

• Three WEF-publications fit perfectly how I perceive AI Governance and the real-life challenges around it: 𝗙𝗶𝗿𝘀𝘁𝗹𝘆, 𝗔𝗜 𝗚𝗼𝘃𝗲𝗿𝗻𝗮𝗻𝗰𝗲 𝗶𝘀 𝗮𝗹𝘀𝗼 𝗮𝗹𝗺𝗼𝘀𝘁 𝗮𝗹𝘄𝗮𝘆𝘀 𝗖𝗵𝗮𝗻𝗴𝗲 𝗠𝗮𝗻𝗮𝗴𝗲𝗺𝗲𝗻𝘁. 𝗦𝗲𝗰𝗼𝗻𝗱𝗹𝘆 (seehere), 𝗶𝗻𝗰𝗿𝗲𝗮𝘀𝗶𝗻𝗴𝗹𝘆, 𝗔𝗜 𝗚𝗼𝘃𝗲𝗿𝗻𝗮𝗻𝗰𝗲 𝗶𝘀 𝘀𝗲𝗲𝗻 𝗳𝗿𝗼𝗺 𝗮 𝘀𝘁𝗿𝗮𝘁𝗲𝗴𝗶𝗰 𝗲𝗻𝗮𝗯𝗹𝗲𝗺𝗲𝗻𝘁 𝗽𝗲𝗿𝘀𝗽𝗲𝗰𝘁𝗶𝘃𝗲 (see here and here), extending far beyond regulatory compliance.

• What has become a common topic in conversations well beyond work – 𝘁𝗵𝗲 𝗳𝘂𝘁𝘂𝗿𝗲 𝗼𝗳 𝘄𝗼𝗿𝗸 𝗮𝗻𝗱 𝗵𝗼𝘄 𝗔𝗜 𝘄𝗶𝗹𝗹 𝗶𝗺𝗽𝗮𝗰𝘁 𝗶𝘁 – has gotten some solid grounding in Stanford Digital Economy Lab’sresearch. To me, mainly the parts on “what we do not know” serve as useful evidence for the warning not to throw too many things together – particularly in economically challenging times.

💡𝗔𝗜 𝗟𝗶𝘁𝗲𝗿𝗮𝗰𝘆:

• 𝗧𝗵𝗲 𝗨𝗦 𝗗𝗲𝗽𝗮𝗿𝘁𝗺𝗲𝗻𝘁 𝗼𝗳 𝗟𝗮𝗯𝗼𝘂𝗿 𝗿𝗲𝗹𝗲𝗮𝘀𝗲𝗱 𝗮𝗳𝗿𝗮𝗺𝗲𝘄𝗼𝗿𝗸 𝗼𝗻 𝗔𝗜 𝗟𝗶𝘁𝗲𝗿𝗮𝗰𝘆stating “every worker will need baseline AI literacy skills to succeed, regardless of industry or occupation” and breaking the concept down in “understand AI principles”, “explore AI uses”, “direct AI effectively”, “evaluate AI outputs” and “use AI responsibly”.

• 𝗢𝗹𝗶𝘃𝗲𝗿 𝗣𝗮𝘁𝗲l, eloquently as always, points out different aspects of AI Literacy, broader than what we usually think about, such as requirements for critical thinking, inthis podcast.

🤖 𝗔𝗜 𝘁𝗲𝗰𝗵 𝗱𝗲𝘃𝗲𝗹𝗼𝗽𝗺𝗲𝗻𝘁𝘀 – 𝘁𝗵𝗲 𝗯𝗶𝗴𝗴𝗲𝗿 𝗽𝗶𝗰𝘁𝘂𝗿𝗲:

• As was to be expected fromPatrick McGee, he wrote an exciting articleon what we might be aware of, but still losing out of sight in current times: 𝗽𝗵𝘆𝘀𝗶𝗰𝗮𝗹 𝗔𝗜.

The longer the review period, the more likely it becomes that relevant twists and turns are missed – happy to see your additions in the comments ⬇️

Also: A ☕goes to the first person to decode the Claude-paragraph icon.