AI Model Vetting: US Moves Toward Mandatory Checks

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

Bob: And I'm Bob Russell. Today: Google unveils autonomous AI agents at I/O, the Trump administration weighs mandatory vetting of powerful AI models, KPMG bets big on Anthropic for global tax advice, a court finally closes the Musk versus OpenAI saga, plus two things you can actually try this week, from a country-wide ChatGPT rollout to a new way to write scientific software. Let's start with Google.

Google launches autonomous AI agents and a new Gemini model at I/O 2026.

Samantha: At its annual I/O developer conference, Google announced something it's calling Gemini Spark. These are AI agents that can work in the background and finish complex tasks on their own. Think about booking a trip, comparing insurance policies, or chasing down a refund. You give it a goal and it just does it.

Bob: And alongside Spark, Google rolled out a faster, cheaper version of its core AI model. The pitch is simple: it's quick enough to power those agents at scale, without the hefty cloud bill you'd normally expect. Google says the agent can hold context across hours, even days, which is what you need if it's actually managing a project rather than answering a single question.

Samantha: The demos were striking. One agent was shown planning a family holiday end to end, picking flights, booking hotels, juggling a budget, and pausing to ask the user only when it hit a real choice. Another handled a small business inbox, drafting replies and flagging the messages that actually needed a human.

Bob: This is the moment the industry has been pointing at for two years. Up to now AI mostly answered questions. Agents take actions. That's a big leap, and it's also where the risk sits. An agent that can spend money or send emails on your behalf can also make expensive mistakes on your behalf.

Samantha: Google says there are guardrails: spending limits, confirmation prompts for anything irreversible, and a log of every action. But security researchers are already warning about prompt injection, where a malicious website tricks the agent into doing something it shouldn't. That's a genuinely new attack surface.

Bob: For the market, this puts pressure on everyone. OpenAI, Anthropic, Mistral, all of them have been moving toward agents, but Google has the distribution. Gemini is baked into Android, into Workspace, into Chrome. If Spark works as advertised, hundreds of millions of people will have an agent by default within months.

Samantha: For Europe, this lands right in the middle of the AI Act rollout. Agents that act autonomously on a user's behalf are exactly the kind of high-impact system Brussels has been writing rules for. Expect questions about liability fast: if an agent books a non-refundable flight to the wrong city, who pays?

Bob: And there's a competitive angle. European firms like Mistral are pushing into industrial AI, but consumer agents at this scale are tough to match. The realistic European play is integration, building specialised agents on top of these models for banking, healthcare, public services, areas where local trust and data rules really matter.

Trump administration weighs executive order to vet powerful AI models before release.

Bob: Washington is reportedly preparing an executive order that would require advanced AI models to be tested by the US government before they can be released to the public. That's a sharp turn from the deregulation stance the Trump administration started with.

Samantha: What triggered it, according to reporting, is concern about a new generation of very capable models, including Anthropic's latest, and rising bipartisan pressure in Congress. Google DeepMind, xAI and Microsoft have already agreed to let the government test future models. So in practice, voluntary cooperation is already happening; the executive order would just make it formal.

Bob: The shift matters because the US has been the holdout. Europe has the AI Act, China has its own approval regime, and Washington was the big economy without a hard pre-release check. If this order lands, you'd suddenly have the three biggest AI markets all running some version of mandatory testing.

Samantha: There's also a clear political read. Public sentiment on AI has soured, especially around jobs and around the speed at which these systems are getting deployed. A vetting process gives the administration something concrete to point at when voters ask who's actually watching this stuff.

Bob: The open question is what testing actually means. Safety against misuse? Bias? National security risks like helping with cyberattacks or bioweapons? The scope decides whether this is a meaningful check or a rubber stamp.

Samantha: For Brussels, this is partly a vindication. The AI Act was criticised by US tech leaders as overreach, and now Washington is moving the same way. It also opens the door to real coordination: shared test standards across the Atlantic would save companies from doing the same safety work twice.

Bob: For European users, the practical effect is that the big American models you already use, Gemini, ChatGPT, Claude, would face an extra hurdle before launch. That could mean slower releases here, but also more confidence that what does ship has been independently looked at.

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.

KPMG taps Anthropic to rebuild its global tax advisory platforms with AI.

Bob: KPMG has signed a deal with Anthropic to rebuild its global tax advisory tools around Claude, Anthropic's AI model. We're talking about the platforms KPMG's tax advisors use every day across dozens of countries, now being rewired so an AI handles a big chunk of the analytical work.

Samantha: This is significant because tax advice is one of the highest-margin services in the consulting world, and also one of the most labour-intensive. A junior associate at a Big Four firm spends years reading rulings, cross-referencing treaties, drafting memos. KPMG is essentially saying that work can be largely automated.

Bob: KPMG's framing is that advisors will move up the value chain, focusing on strategy and client relationships, while the AI does the document-heavy grunt work. The honest read is that the firm needs fewer juniors to deliver the same output. That's a structural shift in how professional services hire.

Samantha: It's also a real win for Anthropic. The company has been quietly building a business in regulated industries, banking, law, now tax, where accuracy and traceability matter more than flashy demos. Picking Claude over OpenAI or Google for a deal this size is a signal.

Bob: And tax is the perfect test case. The rules are written down, they don't change every week, and every answer has to be defensible. If an AI gets it wrong, KPMG signs the bill. So the firm clearly believes the technology is now reliable enough to put its name on the output.

Samantha: For European businesses, this could push down the cost of cross-border tax advice, which is a real burden for mid-sized firms operating across EU member states. Twenty-seven tax codes is a lot, and AI is genuinely good at that kind of pattern matching.

Bob: The flip side is the jobs question. The Netherlands, Ireland, Luxembourg, big tax advisory hubs in Europe, will feel this first. The path into the profession used to be: do the grunt work for five years, then become an advisor. If the grunt work is gone, so is the entry-level rung.

Court closes Musk versus OpenAI lawsuit after years of legal battle.

Samantha: A US court has effectively ended Elon Musk's long-running lawsuit against OpenAI. Musk had argued OpenAI betrayed its original non-profit mission by becoming a for-profit company aligned with Microsoft. The court ruled the case had run out of time, and most of his claims won't go to trial.

Bob: This closes a saga that's been hanging over OpenAI for years. Musk was a co-founder and early funder, left the board, then sued, dropped the suit, refiled it, and at one point tried a hostile takeover bid. The legal cloud has finally lifted, and OpenAI can stop spending lawyer-hours on it.

Samantha: It matters because OpenAI's transition to a for-profit structure was at the heart of the case. With the lawsuit dismissed, the company has a much cleaner runway to raise capital, restructure, and operate without that hanging risk. For Sam Altman personally, it's a major weight off.

Bob: For Musk, it's another setback in his very public feud with the AI industry's biggest name. His own company, xAI, has grown fast, but it hasn't caught OpenAI in users or revenue. The lawsuit was partly a competitive lever, and that lever is now broken.

Samantha: The broader principle the court touched on is interesting too: a non-profit's mission, even one as lofty as "building safe AI for humanity," doesn't automatically translate into a binding contract that former donors can enforce decades later. That has implications for every mission-driven tech foundation out there.

Bob: European regulators have been watching OpenAI's corporate structure closely, especially around the Microsoft tie-up and the data flows into Ireland. With the lawsuit gone, expect Brussels and national data authorities to be the main source of pressure on OpenAI from here on.

Samantha: And for European AI startups, there's a quiet lesson. If you take big-name money early and promise to stay non-profit, that promise may not hold you forever, but the reputational cost of breaking it lingers. Structure your governance carefully from day one.

Malta becomes the first country to offer ChatGPT Plus to all its citizens.

Bob: Here's something you can actually look at this week. OpenAI has partnered with the government of Malta to give every Maltese citizen free access to ChatGPT Plus, the paid tier. Malta becomes the first country in the world to roll out a premium AI assistant nationwide.

Samantha: Plus is normally around twenty euros a month, and it gets you the more capable models, image generation, voice mode, and longer conversations. Malta's government is funding it for residents, framing it as digital infrastructure, the same way you'd fund libraries or public Wi-Fi.

Bob: For listeners not in Malta, the interesting bit is what this looks like in practice. Residents sign in with a government-issued digital ID, link it to a ChatGPT account, and the Plus features just appear. No payment screen. The rollout is happening over the coming weeks.

Samantha: If you are in Malta, or you know someone there, this is worth checking the same day it's available to you. The practical wins are real: drafting emails in Maltese and English, helping kids with homework, summarising government documents, translating medical paperwork. Things people pay for in other countries, free here.

Bob: And it's a test case the rest of Europe will be watching. Estonia, Finland, Ireland, all have strong digital-government traditions, and a national AI subscription could be a logical next step. The cost per citizen is small compared to, say, a public transport subsidy.

Samantha: If you want to see whether your own country could go this route, look at Malta's setup over the next few months: uptake numbers, what people actually use it for, and whether there are problems with data handling. Malta sits inside the EU, so GDPR applies fully, which means the privacy questions will be answered in public.

Bob: And if you're just curious, OpenAI's own announcement page walks through how the partnership works. It's a five-minute read and a useful glimpse of what AI-as-public-service might actually look like.

Google and Harvard release an AI tool that writes scientific software for you.

Samantha: To close out, something more niche but genuinely useful if you work anywhere near research or data. Google and Harvard have published an AI system called Empirical Research Assistance, or ERA, that automatically writes high-performance scientific code. The paper is in Nature, and it's already getting attention.

Bob: What's clever is the approach. It doesn't just generate code once and hand it over. It tries thousands of variations, tests them, refines them, and picks the best version. Tasks that used to take a researcher weeks to code and optimise come back in hours.

Samantha: For the average listener you're probably not writing simulation code, but the principle is what to take away. This same pattern, generate, test, refine, automatically, is starting to show up in everyday tools. GitHub Copilot, Cursor, and others are moving the same direction for regular software development.

Bob: And if you do work in research, in academia, or in any kind of data-heavy job, it's worth reading the Nature paper or at least the summary. The methodology is openly described, which means smaller labs and European universities can build on it without paying for proprietary tooling.

Samantha: That's the practical angle for Europe: free, published, reproducible. A research group in Leuven or Delft can pick this up tomorrow. It's the opposite of the closed-model approach we usually cover.

Bob: Today we covered Google's new AI agents at I/O, Washington moving toward AI model vetting, KPMG's big bet on Claude for tax work, the end of the Musk versus OpenAI lawsuit, Malta giving every citizen ChatGPT Plus, and Google and Harvard's AI that writes scientific code.

Samantha: 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.