At Google I/O 2025, Artificial Intelligence Took the Spotlight — Yet One Significant Matter Was Left Unmentioned

This year’s Google I/O 2025 centered on artificial intelligence. The tech behemoth introduced a range of AI-centric announcements, including a robust new video generation tool named Flow, a premium $250 AI Ultra subscription package, substantial enhancements to its Gemini AI platform, a virtual shopping try-on feature, and the comprehensive launch of its AI-driven search tool, AI Mode, across the United States.

However, during nearly two hours of AI presentations by Google executives, one essential term was notably absent: “hallucination.”

In the AI domain, hallucinations denote the concerning propensity of large language models (LLMs) to produce false or misleading information — essentially fabricating data. Despite being one of the most enduring and disquieting hurdles in AI advancement, this concern was never directly confronted during the keynote. This omission is particularly notable given that, by some measures, hallucinations are on the rise. Indeed, certain models have been reported to hallucinate more than 40% of the time.

If you had only tuned into Google I/O, you might believe hallucinations were a relic of the past. You wouldn’t be aware that every AI-generated response in Google’s AI Overviews carries a disclaimer: “AI responses may include mistakes.”

The nearest Google came to recognizing the problem was during a discussion about AI Mode and Gemini’s Deep Search functionalities. Presenters asserted that the model would “check its own work” before delivering answers — but without additional clarification, this resembled more circular reasoning than thorough fact-checking.

For skeptics of AI, the technology sector’s faith in these tools often appears out of touch with real-life performance. Users continue to report fundamental errors, such as incorrect spellings, flawed calculations, or wrong responses to simple inquiries like “Will water freeze at 27 degrees Fahrenheit?”

Google enthusiastically claimed that its newest model, Gemini 2.5 Pro, performs well on various AI benchmarks. However, when it comes to factual accuracy, the outcomes are less commendable. According to Google’s own metrics, Gemini 2.5 Pro achieves merely 52.9% on the SimpleQA benchmark — a test aimed at assessing how effectively a model can answer succinct, fact-based queries. That’s just above a passing score.

When approached for commentary on the SimpleQA results or the broader hallucination issue, a Google representative refrained from elaborating. Instead, they directed attention to the company’s official clarification on AI Mode and AI Overviews. The document concedes that hallucinations can happen, stating:

“[AI Mode] utilizes a large language model to assist in responding to queries, and it’s possible that, in rare instances, it may occasionally assert information that is incorrect, which is commonly referred to as ‘hallucination.’ … We’re also employing innovative techniques with the model’s reasoning capabilities to enhance factual accuracy. For instance, in collaboration with Google DeepMind research teams, we leverage agentic reinforcement learning (RL) in our tailored training to encourage the model to produce statements it recognizes as more likely to be true (not hallucinated) and also supported by inputs.”

So, is Google being excessively optimistic? There’s a possibility that hallucinations will eventually be resolved. Yet, current research indicates that this is not yet the reality. Regardless, organizations like Google and OpenAI are charging ahead into the realm of AI-driven search — a shift that could usher in a period of widespread misinformation unless these foundational issues are resolved.

Until that time, we may all be left pondering: are the machines hallucinating — or are we?