The New AI Fight Is Not About Chatbots Anymore
The strongest point to make first is simple: Anthropic has not publicly proved that Alibaba copied Claude. What it has done is accuse Alibaba-linked operators of running a massive extraction campaign. That difference matters. Allegation is not evidence. But in the AI business, even an allegation can move markets, stir lawmakers, and turn a corporate rivalry into a geopolitical food fight with extra chili oil.
Anthropic claims operators connected to Alibaba and its Qwen AI lab used nearly 25,000 fraudulent accounts to generate more than 28.8 million Claude interactions. The alleged goal was to extract useful capabilities from Claude and help train competing AI systems, according to reports based on Anthropic’s letter to U.S. lawmakers.
That is not casual tinkering. That is not “I asked Claude to write my cousin’s wedding toast.” That is industrial-scale querying, if Anthropic’s claim is accurate.
And now the big question lands with a thud: can you copy an AI by talking to it enough?
Anthropic Says Alibaba Went Fishing Inside Claude
According to TechRadar, Anthropic accused groups linked to Alibaba and Qwen of trying to extract Claude’s capabilities by asking it enormous numbers of questions. Anthropic reportedly told U.S. lawmakers that Alibaba used nearly 25,000 fraudulent accounts and more than 28.8 million interactions to gather proprietary information about Claude.
The claim centers on a technique called model distillation. In plain English, one model learns by studying the answers of another model. It is like training a student by letting them watch a master solve problems all day. Distillation can be legitimate. AI companies use it to make smaller, cheaper, faster models.
But Anthropic is not complaining about normal internal model compression. It alleges something different: a rival allegedly used Claude’s outputs to train competing systems without permission. That is the digital equivalent of sending 25,000 people into a restaurant, ordering every dish, reverse-engineering the recipes, and opening a suspiciously familiar place next door.
Alibaba has not publicly responded to the allegations, and TechRadar noted there has been no independent confirmation of Anthropic’s claims.
What “Distillation Attack” Actually Means
A distillation attack sounds like something involving whiskey, spies, and a basement full of GPUs. Sadly, it is nerdier than that. But it may also be more important.
Model distillation happens when outputs from a stronger AI model help train another model. The weaker model does not need to see the stronger model’s source code. It does not need the original training data. It only needs enough examples of how the stronger model responds.
Ask enough questions. Save enough answers. Use those answers as training material. Repeat at scale.
That is the alleged playbook here. Anthropic says Alibaba-linked operators targeted Claude’s valuable features, including advanced software engineering and agentic reasoning capabilities. TechRadar reported that the questions were complex and detailed, aimed at learning how Claude handled high-value tasks.
This is why AI labs are suddenly sweating through their hoodies. Their models are designed to answer questions. But every answer reveals a little behavior. At small scale, that is normal use. At 28.8 million interactions, it starts looking less like customer service and more like a forensic autopsy.
The Numbers Are the Headline
The numbers give this story its punch. Anthropic’s allegation includes nearly 25,000 fraudulent accounts and more than 28.8 million Claude interactions. Reuters reported that the alleged activity ran from April 22 to June 5, 2026, according to Anthropic’s letter.
That time window matters. This was not a slow drip over several years. Anthropic describes it as a concentrated campaign. If true, the operation would look organized, expensive, and deliberate.
The Times of Bangladesh report, citing BBC and AP reporting, said Anthropic described the campaign as the largest AI model extraction effort of its kind. It also said Anthropic sent its concerns to U.S. lawmakers, including Senators Tim Scott and Elizabeth Warren.
Again, these are Anthropic’s claims. They are not a court judgment. They are not a technical audit released in full for the public to inspect. But the scale alone explains why the story jumped from tech blogs into national-security territory almost immediately.
Twenty-eight million questions is not curiosity. It is a strategy.
Why Claude Is Worth Copying
Claude is not just a chatbot that can summarize meeting notes and politely refuse weird requests. Anthropic has positioned Claude as a high-end AI assistant for coding, reasoning, business tasks, and increasingly complex workflows. Anthropic describes itself as an AI safety and research company focused on reliable, interpretable, and steerable systems.
That positioning matters. If a model performs well at software engineering, multi-step reasoning, cybersecurity analysis, and agentic work, it becomes economically valuable. Very valuable. “Please write me a poem” is cute. “Please automate parts of my enterprise software workflow” is where the money shows up wearing a nice suit.
Anthropic alleges that operators linked to Alibaba targeted Claude’s most valuable traits, including longer task handling and decision-making processes.
In the AI race, high-end capability is the crown jewel. Training frontier models costs enormous sums. It requires talent, data, compute, engineering discipline, and patience. If a competitor can imitate parts of that capability by harvesting outputs, it may shrink the gap without paying the same bill.
That is why Anthropic is waving the red flag so hard.
Alibaba, Qwen, and the Bigger China Angle
Alibaba is not a random company wandering into the AI arena with a snack and a dream. It is one of China’s largest technology giants, and its Qwen model family has become one of the country’s major AI efforts.
Anthropic’s accusation specifically points to operators linked to Alibaba and its Qwen AI lab. TechRadar and Reuters both reported that Anthropic framed the campaign as an effort to extract Claude’s capabilities and accelerate rival AI development.
This is where the story grows teeth. The U.S. and China are already locked in an AI competition involving chips, cloud access, export controls, talent, security rules, and military concerns. Add alleged model extraction to that stew, and suddenly the pot starts making ominous noises.
Anthropic has also accused other Chinese AI firms of similar behavior earlier this year, including DeepSeek, Moonshot AI, and MiniMax, according to TechRadar. OpenAI has also expressed concern about distillation-style extraction by Chinese groups.
So Anthropic is not treating this as a one-off annoyance. It is framing it as a pattern.
The Pentagon Blacklist Makes the Story Hotter
As if the AI allegations were not spicy enough, Alibaba is also fighting a separate U.S. government designation. The Times of Bangladesh reported that Alibaba sued the U.S. Department of Defense to seek removal from the Pentagon’s list of Chinese military companies. The report said Alibaba argued in a petition filed in federal court in San Jose that the designation had “no basis in fact or law” and that the Pentagon failed to follow a fair process.
That designation carries more than symbolic weight. Listed firms can be barred from U.S. defense contracts and face reputational damage.
This does not prove Anthropic’s allegation. It does, however, explain why Washington may view the claim through a national-security lens rather than as a plain old corporate spat.
Anthropic also urged Congress to impose penalties on companies behind such attacks and strengthen protections against misappropriation of U.S.-developed AI technology.
So we now have two overlapping battles: one over alleged AI extraction, and another over Alibaba’s U.S. defense-related designation. That is not a small bonfire. That is a fireworks warehouse with a faulty extension cord.
The Irony Is Impossible to Ignore
Here is the awkward part. AI companies trained many models on huge piles of internet data, public text, licensed material, user data, code, books, websites, forums, and other sources. Now some of those same companies are arguing that their model outputs represent valuable intellectual property that rivals must not harvest.
TechRadar noted this irony directly: firms that trained models on vast collections of publicly available information are now fighting over how their own models should be protected as proprietary assets.
That does not automatically make Anthropic wrong. It does make the debate messy.
The AI industry wants broad access when it trains foundation models. Then it wants tight control once those models become products. Legally, those may be different questions. Morally, rhetorically, and politically, it is a harder sell. The industry is basically saying, “Scraping was innovation when we did it, but theft when you do it to us.”
That argument may still win in court or policy. But nobody should pretend it sounds clean.
Why This Could Reshape AI Access
If Anthropic’s allegation sticks politically, AI companies may lock down access much harder. Expect stricter account verification. Expect aggressive rate limits. Expect more anomaly detection. Expect enterprise contracts to include tougher anti-distillation language. Expect APIs to behave more like airport security after someone tried to bring a suspiciously GPU-shaped suitcase through customs.
The problem is brutal. AI tools become useful because people can interact with them freely. But open access creates extraction risk. Close the doors too much, and developers complain. Leave them open, and rivals may harvest outputs at scale.
Anthropic’s claim shows the central contradiction of commercial AI. Labs want their models to be widely used but not widely copied. That is a delicate balance. Actually, it is less like a balance and more like juggling knives while riding a Roomba.
The likely result is a more guarded AI ecosystem. The best models may become harder to access, especially for users in sensitive jurisdictions or high-risk sectors.
That would hurt researchers, startups, and legitimate developers too. Security always has collateral damage.
What Alibaba Has and Has Not Said
As of the cited reports, Alibaba had not publicly responded to Anthropic’s extraction allegations. TechRadar stated there was no independent confirmation of Anthropic’s claims. The Times of Bangladesh report said BBC contacted Alibaba for comment and sought more details from Anthropic.
That silence matters. It leaves the public with one side of the story. Anthropic has made the allegation. Alibaba has not publicly answered it in the reports reviewed here.
That does not mean Alibaba is guilty. It means the evidentiary picture is incomplete.
For a serious news article, this distinction is non-negotiable. Anthropic says Alibaba-linked operators did it. Reports repeat that claim. But without public technical evidence, logs, account trails, legal filings focused on the extraction claim, or an independent audit, the allegation remains an allegation.
The smart reading is this: the claim is specific, serious, and politically explosive. It is not yet independently proven in the public record.
Confidence level: moderate that Anthropic made the allegation as reported; low to unknown on whether the alleged extraction happened exactly as described.
Why Lawmakers Care
Anthropic sent its allegations to U.S. lawmakers for a reason. This is not only about one company protecting one product. It is about who controls frontier AI capability.
The Times of Bangladesh report said Anthropic urged Congress to impose penalties on companies behind these attacks and strengthen measures to stop U.S.-developed AI technology from being misappropriated.
That request fits a broader policy mood in Washington. U.S. officials already worry about advanced chips, cloud compute, model weights, cyber capabilities, military applications, and the speed at which Chinese AI firms are catching up. Distillation adds another headache: even if model weights remain private, outputs may leak capability.
That is a nightmare for regulators. Traditional export controls focus on physical goods, software access, cloud infrastructure, and chips. But what do you do when the “export” is millions of answers from a chatbot?
You can restrict accounts. You can monitor usage. You can punish companies after the fact. But you cannot un-answer 28.8 million questions.
That is the policy problem in one sentence.
The Business Stakes Are Huge
Training frontier AI models costs a fortune. Companies spend heavily on GPUs, data centers, researchers, infrastructure, safety testing, and deployment. If a rival can imitate valuable behaviors by paying for access and harvesting outputs, the economics get ugly fast.
Anthropic reportedly argued that distillation attacks can turn American AI research and development into a subsidy for geopolitical competitors.
That line is dramatic. It is also strategically smart. Anthropic is not merely saying, “Someone violated our terms of service.” It is saying, “Someone used our expensive work to help a rival ecosystem catch up.”
This framing may influence lawmakers more than a normal IP complaint would. Corporate theft is one thing. National competitiveness is another. Add China, AI, military concerns, and frontier models, and suddenly everyone in Washington finds a microphone.
For Alibaba, the stakes are also enormous. If the allegations gain traction, they could affect trust in Qwen, cloud partnerships, enterprise adoption, and U.S. regulatory scrutiny. Even without a final legal finding, reputational damage can move quickly.
In AI, perception is not everything. But it is a very loud roommate.
The Next AI War May Be About Outputs
The first AI copyright battles focused heavily on training data. Authors, publishers, artists, coders, and media companies asked whether AI labs had the right to train on their work. That fight is still alive.
This new fight moves one layer up. It asks whether model outputs can become protected strategic assets. If a company uses another model’s answers to train its own system, is that fair competition, breach of contract, theft, or something else?
The answer may depend on contracts, jurisdiction, access methods, technical safeguards, intent, scale, and whether fraudulent accounts were used. The alleged fake-account network is crucial because it suggests evasion rather than normal customer use.
Model output is now the battlefield. Not just data. Not just chips. Not just talent. Outputs.
That changes the competitive map. AI companies may increasingly treat every prompt as both a customer interaction and a possible reconnaissance probe. This could make frontier AI less open, less cheap, and less casually accessible.
The friendly chatbot era may not disappear. But behind the smiley interface, the security team is definitely sharpening pencils.
Where This Story Goes Next
The next step is evidence. Anthropic has made a highly specific allegation. Alibaba has not publicly responded in the reviewed reports. Lawmakers may ask for more details. Regulators may probe. Alibaba may deny, litigate, or stay quiet. Other AI labs will watch closely because every major model provider faces the same risk.
The broader lesson is already clear. Frontier AI is no longer just a product race. It is an extraction race, a security race, a policy race, and a trust race. The companies building the most capable systems want users. They need users. But every user interaction can also become training material for someone else.
That is the uncomfortable genius of the modern AI economy. The product teaches. The customer learns. The rival may learn faster.
If Anthropic’s allegations are accurate, Alibaba-linked operators did not need Claude’s source code. They needed access, scale, persistence, and enough fake accounts to turn conversation into imitation.
That is the nightmare. You do not break into the vault. You ask the vault 28.8 million questions until it starts drawing you a map.
Sources
- TechRadar: “Anthropic accuses Alibaba of copying Claude by asking it millions of questions — and sets the stage for a new AI war”
- Times of Bangladesh: “Alibaba faces AI theft allegations, challenges Pentagon blacklist”
- Reuters: “Anthropic says Alibaba illicitly extracted Claude AI model capabilities”
- 4sysops: “Anthropic accuses Alibaba of massive AI distillation attack”