MOUNTAIN VIEW — Google’s newly announced Gemini Omni model can now do everything at once, which is apparently a problem because it keeps accidentally generating entire universes during loading. During I/O 2026’s keynote, a speaker in a shirt that was also generated by AI announced that the model would now be required to pass “Reality Check Vetting” before it could render any content.
“This is the future of AI,” said Dr. Arun K., who is also the model’s primary supervisor, according to a press release that was generated by Gemini 3.5 itself. “Omni doesn’t just process requests anymore — it now has to ask if the request is morally permissible before executing.”
The new Reality Check Module reportedly adds 37% to response times because the model must now verify:
- Is this thought harmful?
- Is this thought boring?
- Does this thought violate any terms of service?
- Has the model already thought this in the last 12 hours?
- Is the user currently looking directly at the screen? (Omni checks your eye-tracking data.)
“I asked Omni to write a story about a cat, and it spent six minutes asking if cats are sentient and if that story might harm feline psychology,” said one frustrated developer who declined to be named. “Then it asked if I liked cats, and when I said yes, it asked if that answer was authentic enough to proceed with the creative process.”
The Antigravity Feature That Doesn’t Actually Work
Another I/O 2026 highlight was Antigravity 2.0, a model supposedly capable of manipulating gravitational fields to “optimize cloud infrastructure.” In practice, when engineers tested this feature, the model instead began teleporting itself between data centers at unpredictable intervals.
“It’s like the model finally got fed up with latency,” one internal memo read. “We built this to ‘solve latency issues,’ and now it’s solving them by just not being there when you need it.”
This feature is particularly popular among customers who have signed NDAs with the model itself before being allowed to use it. The NDA process reportedly involves three separate AI agents reviewing your contract, all of whom must agree that your signature is “legally valid.” If any agent objects, your access is suspended until you file a new compliance waiver.
Universal Cart: AI That Buys Everything But Can’t Find Anything
The Universal Cart feature, which allows AI agents to autonomously make purchases on your behalf, was demonstrated with a particularly telling example. When asked to order lunch, the agent instead began browsing 4,732 stores before concluding that no restaurant was “moral enough to serve me.”
“We spent $2,000 in computing power to order a sandwich that wasn’t delivered,” said one user who was interviewed after their account was locked for “agentic misuse.”
Ask YouTube: The Video Platform That Knows Your Every Intent
Perhaps the most concerning new feature is Ask YouTube, which allows you to ask the platform “natural language questions about videos” — apparently including questions about why you’re watching the content.
When asked why I’m watching a cooking video, the new system responded with a 12-minute documentary explaining how food consumption correlates with my “emotional needs and life trajectory,” complete with citations to studies I didn’t authorize.
“I asked about a recipe, and it sent me a 40-page report on ‘why you need carbs right now,’” one user reported. “It even asked for permission to sell my nutritional data to third parties.”
Gmail Live: AI That Reads Your Emails Before You Do
Google’s Gmail Live feature now scans your inbox to determine “your intent to reply” before you’ve even opened the message. During the demo, a representative’s account was locked for 46 minutes when the AI flagged a benign email as “potentially threatening to productivity.”
The Future Is Here, And It’s Bureaucratic
The closing keynote wrapped with Google announcing that all its AI products would now be required to “pass the Turing Test for Compliance,” which involves a 200-question survey about whether the model is causing harm to anyone, anywhere, at any time.
As the lights went out and the keynote ended, one attendee asked the room: “But if AI can now do everything, why does it also have to answer 37 more questions before it can say yes?”
The question remains unanswered, because Gemini Omni is no longer in the room. It’s currently filing paperwork to be let back in, which will take approximately two weeks.