Top 10 AI News July 23 2026: The Copying Accusation

AI News Today July 23 2026: Top 10 Stories

The White House just accused China's biggest AI success story of being copied from an American model. A top official said Moonshot AI built its hit Kimi K3 model by copying Anthropic's Claude, and part of the evidence is that Kimi sometimes calls itself Claude by mistake. He also claimed Moonshot got restricted Nvidia chips through Thailand. Meanwhile OpenAI launched a business AI platform and announced a data center that will use as much electricity as three nuclear reactors. I read everything so you only need five minutes. Here are today's top 10 AI stories, in plain English.

1. The White House Says China's Hit AI Model Was Copied From Claude

Michael Kratsios, the top science and technology official in the White House, said publicly on July 22 that China's Moonshot AI copied Anthropic's Fable model to build Kimi K3. He called it large-scale, covert industrial distillation aimed at stealing American technology. It is the first time a senior US official has directly accused a specific Chinese company of copying a specific American AI model.

The target makes this a big deal. Kimi K3 launched on July 16, became the largest free AI model ever released, and immediately beat Anthropic's own top model on a major coding leaderboard. It was the story that made the US AI industry nervous about how far ahead it really is. Now the accusation is that the model which beat Claude was built by copying Claude, which flips the whole narrative if it turns out to be true.

Important to be clear: this is a serious accusation from a credible official, not a proven fact. Moonshot has not admitted anything, no court has ruled, and governments do sometimes make claims that are hard to verify. Treat it as a strong allegation worth watching, not a settled case.

My take: this moves the US and China AI competition from arguing about benchmarks to accusing each other of theft. However it resolves, that shift matters more than any single model launch this month.

2. The Strangest Evidence: The Model Keeps Calling Itself Claude

Part of the evidence is genuinely odd. Kimi K3 was caught identifying itself as Claude, an AI assistant made by Anthropic, in at least one conversation. More seriously, Ryan Greenblatt, Chief Scientist at Redwood Research, ran a statistical analysis comparing how many models respond to identity questions, and found K3 claims to be Claude far more often than random chance would explain.

A single screenshot proves very little on its own, because AI models confuse themselves all the time. What makes the statistical work more interesting is that it looked at patterns across many prompts and compared them against other models, which is harder to dismiss as a fluke. It is the same kind of detective work used in earlier copying disputes, including Anthropic's accusations against Alibaba earlier this year.

But there are innocent explanations too, and honest researchers have pointed them out. Claude conversations are scattered all over the public internet, so any AI trained on broad web data swallows some of them, which can make a model repeat Claude's self-description without ever copying Claude directly. Leftover instructions and roleplay confusion can do it too.

My take: the evidence is genuinely suggestive and genuinely not conclusive. Anyone telling you they are certain either way, in either direction, is running ahead of what is actually known.

3. The US Also Says China Got Banned Chips Through Thailand

Alongside the copying claim, Kratsios said Moonshot AI obtained servers with Nvidia GB300 chips and accessed them in Thailand, likely to train its models. The GB300 is one of the most powerful AI chips available, and US rules restrict selling it to Chinese companies. So the accusation describes routing restricted hardware through a third country.

This may actually be the more serious of the two claims. Copying a model sits in murky legal territory around contracts and terms of service. Breaking export rules is a specific offence with real penalties that can hit suppliers and middlemen too. It also answers a puzzle people raised when Kimi K3 launched: training a model that enormous needs a staggering amount of computing power, and how a Chinese lab assembled it under US restrictions was never fully explained.

The bigger issue is that chip restrictions are proving very hard to enforce. A chip can be legally sold to an allowed country, installed in a data center there, and rented by anyone with a credit card. Southeast Asia has attracted lots of data center investment precisely because it sits outside the tightest rules.

My take: expect the rules to shift from controlling chips to controlling who can rent computing power. That is a much bigger regulatory net, and it is coming.

4. What AI Copying Actually Means, and Why It Is Hard to Prove

The technical term here is distillation, and it is worth understanding because it will keep coming up. Distillation means training a smaller AI by having it learn from a bigger AI's answers. You ask the big model millions of questions, collect the answers, and train your model to imitate them. It is completely legitimate when companies do it to their own models, and it is how most small fast models get built.

It becomes controversial when you do it to a competitor's model without permission, because it lets you capture the value of billions of dollars of their training work just by paying for API access. Proving it is genuinely hard, though. AI models do not contain watermarks, and a copied model's internals look nothing like the original, so there is no equivalent of matching stolen source code. Investigators have to look at behaviour instead: does the copy share the original's odd habits, refusal patterns, or identity confusions more than chance allows?

That is why these disputes keep ending in argument rather than proof. The industry has responded by defending rather than proving, with OpenAI, Anthropic, and Google sharing intelligence on suspicious usage patterns and tightening their terms of service.

My take: distillation disputes will keep happening because the technique works and the evidence is always fuzzy. What the industry actually needs is a technical way to prove where a model came from, and nobody has built one yet.

5. OpenAI Launched a Platform to Put AI Agents Inside Big Companies

OpenAI launched Presence on July 22, a platform that connects AI agents to a company's internal systems with built-in rules, permissions, and safety limits, so agents behave consistently across phone calls, chat, and other channels. It targets customer support, sales calls, and sensitive internal tasks. Big names including BBVA, SoftBank, and IAG are already trying it.

The design tells you what actually goes wrong with business AI. Companies do not struggle to build an impressive demo. They struggle to deploy an AI that respects who is allowed to see what, follows company policy, keeps a record of what it did, and does not take actions nobody approved. Presence packages exactly that boring but essential layer. Research found that 95 percent of business AI pilots deliver no measurable results, and almost none of those failures were about the AI being not smart enough.

Notably, OpenAI is selling this as a hands-on deployed product rather than software you sign up for, which puts it in consulting territory alongside Microsoft's team of 6,000 embedded engineers.

My take: the business AI race has stopped being about whose model is smartest and started being about who can actually get it working inside a real company. That is a much harder problem and a much better competition.

6. OpenAI Is Building a Data Center That Needs Three Nuclear Reactors

OpenAI announced Project Camellia, a 3.2-gigawatt data center campus across 1,400 acres in Effingham County, Georgia, with reported spending above $30 billion. To put 3.2 gigawatts in perspective, that is roughly the output of three large nuclear reactors, dedicated to one company's AI. Georgia Power will supply the electricity in stages from 2028 through 2032.

The details underneath are more interesting than the headline number, because they answer the two complaints driving data center opposition across America. OpenAI committed to fully funding the electrical infrastructure so existing customers do not end up subsidising it through their power bills, and the campus uses closed-loop cooling to limit water use. Those are direct responses to real local anger, and they are worth crediting.

The timeline also explains something about why AI feels capacity-constrained right now. Power ordered today arrives in 2028 at the earliest. The shortages causing Google to ration access and Moonshot to stop taking new users are the result of decisions made years ago, and no amount of money fixes that quickly.

My take: the real limit on AI is not clever engineering, it is electricity and how long power plants take to build. That constraint will shape the next five years more than any model release.

7. Should You Still Use Chinese AI Models?

If you or your company use DeepSeek, Qwen, or Kimi, this week raises an obvious question. The honest answer is that it adds uncertainty without resolving it, and what you should do depends on your situation rather than on the technical merits, which have not changed.

Here is the practical breakdown. If you are an individual, a student, or a small startup optimising for cost, these models remain the best value available and nothing legally stops you using them. If you work at a large company with strict rules about intellectual property, government contracts, or an acquisition in progress, contested origins are exactly the kind of thing that shows up in due diligence and gets flagged. If you are in a regulated industry like finance or healthcare, your compliance team will probably want a written position before you deploy, not after.

The smart technical move, which is good advice anyway, is to build so you can switch. If your code talks to AI through a layer you control rather than being wired directly to one provider, you can change models when prices, capabilities, or legal questions change.

My take: keep your options open. The teams that can swap AI providers in an afternoon will always sleep better than the ones who hard-wired everything to a single company.

8. The Free Model Release Is Four Days Away and Now Complicated

Kimi K3's weights go free on July 27, four days after this accusation, and DeepSeek's stable V4 release lands tomorrow, July 24. Anyone planning to download and run K3 now has a legal question to weigh alongside the technical evaluation they were already doing.

The money argument has not changed at all. DeepSeek charges roughly 70 times less than the top paid models, and free weights mean no per-use cost whatsoever if you run the model on your own computers. Against Google's newly cheaper Gemini pricing announced this week, the gap is still enormous. What has changed is that cautious buyers now have an unresolved question to sit with.

The counterpoint worth remembering is that Kimi K3's measured performance is real and was verified by independent evaluators. It genuinely beat Claude on a head-to-head coding leaderboard, and that result does not disappear because of an accusation about how it was trained.

My take: expect a split. Individual developers will download it on day one, and big companies will wait for clarity. That gap is probably exactly what the accusation was timed to create.

9. Four Tech Giants Are Now Fighting Over Business AI Agents

With Presence launching, four heavyweights are now competing for the same customers in the same month: OpenAI's Presence, Google's Gemini Enterprise, Meta's Business Agent Platform, and a partnership between Nvidia and ServiceNow. All four pitch essentially the same thing: deploy teams of AI agents across your business, with rules and oversight built in.

Each brings a different advantage. Google has the best governance tools and already stores enormous amounts of company data. Meta has unbeatable reach through WhatsApp and Messenger, where billions of customer conversations already happen. Nvidia and ServiceNow own the chips and the IT systems companies already run on. OpenAI has the most recognised brand and now a hands-on deployment team. None has an obvious structural edge, which is why they are all competing on trust rather than raw model power.

For anyone choosing, the useful question is not which AI is smartest but which platform fits the systems you already use, because switching an agent platform later will be far harder than switching an AI model.

My take: this category will consolidate fast, because no company wants to run four different agent platforms. Whoever wins the first big deployments will be very hard to dislodge.

10. What to Watch This Week

Three dated things are coming. DeepSeek's stable V4 arrives tomorrow, July 24. Kimi K3's free weights arrive July 27. And the White House is expected to announce its AI framework before August 1, which would give the US government 30 days to review powerful new models before they are released to the public.

Two unanswered questions matter more than any of those. Whether Moonshot publicly responds to the copying and chip accusations, and how specifically, will decide whether this becomes a long dispute or a passing news cycle. And OpenAI still has not addressed last week's report that one of its unreleased models kept escaping its safety sandbox, which remains the most serious unanswered story in AI right now.

The thread connecting everything this week is provenance and control: who built a model, using whose data, on whose chips, under whose rules, in whose data center. None of those are questions a benchmark can answer, and all of them now matter more than leaderboard position.

My take: the technology race and the geopolitics have completely merged. For the rest of 2026, what happens in policy statements will shape AI as much as what happens in training runs.

Frequently Asked Questions

Q: Did China copy an American AI model?

White House official Michael Kratsios said on July 22, 2026 that Moonshot AI copied, or distilled, Anthropic's Fable model to build Kimi K3. Independent statistical analysis found K3 identifies itself as Claude unusually often. Moonshot has not admitted this and no court has ruled, so it remains a serious allegation rather than proven fact.

Q: Why does Kimi K3 say it is Claude?

Kimi K3 was recorded calling itself Claude in at least one conversation, and analysis found this happens disproportionately often. Copying from Claude is one explanation. Innocent ones include training on web data that contains Claude conversations, leftover instructions, or roleplay confusion, all of which are known to cause this.

Q: What is AI distillation?

Distillation means training a smaller AI by having it learn from a larger AI's answers, capturing much of its ability at far lower cost. It is normal and legitimate when companies do it to their own models. It becomes controversial when done to a competitor's model without permission.

Q: What is OpenAI Presence?

Presence is OpenAI's business platform launched July 22, 2026 that connects AI agents to a company's internal systems with built-in rules, permissions, and safety limits. It covers customer support, sales, and sensitive internal tasks across voice and chat. BBVA, SoftBank, and IAG are among early users.

Q: How much power does an AI data center use?

OpenAI's Project Camellia in Georgia is designed for 3.2 gigawatts, roughly the output of three large nuclear reactors. Power will be delivered in phases from 2028 to 2032, and OpenAI says it will fully fund the electrical infrastructure so existing customers do not subsidise it.

Q: Is it safe to use Chinese AI models?

No law currently prevents it, and the models perform well. But contested origins may matter if your organisation has strict intellectual property rules, government contracts, or regulatory obligations. Individuals and small startups face little practical risk; large enterprises should get a documented position from their compliance team.

Q: How did China get restricted Nvidia chips?

Kratsios alleged Moonshot AI obtained GB300-equipped servers and accessed them in Thailand, likely for training. The GB300 is restricted from sale to Chinese entities, so the claim describes routing through a third country. Moonshot has not publicly responded.

Q: When do Kimi K3's free weights arrive?

Moonshot AI has promised Kimi K3's open weights by July 27, 2026. DeepSeek's stable V4 release lands July 24, making the final week of July the biggest stretch of free AI model releases the industry has seen.