DeepSeek R1 suspected of copying Google Gemini 2.5 Pro
Recently, Chinese laboratory DeepSeek released an updated version of its R1 artificial intelligence model, which handles math and programming tasks excellently.
And developer from Melbourne Sam Peach published curious findings.
The new DeepSeek R1 model prefers words and expressions surprisingly similar to those used by Google Gemini 2.5 Pro. And another researcher noticed that the so-called “thought traces” of this DeepSeek model look practically identical to Gemini’s work traces.
By the way, DeepSeek already had a similar story. Earlier this year, OpenAI reported to Financial Times about evidence of DeepSeek using the distillation method — a technique that allows extracting knowledge from more powerful models. And Bloomberg wrote that Microsoft recorded massive data extraction through OpenAI developer accounts at the end of 2024. And the accounts were supposedly connected to DeepSeek.
Well, this looks like an era of “model cannibalism”. When advanced systems learn not only from human data, but also from output signals of competing models, creating a unique closed loop of digital evolution. Though there’s also a nuance here. All leading companies that have their artificial intelligence models face lawsuits for copyright infringement. Everyone steals from everyone.
