At Google I/O, AI that never hallucinates mistakes

This year, Google I/O 2025 had one concentration: artificial intelligence.

We have already covered all the largest news that came out of the annual developer conference: a new video generation tool from artificial intelligence called Flow. Ultra 250 AI. A ton of new changes on Gemini. The virtual shopping experience feature. It is important to launch the search tool set for all users in the United States.

However, over two hours of Google leaders talking about artificial intelligence, one word we did not hear was “hallucinations”.

Halosa remains one of the most stubborn verbs and Regarding problems with artificial intelligence models. The term refers to the facts that have been invented and inaccurate legs stating that the models of the large language “hallucinations” are in their responses. According to the special Big Ai Brands standards, hallucinations increase – with some models that calm down more than 40 percent of the time.

But if you are watching Google I/O 2025, you will not know that this problem was present. You believed that models such as Gemini have never checked; You will definitely be surprised by the vision of the Google Ai warning. (“Artificial intelligence responses may include errors.”

The closest Google came to acknowledging the problem of hallucinations during part of the presentation on artificial intelligence and deep research capabilities in Gemini. The model will be achieved from his work before providing an answer, we were told-but without further details about this process, it seems like a blind person who leads the blind more than examining the original facts.

For skeptics of artificial intelligence, it appears that the degree of confidence in the silicon valley in these tools is absolute of actual results. Real users notice when artificial intelligence tools fail in simple tasks such as counting, explaining or answering questions such as “Does water freeze at 27 degrees Fahrenheit?“

Google was eager to remind viewers that the latest model of artificial intelligence, Gemini 2.5 Pro, sits on top of many leaders of artificial intelligence. But when it comes to honesty and ability to answer simple questions, Chatbots AI is classified as a curve.

Gemini 2.5 Pro is the most intelligent artificial intelligence model in Google (according to Google), but it is The grades are only 52.9 percent On Simpleqa Simpleqa function test. According to Openai research paperSimpleqa test is “a standard of evaluation The ability of language models to answer short questions seeking facts.(Emphasizing us.)

Google’s actor refused to discuss the Simpleqa standard, or hallucinations in general – but we referred to the Google official Explanation of artificial intelligence and a general view of artificial intelligence. This is what he says:

[AI Mode] A large language model is used to help answer inquiries, and it is possible, in rare cases, sometimes providing inaccurate information at times, known as “hallucinations”. As with an artificial intelligence overview, in some cases, this experiment may offend the interpretation of the web content or the context of missing, as it can happen with any automatic research system …

We also use new approaches with the action capabilities of the model to improve realism. For example, in cooperation with the Google DeepMind research teams, we use the reinforcement of agents (RL) in our training for the model reward to create data known to be more likely to accurately (unclean) and also inputs.

Is it wrong to be Google optimistic? Hallus may yet prove a solution that can be solved, after all. But it seems increasingly clear from research that hallucinations from LLMS is not a solutionable problem now.

This did not prevent companies such as Google and Openai from running forward in the era of searching for artificial intelligence-and this is likely to be an era full of mistakes, unless we are hallucinations.