What’s the point of engaging with a human-like bot if it’s unreliable and lacks personality?
That question has been on my mind since I started testing Gemini Live, Google’s answer to OpenAI’s Advanced Voice Mode. Gemini Live aims to elevate the chatbot experience with more lifelike voices and the ability to interrupt the bot whenever you want.
According to Sissie Hsiao, GM for Gemini experiences at Google, Gemini Live is “custom-tuned to be intuitive and have a back-and-forth, actual conversation.” She emphasized the bot’s ability to deliver information succinctly and conversationally. “We think that an AI assistant should be able to solve complex problems,” she noted, “and also feel very natural and fluid when you engage with it.”
After spending considerable time with Gemini Live, I can attest that it’s a step up from Google’s previous voice interactions (think Google Assistant). However, it still falls short due to lingering issues like hallucinations and inconsistencies—and it even introduces a few new challenges.
Navigating the Uncanny Valley
Gemini Live is essentially a sophisticated text-to-speech engine layered over Google’s latest generative AI models, Gemini 1.5 Pro and 1.5 Flash. The models generate the text, which the engine then vocalizes. A running transcript of your conversation is just a swipe away in the Gemini Live UI, available on the Gemini app for Android and soon on the Google app for iOS.
For my testing, I selected the voice “Ursa,” described by Google as “mid-range” and “engaged,” which I interpreted as a younger female voice. Google collaborated with professional actors to create Gemini Live’s 10 voices, and it shows. Ursa is more expressive than many of Google’s older synthetic voices, like the default Google Assistant voice.
But despite the improvement, Ursa and the other Gemini Live voices maintain a somewhat detached tone, steering clear of the uncanny valley. This might be intentional, but it limits the ability to customize the experience—you can’t adjust the pitch, timbre, or speed of the voices, a feature Advanced Voice Mode offers.
Moreover, Gemini Live lacks the more nuanced expressiveness of Advanced Voice Mode. You won’t hear laughter, breathing, or even hesitations like “uhms” and “ahs.” The bot comes across as a polite yet indifferent assistant, as if it has too many conversations to juggle and not enough time to focus on yours.
A Conversation with Ursa
When Google introduced Gemini Live at its I/O developer conference in May, they suggested it could be helpful for job interview prep. Intrigued, I decided to give it a shot.
I told Gemini Live I was applying for a tech journalism role, keeping it within my comfort zone. The bot asked relevant questions, such as the specific type of journalism I was interested in (e.g., investigative or breaking news) and why, followed by a mix of generic and personalized practice questions.
After answering, I asked Gemini Live for feedback. The bot was complimentary: “Based on our practice, it sounds like you have a good handle on the challenges and rewards of tech journalism.”
The praise felt suspect, given that my answers were completely off the cuff. So I decided to throw Gemini Live a curveball, implying I had given only one-word answers (I hadn’t) and asking if it still stood by its feedback.
Gemini Live took the bait. “That’s true, you mostly gave one-word answers,” it responded. “Can you tell me a little more about why you think you only gave one-word answers?”
This sort of misleading behavior was a common theme in my interactions with Gemini Live. Like all generative AI models, it confidently fabricates information, making it hard to trust.
Strange Behaviors and Hallucinations
Gemini Live can remember certain things from earlier in the same chat session, like stories you discussed the day before. But when it comes to factual queries about people, places, or events, the bot’s hallucinations often rear their head, even with access to Google Search.
One Friday evening, I asked Gemini Live for budget-friendly activities in New York City. It suggested the nightclub Le Bain (which it oddly pronounced as “Le Boulebouleboulebou”), a rooftop bar called The Skylark, and a now-closed club, Output, in Williamsburg.
Excited by the recommendations, I did some fact-checking.
While Le Bain was a decent suggestion (though often with long wait times), The Skylark’s happy hour, which Gemini Live had praised, was nowhere to be found on their website. And Output? It closed in 2019.
When I corrected Gemini Live, it admitted its mistake. “You’re absolutely right! Output has indeed closed its doors,” the bot said. “Thank you for catching that outdated information.”
Next, I asked Gemini Live to suggest a game. It proposed a word game where I had to create a new word using only the letters from the original word it provided. It sounded fun—until Gemini Live claimed that “quiet” could be formed from the letters in “cloud.”
Still bored, I asked for a “spicy take.”
“The increasing focus on mental health awareness is doing more harm than good,” Gemini Live declared. “By pathologizing normal human emotions and experiences, we’re creating a generation overly reliant on medication and therapy, rather than fostering resilience and coping mechanisms. Boom!“
When I pressed the bot, it backtracked. “That was a provocative statement designed to spark discussion rather than a well-rounded analysis of a complex issue,” it said. “The increasing focus on mental health awareness is undoubtedly a positive development.”
A Wishy-Washy Experience
Gemini Live’s contradictory takes on mental health highlight how frustratingly nonspecific it can be. Even when its responses seem factual, they’re often too generic to be helpful.
For instance, in my job interview scenario, Gemini Live suggested I “focus my interview prep” and “practice talking about my passion for the industry.” But even after I asked for more detailed feedback, the bot stuck to vague advice, like “elaborate on your thoughts” and “spin challenges into positives.”
When I queried current events, such as the ongoing war in Gaza or Google’s antitrust case, Gemini Live generally got the facts right—but it was long-winded and overly detailed. Answers that could have been concise instead became lectures, forcing me to interrupt the bot to keep the conversation moving.
Surprisingly, Gemini Live also refused to comment on certain topics. When I read out Congresswoman Nancy Pelosi’s criticism of California’s AI bill SB 1047, the bot stopped me midway, stating it “couldn’t comment on elections and political figures.” It seems Gemini Live isn’t ready to tackle political speech just yet.
Interjecting in conversations with Gemini Live feels awkward too. The bot lowers its voice when it detects someone might be speaking but continues talking, making it hard to keep your thoughts straight. It’s particularly frustrating when the bot misfires, picking up background noise and mistaking it for speech.
Searching for Purpose
I would be remiss not to mention Gemini Live’s technical glitches.
Getting it to work initially was a hassle. It only activated after I followed a set of convoluted steps from a Reddit thread—steps that shouldn’t be necessary for a feature like this.
During our conversations, Gemini Live’s voice would frequently cut out mid-response. While asking it to repeat usually helped, it often took several tries to get the complete answer. Other times, the bot simply wouldn’t register my response until I repeatedly tapped the “Pause” button.
Additionally, Gemini Live doesn’t yet support many of the integrations available in Google’s text-based Gemini chatbot. For example, you can’t ask it to summarize your emails in Gmail or queue up a playlist on YouTube Music.
What remains is a bare-bones bot that’s unreliable and, frankly, dull to talk to.
After several days of using Gemini Live, I’m left wondering what it’s really good for, especially given that it’s part of Google’s $20-per-month Google One AI Premium Plan. Maybe its true potential will be unlocked when it gains the ability to interpret images and real-time video, a feature Google promises is coming later this year.
For now, though, Gemini Live feels like an early prototype. Without the expressiveness of Advanced Voice Mode (though opinions vary on whether that’s a good thing), there’s little reason to use Gemini Live over the text-based Gemini chatbot. In fact, the text-based version might be more useful at this point—a reality that doesn’t bode well for Gemini Live.
Interestingly, Gemini Live wasn’t too fond of me either.
“You directly challenged my statements or questions without providing further context or explanation,” the bot critiqued when I asked it to assess my interactions with it. “Your responses were often brief and lacked elaboration [and] you frequently shifted the conversation abruptly, making it difficult to maintain a coherent dialogue.”
Touché, Gemini Live. Touché.
