California's Hidden AI Systems: A Warning for EU AI Act

Samantha: Welcome to The State of Tech, The European Edition, Sunday June twenty-first, 2026. I'm Samantha Lawrence.

Bob: And I'm Bob Russell. Today: Paris wraps Europe's biggest tech show with India taking centre stage on AI, Taiwan's memory chip designers heading for an eye-watering revenue jump, the Philippines fast-tracking semiconductor and AI investments, California admitting it quietly runs six high-risk AI systems on its own residents, and to close out, two things you can actually try today: the open-source photo organiser PhotoPrism and the AI transcription tool MacWhisper. Let's start in Paris.

VivaTech 2026 closes with two hundred thousand visitors as Europe stakes its claim in the global AI race.

Samantha: VivaTech wrapped up yesterday in Paris, and the numbers tell their own story. Two hundred thousand visitors from one hundred and sixty-five nationalities, more than fifteen thousand startups on the floor, and a tenth-anniversary edition that very deliberately put artificial intelligence at the centre of everything. India was the AI Country Partner this year, with Prime Minister Narendra Modi leading the delegation in person. That is a notable signal, because for years the conversation about AI hubs has bounced between San Francisco, Beijing, and, occasionally, London. Paris is now firmly in that conversation too.

Bob: And the line-up backed that up. Jeff Bezos was on stage, Yann LeCun was on stage, and the first VivaTech-Bloomberg Awards handed out prizes to figures shaping global tech. What stood out on the show floor was the breadth: smart contact lenses, humanoid robots controlled by thought signals, and three-D printed medical implants that the body absorbs over time. That mix matters, because the European pitch has always been that AI is one technology among many, not the only game in town. Robotics, biotech, and hardware were given equal billing.

Samantha: The European angle is really about positioning. France has been pushing hard to make Paris the centre of gravity for European AI, with Mistral as the home-grown champion and a steady drumbeat of state-backed announcements. VivaTech is the showcase that ties that strategy together. For European companies and investors, the takeaway is that the funding pipeline, the talent, and the political backing are now aligning in a way they were not three years ago. The question is whether that converts into companies that can compete at the scale of the US and Chinese giants.

Bob: There is also a soft-power story here. Hosting Modi and signing India up as the AI partner is a deliberate move to anchor Europe inside the wider, non-American AI supply chain, especially on talent and services. With Washington's export rules tightening and Beijing building its own stack, Brussels and Paris want optionality. For listeners in Europe, the practical effect is that more of the AI tools you use over the next few years may come from European-Indian partnerships, not only from the usual American names. That is a quieter shift than the headline numbers, but probably the more important one.

Taiwan's memory chip designers brace for a two hundred and forty-five percent revenue jump as AI data centres devour every chip they can make.

Samantha: Now to Taiwan, where memory chip designers are looking at a year that almost defies belief. The industry is projecting a two hundred and forty-five percent jump in revenue for 2026, driven almost entirely by demand from AI data centres. Memory prices are climbing, supply is tight, and the buyers are the hyperscalers and specialised AI firms building out training and inference capacity. For context, memory is the unglamorous cousin of the AI chip story, but every Nvidia accelerator needs huge stacks of high-bandwidth memory next to it, and that is where Taiwan sits.

Bob: What is interesting is the pivot inside the industry. For decades, memory chips were tied to the consumer cycle: phones, laptops, gaming PCs. Now the centre of gravity has shifted to AI infrastructure, which pays more and orders in much larger volumes. That changes how these companies invest, who they hire, and how exposed they are to a possible AI slowdown later on. If the AI build-out cools, memory makers will feel it faster and harder than they used to.

Samantha: For Europe, this has two direct consequences. First, anyone building AI infrastructure here, whether that is a sovereign cloud project, a research lab, or a corporate data centre, is going to pay more and wait longer for the memory components. Second, European policymakers who have been trying to reduce dependence on Asian semiconductor supply now have another data point. The Chips Act money has been flowing toward logic chips and packaging, but memory remains a gap. Expect that gap to move up the political agenda over the coming months, especially as price pressure starts showing up in cloud bills.

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 Philippines fast-tracks incentives to lure semiconductor and AI investment, betting on a Southeast Asian tech corridor.

Bob: The Philippines is making a serious push to become a semiconductor and AI hub in Southeast Asia. The Marcos government is accelerating incentives and projects aimed at chip companies and AI firms, with the Luzon Economic Corridor as the main investment channel. They are also activating an industry advisory council to coordinate between government, universities, and companies already operating there, including Texas Instruments. The pitch is straightforward: existing manufacturing experience, a large pool of English-speaking engineers, and now a government willing to move fast on permits and tax breaks.

Samantha: This fits a wider regional pattern. Vietnam, Malaysia, and Thailand have all been positioning themselves as alternatives to Taiwan and South Korea for parts of the chip supply chain. The Philippines was a bit late to that conversation, partly because of infrastructure constraints and partly because the policy was unclear. What is new is the speed and the explicit linking of semiconductors to AI as a single industrial policy, rather than treating them as separate tracks.

Bob: For European companies, this opens a concrete door. Firms like ASML, Infineon, and STMicroelectronics already have Southeast Asian operations, and a government actively offering incentives changes the calculation for where to put the next packaging plant or design centre. There is also a geopolitical angle. Europe has been talking about diversifying away from concentrated dependencies in East Asia, and a friendly, English-speaking partner that fits inside US-aligned trade frameworks is attractive. The risk, as always, is that incentive races to the bottom end up subsidising investments that would have happened anyway.

Samantha: The bottom line is that the map of where chips and AI hardware get made is being redrawn in real time, and Manila has decided it does not want to sit this round out. For European listeners, watch which companies announce Philippine investments over the next six months. That will tell you whether the incentives are working or whether this stays a press release.

California admits to running six high-risk AI systems on its own residents, some of them never disclosed before.

Bob: California has quietly admitted to using six high-risk automated systems that make significant decisions about residents' lives, and several of them had never been publicly disclosed. The list includes tools that predict recidivism for incarcerated people, systems that flag potentially fraudulent unemployment claims, software that remotely monitors students during university exams, and detectors meant to spot AI-generated text in student assignments. The disclosure came because of a state law that requires agencies to report this kind of usage, which is the only reason the public knows about it at all.

Samantha: The accountability story here is uncomfortable. These are systems making consequential calls, whether someone gets parole, whether a jobless claim gets paid, whether a student is accused of cheating, and they were running without public knowledge. The technical accuracy of recidivism prediction and AI-generated text detection is contested in the academic literature, with multiple studies showing high false-positive rates. Running those tools without disclosure means people affected by their decisions had no real way to challenge them.

Bob: For Europe, this lands right in the middle of the AI Act implementation debate. The Act explicitly classifies several of these use cases, criminal justice scoring, benefits decisions, education monitoring, as high-risk, with transparency and human oversight requirements attached. California is essentially showing what happens when those obligations do not exist or are weakly enforced: agencies deploy first and disclose later, if at all. European regulators will be watching this as a real-world example of why the registration and audit requirements in the AI Act exist.

Samantha: There is also a procurement lesson. Many of these systems were bought from private vendors, which means accountability is split between the agency and the supplier. European public bodies are about to go through the same process under the AI Act, and the California case is a useful warning that without strict procurement rules, the same opacity can build up here. Expect civil liberties groups in Europe to cite this disclosure in the coming weeks.

PhotoPrism is the open-source photo library that finally lets you organise decades of pictures without handing them to Google.

Bob: To close out, two things you can actually use today. The first is PhotoPrism, a free, open-source photo management tool that has been gaining real traction with people trying to escape the big cloud photo services. The idea is simple: you point it at your photo collection, whether that is on a home server, a network drive, or a small computer like a Raspberry Pi, and it builds a searchable library with face recognition, location maps, and automatic tagging. All of it runs locally, so your photos never leave your network.

Samantha: What makes PhotoPrism interesting right now is that it has matured to the point where it is genuinely competitive with Google Photos and Apple Photos on the features that matter. Users report that the search works well, you can type "beach 2019" or "dog in snow" and it finds the right pictures, the face grouping is accurate, and the mobile interface is solid. It is not a one-click install, you do need to be comfortable running it in a container or on a small home server, but the documentation has improved a lot.

Bob: The European angle is practical. Anyone uncomfortable with the data practices of the big US photo services, and that is a growing group given the AI training debates of the past year, now has a credible alternative that keeps everything on your own hardware. It works on Linux, Windows, and Mac, the core version is free, and there is a paid tier for households that want extras like more advanced face recognition. For families sitting on twenty years of photos that are currently scattered across phones, laptops, and old hard drives, PhotoPrism is the tool that pulls it all into one place and keeps it private.

Samantha: One tip from the community: start with a copy of your photo library, not the original, while you get the hang of it. Once you trust the setup, you can point it at the real collection. The project is actively maintained and the community forums are helpful, which matters for software you are going to entrust with family memories.

MacWhisper turns hours of interviews and meetings into searchable text on your own laptop, no cloud upload required.

Bob: The second find is MacWhisper, a transcription app for Mac users that has become something of a quiet favourite among journalists, researchers, and anyone who has to deal with recorded audio. It uses OpenAI's Whisper speech-recognition model, but runs it entirely on your own machine. You drop in an audio or video file, and you get back a clean transcript, with speaker separation in the paid version. No upload, no subscription required for the basic version, and no waiting in a cloud queue.

Samantha: The reason this matters now is that transcription used to mean either paying per minute to a cloud service or accepting low-quality automatic captions. MacWhisper changes that calculation. Reviewers consistently praise the accuracy, especially for English, and the app handles long files without falling over. There is a free tier that covers most casual needs, and a one-time paid version that adds features like speaker labels, batch processing, and export to subtitle formats.

Bob: For European users, the privacy angle is the selling point. If you are a lawyer, a doctor, a journalist with confidential sources, or just someone who does not want meeting recordings sitting on a server in another country, local transcription is the only sensible option. MacWhisper is Mac-only, which is the obvious limitation, but Windows and Linux users have similar open-source projects like WhisperDesktop that do roughly the same job. The whole category of local AI transcription has matured to the point where there is really no good reason to upload sensitive audio to a cloud service anymore.

Samantha: Today we covered: VivaTech wrapping up in Paris with India as AI partner, Taiwan's memory chip designers heading for a huge revenue surge, the Philippines pushing incentives for semiconductors and AI, California's disclosure of six high-risk AI systems, and two things to try yourself: PhotoPrism for organising your photo library at home, and MacWhisper for private, local transcription.

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.