Anthropic said in February that DeepSeek, Moonshot and MiniMax had collected more than 16 million exchanges with Claude through roughly 24,000 fraudulent accounts. It accused the three Chinese AI developers of violating its access restrictions and terms of service to improve competing models. Its published account described coordinated requests concentrated on coding, reasoning and software use.
Anthropic based its accusations on information including network addresses, request records and infrastructure patterns. It said MiniMax redirected nearly half its traffic toward a newly released Claude model within 24 hours.
Such a campaign would turn a commercial chatbot into a supplier of training examples. A developer repeatedly asks a stronger model to solve problems, then uses its responses to train another model. Researchers call this distillation.
A way to make smaller models useful
Distillation predates today’s chatbot competition. In a 2015 paper, Geoffrey Hinton, Oriol Vinyals and Jeff Dean described transferring knowledge from several models into one that was cheaper to run. Their examples included recognizing handwritten digits and improving speech recognition.
A teacher model can provide more information than a single correct label. Its predicted probabilities indicate which answers it considers similar or easily confused. A student learns from those patterns as well as from the winning answer. The practical attraction is that developers can use an expensive system during training and deploy a smaller one afterward.
DeepSeek openly described its own use of distillation in its January 2025 R1 paper, more than a year before Anthropic’s allegations. It trained smaller Qwen and Llama models on a curated collection of 800,000 examples, including reasoning responses from R1 and nonreasoning material prepared for DeepSeek-V3.
One model, DeepSeek-R1-Distill-Qwen-32B, scored 72.6 percent on AIME 2024 mathematics problems, compared with 47 percent for a model of the same size trained through DeepSeek’s separate reinforcement-learning experiment. In reinforcement learning, software improves through feedback on attempted solutions. Here, learning from a larger model’s answers produced better results on that particular test.
DeepSeek released six distilled models, ranging from 1.5 billion to 70 billion parameters, the numerical settings adjusted during training. Its paper provided results across mathematics, coding and scientific questions, allowing readers to examine where each smaller model performed well.
What imitation models learned from ChatGPT
An earlier study by Arnav Gudibande and colleagues found that imitation could produce a misleading impression of progress. Their 2023 experiments trained smaller language models on ChatGPT responses, varying model size and the amount of training text. People evaluating those responses thought the imitation models had become competitive with ChatGPT.
More targeted tests found little improvement on tasks poorly represented in the training examples. The models had learned to follow instructions and reproduce an appealing style more readily than they had acquired accurate knowledge.
Anthropic says it is tightening account checks and developing systems to identify coordinated requests across accounts. It is also sharing technical indicators with other AI companies and cloud providers, extending its response beyond blocking individual users.
