OpenAI and Anthropic Accuse DeepSeek of Model Distillation and IP Infringement: Industry Reactions and Technical Overview

OpenAI announced that it has found evidence DeepSeek used its models for training, accusing the company of intellectual‑property infringement through model distillation, while Microsoft also began probing DeepSeek's possible use of OpenAI's API.

OpenAI: "We need free use of all artists' and writers' works to train models, so we can save money to sue DeepSeek for stealing our stuff!"

Tech media founder Jason of 404 Media mocked OpenAI's stance, and Anthropic founder Dario Amodei published a lengthy post downplaying DeepSeek as a threat, noting its performance is comparable to Claude 3.5 Sonnet from 7‑10 months ago.

"To stay ahead, should we set more constraints?"

Shortly after the accusations, Microsoft integrated DeepSeek models into its AI platform, prompting online commentary that denial is the first step of acceptance.

Common AI Techniques That Violate OpenAI Terms

Investigations suggest DeepSeek may have called OpenAI's API last autumn, potentially leaking data, and OpenAI's terms prohibit using output data to train competing models.

Model distillation, a compression method where a large "teacher" model transfers knowledge to a smaller "student" model, is highlighted as a key technology.

Distillation is highly effective for transferring knowledge from large or ensemble models to smaller ones.

Examples include Together AI distilling Llama 3 into Mamba for 1.6× faster inference, and IBM noting that distillation helps move advanced capabilities into smaller open‑source models, making generative AI more accessible.

OpenAI’s chief scientist Mark Chen acknowledged DeepSeek independently discovered ideas used in OpenAI's o1 project.

Despite these technical merits, OpenAI itself faces criticism for training on copyrighted data, including a lawsuit by The New York Times.

Legal arguments claim that using publicly available internet data for training is reasonable, that large‑scale models rely on scaling rather than any single stolen piece, and that OpenAI’s own practices may be non‑compliant.

"DeepSeek Models Lead Only on Cost"

Anthropic’s Dario Amodei stated DeepSeek’s latest model matches Claude’s performance from 7‑10 months ago but at a lower cost, and that the company’s overall investment is comparable to Anthropic’s AI lab.

Analyst Guo Mingqi highlighted two trends accelerated by DeepSeek R1: continued AI compute growth via optimized training despite slowing scaling laws, and a significant drop in API/token pricing that could broaden AI application adoption.

He warned that while lower costs may boost usage, it remains uncertain whether increased volume will offset price reductions, and that Nvidia will stay a winner when scaling benefits resume.

Reference links:

404 Media article

Financial Times report

Gary Marcus tweet

Additional source