Scientific and commercial interests clash within Google DeepMind. Researchers are not allowed to publish papers that reveal secrets about Gemini.
DeepMind, a subsidiary of Google, has been a leading institute for AI research for years. The rules around publishing papers that share that research with the academic world have been significantly tightened. Several (former) employees of the DeepMind division testify to the Financial Times how they are not allowed to publish papers or can only do so with delay.
Commercial interests are gradually gaining the upper hand over academic ones within DeepMind. Parent company Google has tied its fate to winning or losing the AI race, and that influences how it approaches scientific research. Papers containing information that could also be interesting for competitors now have to go through a stricter review process.
read also
Google Phases Out Assistant: Gemini Takes Its Place, Without Compatibility Guarantees
Under Embargo
Researchers testify to the Financial Times that their papers sometimes had to remain ‘under embargo’ for up to six months by DeepMind’s management. Other papers are not allowed to be published at all. This applies to papers containing sensitive information about the Gemini models, or comparisons with competitors that put Gemini in a bad light.
DeepMind is also very cautious about sharing information on competing models, especially when it comes to generative AI. In some cases, publications revealing vulnerabilities of other AI systems were also blocked to avoid appearing ‘hostile’.
This causes frustration among researchers: “If you’re not allowed to publish, that’s a career killer for a researcher,” the Financial Times quotes an anonymous witness. For some, it was even the direct reason to leave DeepMind. “The company has started to care more about the product and less about research for the public good,” said another former DeepMind scientist.
Products Over Research
DeepMind responds to the article stating that it still shares a lot of research with the academic world, but acknowledges that the policy has changed. The change of course was initiated by the merger of DeepMind with Google Brain in 2023. Product development takes precedence over scientific research: teams working on commercial Gemini products have more access to resources than teams doing research.
The internal tensions have no impact on product development, as Google is churning out new AI services under the Gemini flag. Last week, Gemini 2.5 with reasoning capabilities was launched. The fierce competition compels Google to step up its game to keep up with OpenAI in the AI race.