EU AI Act Article 50 live: transparency rules reset

Samantha: Welcome to The State of Tech, The European Edition, Thursday August twenty, 2026. I'm Samantha Lawrence.

Bob: And I'm Bob Russell. Today: sweeping new AI transparency rules go live on both sides of the Atlantic, US authorities warn of AI-powered attacks on factory floors, the White House pulls open-source AI into its safety net, a memory chip crunch that's rippling out from AI into everyday electronics, and to close out, two things you can actually use today: the free European writing tool Ellipsus, and the open-source drawing app Excalidraw. Let's start with the biggest regulatory shift of the week.

Europe and California switch on landmark AI transparency laws on the same day.

Samantha: As of August second, two thousand twenty-six, two major pieces of AI legislation are now operational, and together they set a new global baseline. In Europe, Article fifty of the AI Act kicks in, forcing companies to clearly tell users when they're interacting with an AI system, and to digitally mark AI-generated content so it can be detected. In California, the AI Transparency Act, known as CATA, takes effect with similar requirements. The two rulebooks are not identical, but the direction is unmistakable: synthetic content has to be labelled, and users have to know when a machine is on the other side of the conversation.

Bob: And what makes this significant is the scope. This is not just about high-risk uses like healthcare or hiring anymore. It applies to nearly all generative AI outputs, from chatbots on retail websites to AI-generated images posted on social platforms. Companies now have to build detectability into their systems, so that a piece of AI-generated text, audio, or video carries an invisible marker that other software can pick up. The Commission has published guidance to help firms comply, and Amsterdam has even released a handbook specifically for civil servants. This is regulation moving from paper to code.

Samantha: The enforcement question is where it gets tricky. Labelling a chatbot with "you're talking to an AI" is straightforward. Reliably watermarking a stream of AI-generated images that get cropped, screenshotted, and reposted across a dozen platforms is a very different engineering challenge. Detection tools today are still patchy, and researchers have shown that many watermarks can be stripped with basic editing. So the rules are ambitious, but the technology underneath still needs to catch up.

Bob: For European users, the practical effect starts showing up in the coming weeks. Expect to see more disclosures at the top of chat windows, more labels on AI-generated marketing images, and more platforms rolling out detection features. For smaller European AI companies, compliance is going to be a real cost, and industry groups have already warned that the burden falls disproportionately on startups. The bigger players, they've been preparing for this for over a year. And with California moving in parallel, US firms serving European customers now face a fairly consistent set of expectations, which honestly makes global rollout easier than a fragmented patchwork would have been.

US authorities warn hackers are using AI to attack factory control systems across critical sectors.

Samantha: The FBI, NSA, and CISA jointly issued an urgent advisory yesterday warning of an AI-fuelled campaign targeting Siemens S7 industrial control devices. These are the small computers that run pumps, valves, and production lines in energy plants, water utilities, and factories across Europe and North America. Attackers are using AI to generate exploitation scripts disguised as legitimate software, then using those scripts for reconnaissance, credential theft, and in some cases actively disrupting industrial processes.

Bob: The concerning part is what AI changes in this equation. Attacks on industrial control systems used to require deep specialist knowledge. You needed to understand the specific hardware, the specific protocols, and how a specific factory was wired up. AI lowers that skill barrier dramatically. A less sophisticated attacker can now generate a working exploit by describing what they want. The advisory specifically calls out disguised scripts that look like ordinary maintenance software, which makes them much harder for defenders to spot.

Samantha: For European critical infrastructure, this is not theoretical. Siemens is a European company, and its S7 controllers are everywhere in European industry, from water treatment plants in the Netherlands to manufacturing lines in Germany and Italy. National cybersecurity agencies across the EU are expected to issue their own follow-up guidance in the coming days. The immediate advice is straightforward: segment industrial networks from office networks, patch known vulnerabilities, and monitor for unusual traffic patterns. But the harder truth is that many of these devices were installed years ago and were never designed for a world where AI can generate custom attacks on demand.

Bob: And there's a broader signal here. This is one of the first official advisories that frames AI as a specific accelerant in cyber warfare against physical infrastructure. Until now, most AI cybersecurity conversations have been about phishing emails or deepfake fraud. Attacks on the machines that keep the lights on and the water flowing raise the stakes considerably. Insurance costs for critical infrastructure operators are already climbing, and this advisory will not help that trend.

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 White House extends AI safety testing to powerful open-source models for the first time.

Bob: The White House has confirmed it will expand its AI safety framework to cover open-source AI models once they reach what officials call frontier capabilities. Until now, the framework applied only to closed models from companies like OpenAI and Anthropic. Under the expansion, open-weight models, including those released by Meta and by Chinese research labs, would be subject to federal pre-release safety testing once they hit that capability threshold.

Samantha: This is a meaningful shift in how Washington thinks about AI risk. The open-source AI community has long argued that openness is itself a safety feature, because researchers around the world can inspect the models and find flaws. The counter-argument, which the White House is now leaning into, is that once a model is capable enough to help with, say, designing a bioweapon or writing sophisticated malware, the fact that anyone can download it becomes a serious problem. There's no recall button for an open-weight model. Once it's out, it's out.

Bob: The practical questions are enormous. Who defines frontier capability? How do you test a model before release when the whole point of open source is that anyone can fork it and modify it? And how do you enforce this on labs based in jurisdictions that don't share Washington's approach? The policy statement is short on those specifics for now.

Samantha: For Europe, this creates an interesting dynamic. The EU AI Act already includes provisions for general-purpose AI models with systemic risk, and open-source models are partly exempted, though with conditions. A US move to tighten oversight of open frontier models could push European policymakers to revisit that balance. It also affects European companies and universities that build on top of open models, because if a foundation model suddenly becomes subject to US export controls or pre-release testing delays, everything downstream gets slower.

A global memory chip shortage driven by AI demand pushes prices up as much as four hundred percent.

Bob: Now to a story that isn't about AI algorithms, but about the physical stuff underneath. The global memory chip market is in what the industry is calling RAMmageddon. Prices for DRAM and NAND flash memory have climbed between two hundred and four hundred percent over the past year, driven by explosive demand for high-bandwidth memory used in AI servers. Manufacturers have reallocated capacity toward those higher-margin AI components, and the ripple effects are hitting everyone else.

Samantha: The knock-on effects are showing up in consumer products already. PC makers are quietly raising prices on laptops and desktops. Smartphone manufacturers are absorbing costs on their premium models but passing them through on mid-range devices. Gaming consoles, smart TVs, cameras, even car infotainment systems, they all use the same memory chips that are now being diverted to AI data centres. Analysts expect the shortage to persist for at least two more years, because building new fabrication capacity takes that long.

Bob: For European electronics makers, this is a genuine problem. Companies that assemble consumer devices in Europe are seeing input costs surge, and they don't have the same procurement leverage as the American hyperscalers who are locking in multi-year contracts with memory suppliers. Smaller European brands, in particular, are getting squeezed. Some are switching to older memory generations to keep prices down, which means slower devices for consumers.

Samantha: There's also a strategic angle. Europe has been investing in its own semiconductor capacity through the Chips Act, but memory is not where that investment has been focused. Most memory production is concentrated in South Korea, Taiwan, Japan, and the United States. This shortage is a reminder that when a critical component becomes scarce, geography matters. Expect European industry groups to raise this in Brussels in the coming months, arguing for a more targeted push into memory manufacturing. Whether that translates into actual policy movement is another question, but the case is easier to make when consumer prices are visibly climbing.

The free European writing tool Ellipsus lets teams draft and edit long documents together without lock-in.

Bob: To close out, two things you can actually use today. First up, a tool called Ellipsus. It's a browser-based writing environment built by a small European team, and it's designed for people working on longer texts collaboratively. Think novelists, academic co-authors, journalists working on long features, or teams putting together policy documents. It's free to use, and unlike the big American document platforms, it doesn't mine your text for AI training or lock your files behind a proprietary format.

Samantha: What makes it stand out is how it handles versions and collaboration. Ellipsus keeps a full history of every draft, so you can branch a document, try a different direction, and merge it back later, similar to how software developers work with code. Users report that it makes editing a long manuscript with multiple contributors far less painful than juggling comments in a shared cloud document. Export is straightforward, so you can pull your work into a standard word processor or a publishing tool whenever you want. Nothing is trapped.

Bob: It runs entirely in the browser, so there's no installation, and it works on any operating system, whether you're on a Windows laptop, a Mac, or a Linux machine. There's a free tier that covers most individual needs, and paid tiers for teams that want more storage and advanced features. The team is based in Europe, which for anyone concerned about where their draft manuscript lives, is a meaningful detail. Reviewers particularly praise it for fiction writing, but it's genuinely useful for any long-form project.

Samantha: A quick note on the practical side. If you're currently wrestling with a shared cloud document that has fifteen people commenting on the same paragraph, Ellipsus is worth an afternoon of your time. Import your existing draft, invite one or two collaborators, and try the branching feature on a section you're not sure about. That's the workflow where it really shines.

Excalidraw turns any browser into a shared whiteboard for sketches and diagrams.

Bob: Our second pick is Excalidraw. It's a free, open-source whiteboard tool that runs in your browser, and it's designed to look like you sketched everything by hand on a piece of paper. The visual style is deliberately rough, which sounds odd until you use it. It removes the pressure to make things look polished, so you focus on ideas instead of formatting.

Samantha: It's genuinely useful for a wide range of tasks. Teams use it for architecture diagrams, teachers use it for explaining concepts to students, product managers use it for user journey sketches, and plenty of people just use it for casual brainstorming. Multiple people can collaborate on the same board in real time, and you can export your sketches as images or embed them in documents.

Bob: There's no account required to get started. You just open the site in your browser and start drawing. If you want to save work across sessions or collaborate with others, there's a paid tier that adds cloud storage, but the core tool is free forever. It also works offline once loaded, which is unusual for a browser tool, and there's a self-hosted version for organisations that want to run it on their own servers. That makes it a favourite among European teams with stricter data policies.

Samantha: Give it fifteen minutes with a colleague on a video call. Share a link to a fresh board, sketch out whatever you're trying to explain, and see how much faster you reach a shared understanding than passing slides back and forth. That's the moment most users get hooked on it.

Samantha: Today we covered: Europe and California switching on major AI transparency rules, US authorities warning of AI-driven attacks on industrial control systems, the White House extending safety oversight to open-source frontier AI, a global memory chip shortage rippling out from AI into consumer electronics, and two things to try yourself: the European writing tool Ellipsus, and the collaborative sketching app Excalidraw.

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.