Top 10 AI News July 20 2026: Europe Cracks Open Google

AI News Today July 20 2026: Top 10 Stories

European regulators just did something to Google that no competitor has managed: they are forcing it to let rival AI assistants onto Android phones and to hand over chunks of its search data. On top of that, Google's big model missed its deadline for a third time. Elsewhere, Oracle is cutting up to 30,000 jobs to pay for AI data centers, and a study found AI writing detectors miss nearly one in five AI passages. I read everything so you only need five minutes. Here are today's top 10 AI stories, in plain English.

1. Europe Forces Google to Open Android to Rival AI

The European Commission issued binding orders requiring Google to let rival AI assistants work properly on Android phones and to share parts of its search data with competitors. Under the rules, approved third-party assistants get voice activation and the ability to work across apps on Android, and Google must hand over anonymized data about what people search, click, and view. The search data sharing starts January 2027, and the Android changes are due by July 2027.

To understand why this is huge, think about Google's two biggest advantages. First, Gemini comes preinstalled on billions of Android phones, so most people use it simply because it is there. Second, Google has twenty years of data about what humans search for, which no competitor can buy or copy. Europe just ordered Google to share both. A rival assistant will be able to answer when you say a wake word and act across your apps, and AI developers will get search data they could never otherwise access.

Google is not happy. Its policy chief Kent Walker argued the decisions risk weakening privacy and security protections for millions of Europeans. For every other AI company, though, this is the best news of the month, and it lands exactly when Google looks vulnerable after its model delays.

My take: competitors spent years failing to break Google's grip on Android. Regulators did it in one decision. If you use an Android phone in Europe, you may actually get a real choice of AI assistant by 2027.

2. Google's Big Model Misses Its Deadline for a Third Time

Gemini 3.5 Pro, the flagship model Google has been promising since May, reportedly missed its July 17 target for the third time, and the company is now said to be considering releasing a smaller stopgap model called Gemini 3.6 Flash just to have something new in the market. Google still has not published any official details, pricing, or test scores, so everything circulating is unconfirmed reporting.

Three misses is a different problem than one. A single delay looks like careful engineering, which is genuinely respectable. Three suggests something deeper is wrong, either with the training run itself or with the standard Google set and keeps failing to hit. The talk of a stopgap release is the revealing detail, because putting out a smaller, faster model to fill the gap left by your flagship is basically admitting the flagship is not close to ready.

The cost adds up daily. Companies choosing an AI model this quarter are picking between OpenAI's GPT-5.6, Anthropic's Claude, Grok, and now the free Kimi K3, and every week Gemini is missing is a week those deals get signed with someone else. We covered the fallout in Saturday's roundup.

My take: Google's research team is still one of the best in the world, so this is fixable. But staying quiet while missing deadlines is the worst combination for the businesses trying to decide whether to trust you.

3. Oracle Is Cutting 30,000 Jobs to Pay for AI Data Centers

Oracle is cutting up to 30,000 employees, about 18 percent of its entire workforce, to free up an estimated $8 to $10 billion a year to build AI data centers. The money funds Oracle's part in Stargate, a $500 billion AI infrastructure project with OpenAI and SoftBank, anchored by a $300 billion five-year cloud deal with OpenAI covering 4.5 gigawatts of computing capacity.

This is the most honest look anyone has given us at what the AI boom actually costs. All those enormous data center announcements have to be paid for somehow, and Oracle is paying for its share with the salaries of 30,000 people. The cuts hit its healthcare, cloud, and consulting teams hardest, while deliberately sparing the teams building the AI data centers, which Oracle is hiring for as fast as it can. That is a company converting itself into an AI infrastructure provider, one department at a time.

The risk is that Oracle has bet almost everything on one customer. A $300 billion contract with OpenAI means Oracle's future depends on OpenAI growing, staying able to pay, and surviving its current legal fights and IPO. That is an enormous amount riding on a single relationship.

My take: when people talk about the AI boom, they usually mean stock charts. This is what it looks like on the ground: 30,000 people losing jobs so a company can afford to build data centers. Both things are the same story.

4. Kimi K3 Has the US AI Industry Rattled

Moonshot AI's Kimi K3, the free Chinese model that launched Thursday night and immediately took the top spot on a major coding leaderboard, has genuinely unsettled the US technology industry over the weekend, reopening the debate about how far ahead American AI really is. American labs and investors spent the weekend publicly reassessing, which is not something a routine model release causes.

What makes K3 different from earlier Chinese models is the order it did things. Previous releases competed on being cheaper. K3 competed on being better at coding, beat Anthropic's top model on that leaderboard, and then announced it would give its weights away free on July 27. That combination removes the two comfortable arguments people used to make, that free models are not as good and that Chinese models are budget substitutes rather than real frontier systems.

The honest caveat is that K3 is a specialist, ranking around ninth on general conversation, so it is not a full replacement for the best all-around models. But most business AI spending goes on high-volume coding and agent work, which is exactly what K3 is good at, and it is about to be free.

My take: the July 27 date is the one to circle. That is when a model that just beat a top paid competitor at coding becomes something anyone can download and run for nothing.

5. Microsoft Built an AI Security Tool That Is Cheap Enough to Run Nonstop

Microsoft is preparing Project Perception, an AI security tool that hunts for software vulnerabilities and suggests fixes, and it does something clever: it uses models from Microsoft, OpenAI, and Anthropic together, picking the right one for each job. It scans a company's code, cloud systems, and devices, finds weak points, explains why they matter, and proposes fixes. Microsoft has not confirmed pricing or availability yet.

The clever part is the cost engineering, and it is genuinely good news. Instead of sending every task to the most powerful and most expensive AI, the system routes simple work like log parsing and basic checks to a cheap model, and only calls in a top-tier model when it needs to reason through a complicated attack chain or write a fix that touches live systems. Running frontier AI across an entire codebase used to be far too expensive to do continuously, and smart routing is what makes always-on security scanning realistic.

It also puts Microsoft head to head with Anthropic, whose own AI security program expanded to 150 critical organizations across 15 countries this month. Two well-funded competitors racing to make machine-speed security affordable is genuinely healthy for everyone who needs protecting.

My take: Microsoft using Anthropic's AI inside a product built to compete with Anthropic is peak 2026. But the real story is the cost trick, and it is a technique more teams should copy.

6. SAP Just Spent a Billion Euros on AI That Is Not a Chatbot

German software giant SAP completed its purchase of Prior Labs, a Freiburg startup only about 18 months old, and committed over 1 billion euros across four years to grow it into a leading European AI lab. Prior Labs builds tabular foundation models, which are AI systems designed for spreadsheets and databases rather than text and conversation. Its TabPFN model was published in the scientific journal Nature and beat existing methods across hundreds of independent studies.

SAP's reasoning is refreshingly contrarian. It decided the biggest untapped opportunity in business AI was not chatbots at all, but AI built specifically for the structured data that actually runs companies: sales records, inventories, financial ledgers, transaction tables. Chatbot-style models handle documents well and handle a million-row spreadsheet badly. SAP sits on more business data of that kind than almost anyone, so buying the leading lab in that field and funding it heavily is a serious bet.

It is also a genuinely good European AI story at a time when Europe usually gets described as regulating rather than building. An 18-month-old German startup with a Nature paper being scaled into a frontier lab with a billion euros is exactly the outcome European tech policy has been chasing for a decade.

My take: everyone is fighting over chatbots while an entire adjacent frontier sat mostly ignored. I suspect boring spreadsheet AI will deliver more real business value this decade than another point on a chatbot benchmark.

7. AI Writing Detectors Miss Nearly One in Five AI Passages

Researchers at Epoch AI tested three leading AI writing detectors, Pangram, GPTZero, and Originality.ai, against text written by AI imitating a specific person's writing style. Up to 18 percent of AI-generated passages slipped through undetected, and scientific writing was the most vulnerable category of all.

That failure rate matters enormously because of where these tools get used. Universities use them to catch cheating, publishers use them to screen submissions, and employers use them to check written work, often treating the detector's verdict as proof. A tool that misses almost one in five AI passages when someone simply asks the AI to write in a particular style is not a safe basis for accusing a student or rejecting a candidate. And the weakness in scientific writing is especially worrying given how much academic screening now relies on this software.

The underlying problem is that this is an unfair race. Making AI copy a writing style takes one sentence in a prompt, while detecting it is a genuinely hard statistical problem that gets harder as models improve. Detection is losing, and the gap is widening.

My take: if you are a student or a writer, know that these tools produce false results in both directions. Schools and employers treating detector scores as evidence are making decisions on much shakier ground than they realize.

8. AI Reading Your X-Ray Can Be Confidently Wrong

A new medical benchmark called RadLE 2.0 tested AI models on radiology tasks and found they often give wrong findings with complete confidence. The models do not hedge or flag uncertainty when they are mistaken, which is the specific danger: a hesitant wrong answer invites a second opinion, while a confident wrong answer usually does not.

This lands right as AI pushes deep into healthcare. Just this month, Neko Health raised $700 million for AI-analyzed body scans, Hemispheric raised $52 million for brain-activity AI, and the US government started using ChatGPT to review Medicare and Medicaid records. All of those depend on AI either being right or clearly signalling when it is unsure. A model that is confidently wrong defeats the human double-check that is supposed to catch errors, which makes miscalibrated confidence arguably more dangerous than the error rate itself.

The constructive side is that benchmarks like this are exactly what medical AI needs. You cannot fix what nobody measures, and publishing failure modes openly is how these tools eventually earn the trust to be used safely.

My take: AI in medicine has real promise, and I want it to work. But any system used on patients should be required to say when it is unsure, and until it can, no doctor should treat its output as an answer.

9. China's Big AI Conference Closes With a New Global Club

The World AI Conference in Shanghai closes today after four days that included Xi Jinping's first-ever keynote and the launch of WAICO, the World Artificial Intelligence Cooperation Organization, an international body headquartered in Shanghai with 29 founding countries including Pakistan, Russia, and Kazakhstan. The event ran more than 140 forums with over 1,100 exhibitors, and Huawei used the floor to show off its homegrown AI computing systems.

What matters now is what survives after everyone goes home. Organizations announced with fanfare either turn into real institutions with staff, rules, and a schedule, or they become a press release nobody mentions again. The things to watch are whether WAICO publishes a founding charter, names leaders, and attracts members beyond the original 29, especially countries not already close to Beijing. Xi paired the launch with strong support for open-source AI and promises to help developing countries, which is essentially the recruitment pitch.

The Western response is conspicuously missing. Google's own AI chief called for an international watchdog and a US-led coalition the same week, which quietly admits no such group exists while China's now does.

My take: institutions get built slowly and then shape the rules for decades. Whatever you think of the motives, showing up first with a charter and a headquarters is a real advantage, and right now only one side has done that.

10. Two Dates This Week Could Reshape AI Pricing

Two things happen in the next seven days that matter more than most model launches. On July 24, DeepSeek releases the stable version of its V4 model, ending the constant updates that have kept cautious companies from using it in production. On July 27, Kimi K3's weights go free, meaning the model that just topped a coding leaderboard becomes something anyone can download and run themselves.

The money angle is simple. DeepSeek already charges around 70 times less than the top paid models for the same kind of output, and a stable release removes the last technical excuse not to use it. Kimi K3's free weights go further: no per-use cost at all if you run it on your own machines. For any company spending heavily on AI for coding or automation, the last week of July is the moment to actually test the free options against what they currently pay for.

The sensible advice is to test rather than switch on faith. Run your real work through the free models and your current paid one, compare quality and total cost including running your own servers, and let the numbers decide. The honest answer is usually mixed, with paid models still winning the hardest reasoning.

My take: this is the week the free-versus-paid AI question stops being theoretical for businesses. If free models hold up in real testing, a lot of AI budgets are about to get rewritten.

Frequently Asked Questions

Q: What did the EU order Google to do?

The European Commission issued binding orders requiring Google to let rival AI assistants work across Android with voice activation and cross-app access, and to share anonymized search data including query, click, and view data with competitors. Search data sharing starts January 2027, and Android changes are due by July 2027.

Q: Why is Gemini 3.5 Pro delayed again?

Gemini 3.5 Pro reportedly missed its July 17 target for the third time after falling short on coding and reasoning in testing. Google had already scrapped the original version in June and restarted training. The company is reportedly considering a stopgap Gemini 3.6 Flash release, and has published no official details or benchmarks.

Q: Why is Oracle cutting 30,000 jobs?

Oracle is cutting up to 30,000 employees, roughly 18 percent of its workforce, to free an estimated $8 to $10 billion a year for AI data center construction. The money funds its role in Stargate, a $500 billion project with OpenAI and SoftBank, anchored by a $300 billion five-year cloud contract with OpenAI.

Q: What is Microsoft Project Perception?

Project Perception is Microsoft's AI security tool that finds and fixes software vulnerabilities using models from Microsoft, OpenAI, and Anthropic together. It routes simple tasks to cheap models and complex reasoning to powerful ones, which cuts costs enough to make continuous security scanning practical. It competes with Anthropic's security offering.

Q: Why did SAP buy Prior Labs?

SAP completed its acquisition of Prior Labs and committed over 1 billion euros across four years to build a European frontier AI lab. Prior Labs pioneered tabular foundation models, AI built for spreadsheets and databases rather than text. SAP decided structured business data was a bigger untapped opportunity than chatbots.

Q: Can AI detectors catch AI writing?

Not reliably. Epoch AI tested Pangram, GPTZero, and Originality.ai against AI text imitating a specific writing style and found up to 18 percent of AI passages went undetected, with scientific writing most vulnerable. Detector results should be treated as weak signals, not evidence.

Q: Are AI models safe for reading X-rays?

Not yet without human oversight. The RadLE 2.0 benchmark found AI models frequently deliver wrong radiology findings with full confidence and no signal of uncertainty. That miscalibrated confidence is especially risky because it undermines the human review meant to catch mistakes.

Q: When are Kimi K3's weights free?

Moonshot AI has promised Kimi K3's open weights by July 27, 2026, about eleven days after its API launch. Combined with DeepSeek V4's stable release on July 24, the final week of July is the biggest stretch of free-model releases the industry has seen.