'All models are biased, some are...': Zoho's Sridhar Vembu after concerns emerge over hidden biases by DeepSeek

Zoho founder and former CEO Sridhar Vembu on Tuesday weighed in on the growing debate around AI models like DeepSeek, raising concerns about their inherent biases. In a post on X, Vembu pointed out that even so-called “open weight” models, such as DeepSeek, operate as black boxes with two significant unknowns.

“We do not know what kind of training data went into them and second, we do not know what kind of ‘reinforcement learning’ was done to make them ‘behave well,’” Vembu wrote. He explained that “behaving well” could range from simple politeness in user interactions to more complex and contentious practices, such as active censorship of ideas the model’s trainers might want to suppress.

AI models are inherently shaped by the value systems of those who train them, Vembu said. “US-trained models show sensitivity in certain areas while Chinese-trained models show in other areas,” he noted, suggesting that cultural and political contexts influence how these AI systems operate. Drawing from George Box’s famous quote, he paraphrased, “All models are biased, some are useful.”

The Zoho founder further elaborated that while AI models can be evaluated objectively in areas like coding or playing chess—where performance benchmarks are clear—even here, subtle biases might emerge, such as a preference for certain programming languages like JavaScript over C.

“This is why I would trust these models only when there are other objective ways to evaluate their output. This is true for fields like programming, computer-aided design, etc.,” he said. However, he expressed skepticism about relying on AI for complex areas like economic development advice. “I don’t find most academic economics useful, and these models have memorized the textbooks and the value-neutral economics they teach. But that is my own bias.”

On the future of AI, Vembu suggested that as AI evolves, it may become highly personalised. “The ultimate model is going to differ from person to person, and that personal model will train itself on the person’s own system of values, acting as a faithful mirror of their thinking process,” he concluded.

Vembu’s comments came in response to remarks by writer and thinker S Gurumurthy, who criticised DeepSeek for potential biases, stating, “If DeepSeek is designed to suppress the truth, it is also designed to spread lies. Not only on politics, science, sociology, economics—in everything. QED: DeepSeek cannot be universally valid AI. It can spy but it can't reveal.”

China's DeepSeek recently made headlines after surpassing OpenAI’s models on several performance metrics and becoming the top AI app on the iOS store, raising both admiration and concerns within the tech community.