29 juni 2026 · 13:30
EU AI Act Amended: Deepfake Ban, Extended Deadlines
The European Parliament has formally amended the EU AI Act, extending high-risk compliance deadlines while fast-tracking an outright ban on AI-generated deepfake abuse material from December 2026. The amendment puts accountability squarely on developers and platforms, not just end users. This episode also covers China reclaiming the global supercomputing lead on homegrown chips, a stark BIS warning on AI investment risk, and an open humanoid robot simulator now available to researchers.
Beluister deze aflevering:
Transcript
Samantha: Welcome to The State of Tech, The European Edition, Monday June twenty-nine, 2026. I'm Samantha Lawrence.
Bob: And I'm Bob Russell. Today: the European Parliament rewrites the AI Act and bans deepfake abuse tools, China retakes the supercomputing crown with homegrown chips, the world's central bank advisor warns the AI spending boom could end in a global bust, a robotics company opens up its humanoid simulation to developers everywhere, and to close out, two things you can actually use today: the open photo editor Darktable, and the open-source offline reference reader Kiwix. Let's get into it.
European Parliament rewrites the AI Act, delays high-risk deadlines, and bans deepfake abuse tools outright.
Samantha: We start in Brussels, because the European Parliament has formally amended the AI Act, and the changes are bigger than a simple delay. The main compliance deadline for high-risk AI systems moves from August next year to December 2027. For AI built into regulated products, think medical devices or cars, the deadline shifts further still, to August 2028. Companies that were sprinting toward a summer cutoff just got a serious extension.
Bob: But there is a sharp edge to this. From December this year, the Parliament has added an outright ban on AI systems that generate nonconsensual intimate imagery or child sexual abuse material. And critically, that ban targets both the developers who build these tools and the platforms that deploy them. So a model maker can no longer hide behind the argument that someone else misused their tech. If your system can produce this material, you are on the hook.
Samantha: That framing matters, because until now the debate has often been about output moderation after the fact. This pulls accountability back to the design stage. The delay on high-risk systems is partly a response to industry pressure about readiness, but Parliament is also signalling that delay does not mean softness. Where actual harm is happening today, they are moving faster, not slower.
Bob: For European companies, the practical message is split. If you build enterprise AI tools, you get roughly fifteen extra months to get your documentation, risk assessments, and human oversight in order. That is real breathing room. But if you operate any platform that allows image generation, you now have six months to make sure your guardrails actually hold. Lawyers across the continent will be busy. And expect this December ban to ripple outward, because companies that serve European users tend to apply European rules globally rather than maintain two products.
Samantha: The other signal here is political. Parliament is showing it can move quickly when public outrage builds, and recent deepfake abuse cases involving minors clearly accelerated this. The slower track for industry compliance and the faster track for criminal misuse are now formally in the same law.
China retakes the supercomputing crown with a fully homegrown machine, ending nearly a decade of US dominance.
Bob: Big shift in the global computing map. China has officially overtaken the United States in supercomputing performance for the first time since 2017. Beijing announced that its newest supercomputer outperforms every US machine on the top benchmarks, and crucially, it is built entirely on domestically designed and manufactured chips.
Samantha: That last part is what makes this more than a leaderboard story. For years, US export controls have been designed precisely to prevent this moment. Restrictions on advanced chip exports, on lithography equipment, on design software, all aimed at slowing Chinese high-end computing. This announcement says, in effect, those controls did not stop the climb, they just forced China to build its own stack.
Bob: And a supercomputer is not just a vanity project. These machines drive climate modelling, drug discovery, materials science, and increasingly the training of very large AI models. If China can train frontier AI on entirely domestic hardware, the strategic calculation in Washington changes considerably. The sanctions regime was built on the assumption of dependency, and that assumption is weakening.
Samantha: For Europe, this lands awkwardly. The continent has neither US-scale capital nor Chinese-style state coordination behind its computing infrastructure. The EuroHPC programme has delivered some serious machines, but Europe still imports most of the underlying silicon. This week's news will likely sharpen calls for a European chip strategy with actual teeth, rather than the largely symbolic funding announcements of recent years.
Bob: One caveat worth flagging: independent verification of the benchmark numbers will matter. The official Top500 list updates twice a year, and analysts will want to see this machine run the standard tests under standard conditions. So treat the headline as a strong signal, but the full picture will firm up in the weeks ahead.
Samantha: Quick interruption. If you listen to The State of Tech regularly, hit that like button and subscribe, that way you'll never miss an episode. Okay, moving on.
The Bank for International Settlements warns the AI investment boom could end in a global economic bust.
Bob: The Bank for International Settlements, the institution that advises the world's central banks, has issued its annual report, and it does not pull punches on AI. The BIS warns that the current scale of AI investment could end in what it calls an extended investment bust, with consequences for the broader global economy.
Samantha: The comparison they draw is striking. The BIS argues that the current AI spending wave dwarfs historical bubbles like the railway mania of the nineteenth century and the dot-com boom of the late nineties. We are talking about hundreds of billions of dollars flowing into data centres, chips, and model training, largely from a handful of US tech giants. And the report's central concern is that returns are not yet matching the spending.
Bob: The mechanism they worry about is fairly classic. If lackluster profitability triggers a sudden pullback in financing, the entire chain feels it, the chip designers, the cloud providers, the data centre builders, the energy suppliers. Recent selloffs in tech stocks, which we have seen ripple through markets in recent weeks, are exactly the kind of early warning the BIS is pointing at.
Samantha: For European investors and pension funds, the warning has direct relevance. A lot of European retirement money sits in global equity indices that are now heavily weighted toward a small group of AI-exposed US companies. A sharp correction there does not stay in California, it shows up in Dutch, German, and French pension statements.
Bob: The BIS is not predicting a crash, to be clear. It is flagging asymmetric risk. The upside scenarios for AI are well-publicised; the downside scenarios less so. For European policymakers, the message lines up with growing calls for the continent to build its own AI infrastructure rather than simply consume what the hyperscalers offer. If the boom cools, dependency on a few US firms becomes a strategic vulnerability rather than just a commercial choice.
WIRobotics opens up its humanoid robot simulator, inviting researchers worldwide to build on its work.
Samantha: Closing out the news block with a shift in tone. The robotics company WIRobotics has released the simulation model for its humanoid robot, called ALLEX, and announced a roadmap to gradually open up more of its core technology to outside researchers and developers. The aim is to seed what they call an open ecosystem for physical AI.
Bob: Physical AI is the term the field is settling on for systems that do not just generate text or images, but perceive the world, reason about it, and physically act in it. Until recently, serious humanoid robotics research required either a multi-million-euro lab or access to a commercial robot most universities cannot afford. A high-fidelity simulator changes that equation.
Samantha: The key technical claim here is what is called sim-to-real performance. In plain terms, that means an algorithm trained in the simulation should behave the same way when transferred to the actual physical robot. If that holds up, a researcher in Munich or Delft can develop and test control software for a humanoid without ever touching the hardware, and then deploy it confidently when a physical unit becomes available.
Bob: For European robotics labs, this is genuinely useful. The continent has strong traditions in industrial robotics, think KUKA in Germany or ABB across the Nordics, but humanoid robotics has been dominated by American and Asian players. An open simulator lowers the barrier for European teams to do meaningful work in the field, and to publish results that compete on equal footing.
Samantha: There is a strategic dimension too. By releasing the simulator first, WIRobotics is essentially betting that whoever gets the most researchers building on their platform sets the de facto standards for how humanoid robots are programmed. It is the same playbook that worked for certain AI frameworks a decade ago. Worth watching whether European institutions adopt it widely or stick with existing open robotics platforms.
Darktable is the free, open-source photo editor that gives you serious raw processing without a subscription.
Bob: To close out, two things you can actually use today. First up is Darktable, an open-source photo editing application that has quietly become one of the strongest free alternatives to Adobe Lightroom. It runs on Windows, Mac, and Linux, and it costs nothing.
Samantha: Darktable is built for photographers who shoot in raw format, the unprocessed files that come straight off a camera sensor. It handles colour grading, exposure adjustments, lens corrections, and noise reduction, and it organises your photo library at the same time. Users praise the depth of its tools, particularly the colour science, which many reviewers say rivals or exceeds the paid alternatives for landscape and portrait work.
Bob: There is a learning curve, that is the honest caveat. The interface is built for control, not simplicity, so the first hour can feel intimidating if you are coming from a phone editing app. But there are excellent free tutorials on the project's own website, and once the basic workflow clicks, the speed of editing is genuinely impressive. Everything stays on your own machine, no cloud upload, no subscription, no account.
Samantha: For European users, that local-only approach has real appeal. Your wedding photos, your family archive, your travel pictures, none of it sits on a server in another jurisdiction. The project is community-developed and has been actively maintained for over fifteen years, which is reassuring for anyone who has watched a favourite app get acquired and shut down.
Bob: Quick practical note: budget some time to import an existing collection. Darktable reads from a folder structure you already have, so you do not need to migrate files, but tagging and rating a backlog takes patience. Start with your most recent shoot, get comfortable, then work backwards. Available from darktable dot org, no installer tricks, no upsells.
Kiwix lets you download entire offline knowledge libraries, from Wikipedia to medical guides, for use anywhere.
Samantha: Second find is Kiwix, a free open-source app that lets you download huge reference libraries and read them entirely offline. The most popular package is the full English Wikipedia, but the catalogue also includes medical references, language courses, Project Gutenberg books, and Stack Exchange archives, in dozens of languages.
Bob: The use cases are surprisingly broad. Travellers heading somewhere with patchy mobile coverage can carry the equivalent of a library in their pocket. Teachers in regions with unreliable internet use it in classrooms. And for anyone who simply does not want to be dependent on a connection to look something up, it offers genuine independence from the live web.
Samantha: Kiwix runs on Windows, Mac, Linux, Android, and iOS. The full English Wikipedia with images is a sizable download, well over one hundred gigabytes if you want everything, but you can pick smaller packages, a text-only version, or specific topics like medicine or cooking. Users in humanitarian and educational settings have relied on it for years, which speaks to its stability.
Bob: For European listeners, a nice touch is the language coverage. You can download Wikipedia in French, German, Italian, Dutch, Polish, and many smaller languages, each as a self-contained library. Useful for households where multiple languages matter, or for students preparing for exams without endless tab-switching. Available at kiwix dot org. Completely free, no account needed, and once downloaded the content is yours to keep.
Samantha: Today we covered: the European Parliament rewriting the AI Act with delays and a deepfake ban, China retaking the supercomputing crown on homegrown chips, the Bank for International Settlements warning of an AI investment bust, WIRobotics opening up its humanoid simulator, and two things to try yourself: Darktable for serious photo editing, and Kiwix for an offline library in your pocket.
Bob: Want to know more or react? Visit stateoftech.eu or email us at info@doorzetters.net.
Bob: State of Tech, the tech world in 15 minutes.