AI Governance and Tech Review October 2025 🤖📊
🇪🇺 𝗧𝗵𝗲 𝗘𝗨 𝗔𝗜 𝗔𝗰𝘁 𝗮𝗻𝗱 𝗿𝗲𝗹𝗮𝘁𝗲𝗱 𝗘𝘂𝗿𝗼𝗽𝗲𝗮𝗻 𝗿𝗲𝗴𝘂𝗹𝗮𝘁𝗶𝗼𝗻:
• The EU AI Office now provides a single source of knowledge, guidance and a new risk classification tool: TheAI Act Service Desk.
• Two new strategies were launched: The Apply AI and the AI in Science Strategy:Commission launches two strategies to speed up AI uptake in European industry and science
• The paperSimplifying European AI Regulation provides approaches on how not only simplification but also harmonisation could be addressed.
• After GPAI provisions entered into force in August, aDutch warning– in the context of the country’s recent election – could be interpreted as a first nudge towards actually enforcing them.
📚 𝗔𝗜 𝗹𝗶𝘁𝗲𝗿𝗮𝗰𝘆:
• Boris Johnsonexhibits in an almost caricature-like waywhy AI Literacy is so important.
• While I still believe that pushing AI Literacy is, above all, a strategic interest,this paper, AI literacy requirements according to Art. 4 AI Act. Legal analysis and practical measures, among others by Janine Wendt sheds light on the legal side of AI Literacy, is worth exploring for anyone concerned with compliance, training and awareness, enablement and tech adoption.
🇺🇸 𝗨𝗦 𝗥𝗲𝗴𝘂𝗹𝗮𝘁𝗶𝗼𝗻 & 𝗚𝗼𝘃𝗲𝗿𝗻𝗮𝗻𝗰𝗲:
• Interesting in many ways: one more manifestation ofAI Regulation in the US happening on federal instead of state level, partly parallel to some of the EU AI Act’s requirements on Transparency and hence a possible inspiration for the Code of Practice to be formulated.
• After launching a dedicated user space for minors in September, it took OpenAI not very long to come up with their interpretation of “treating grownups like grownups”:Allowing erotic content.
• Shortly after, Open AI publishedan article on its treating of psychological harms, meant to calm the public, but looking at the facts still revealing more than a Million people globally expressing thoughts of self-harm or suicide when using ChatGPT.
• On a related topic,Character AI bans minors after being sued for teenage suicide.
📄 𝗣𝗮𝗽𝗲𝗿𝘀 & 𝘀𝘁𝘂𝗱𝗶𝗲𝘀:
• Nature published an attention-worthy paper on“The impact of advanced AI systems on democracy”: The impact of advanced AI systems on democracy | Nature Human Behaviour.
•Trust in one’s own country to regulate AI use | Pew Research Center –Surprising to see German’s trust so far at the top!
•Safety Frameworks and Standards: A comparative analysis to advance risk management of frontier AI– no long review needed, but the recommendation: Take the time to read this.
•Economic effects of AI agents – or one scenario of how the interdependencies could play out.
• The IAPP, offering the continiously evolving AI Governance Professional training, published itsnew Body of Knowledge (BoK), effective from February 2026 onwards – a sign of a maturing field.
🚀 𝗧𝗲𝗰𝗵-𝘁𝗿𝗲𝗻𝗱𝘀:
• Open AI made available two large products this month:The Agent Kit and Atlas Introducing ChatGPT Atlas | OpenAI.
• Atlas brought what many had expected but was still impressive when it actually arrived. Many call it the end of the internet – more specifically, it might be the next stage in the decline of the web browser as we know (knew) it. The next step (Atlas taking over workflows) has been thought through by my colleague Thordur.
• Open AI’s Sora reached 1 Million users in 5 days – faster than ChatGPT – but then, what does this number tell us in times a technology is globally available and hyped and provided by a company with market dominance?OpenAI’s Sora hit 1 million downloads in less than five days
• Anthropic thankfully keeps publishing their test results and experiences, here with a new episode ofLLMs reacting to being observed:
• … and a study on LLM poisoning I would like to quote here:A small number of samples can poison LLMs of any size \ Anthropic – “This new study—a collaboration between Anthropic’s Alignment Science team, the UK AISI’s Safeguards team, and The Alan Turing Institute—is the largest poisoning investigation to date. It reveals a surprising finding: in our experimental setup with simple backdoors designed to trigger low-stakes behaviors, poisoning attacks require a near-constant number of documents regardless of model and training data size. This finding challenges the existing assumption that larger models require proportionally more poisoned data. Specifically, we demonstrate that by injecting just 250 malicious documents into pretraining data, adversaries can successfully backdoor LLMs ranging from 600M to 13B parameters. (…) If attackers only need to inject a fixed, small number of documents rather than a percentage of training data, poisoning attacks may be more feasible than previously believed.”
📅 𝗘𝘃𝗲𝗻𝘁𝘀:
• Under the theme“Digital Rerum Novarum: Artificial Intelligence for Peace, Social Justice, and Integral Human Development”, experts from across the world gathered to discuss the future of AI – something happening almost everywhere all the time, but still remarkable in that the Vatican so explicitly addresses responsible AI when most countries still have not come past the point of promoting adaption.
🎧 𝗣𝗼𝗱𝗰𝗮𝘀𝘁𝘀 / 𝗧𝗩-𝘀𝗵𝗼𝘄𝘀:
• My years-long “always in the top 3” podcast,The Hard Fork, on Sora, AI-slop and more.
Jon Steward got sucked into the AI-topic as well – interviewing two important figures of the scene:
• Finishing off with a fairly new discovery:Sinead Bovell‘s now 6 episodes old podcast “I’ve got questions” I’ve Got Questions with Sinead Bovell | Podcast on Spotify – I like the smart questions she asks as much as the interesting guests she picks!
