Gemini, Three Years On: Is This the AI Future We Hoped For?

It’s genuinely hard to believe, but a full three years have now passed since Google’s Gemini assistant made its initial, rather clumsy and clearly underdeveloped, appearance on the global stage.

Google unveiled Gemini — then known as Bard — in February of 2023. True to Google’s typical pattern, the “Gemini” designation emerged several months later, first as the name of an underlying “model” for Bard, then eventually taking over as the primary brand for the entire platform a few months subsequent.

Describing its journey since as anything but smooth would be a colossal understatement. Google has aggressively pushed Gemini into nearly every corner of our digital lives, a strategy reminiscent of the infamous Google+ debacle. This often happened while the product was still too immature to be truly effective, simultaneously sowing widespread confusion between Gemini and the long-standing Google Assistant platform it is gradually displacing on Android and elsewhere.

And, naturally, this doesn’t even touch upon the issue of confidently presented misinformation that Gemini — much like most other large language model systems — frequently provides. Simply put, these systems often get facts wrong with alarming regularity, posing a significant risk for businesses and an even more concerning problem for society at large. Despite this, Google and other developers of similar systems continue their attempts to persuade us that this technology offers a universal solution and an instant answer engine poised to replace traditional search as our primary source of information.

However, despite (deep breath) all of that, as I’ve reflected on Gemini’s three-year mark, I’ve had a sudden realization — an understanding of precisely why its pervasive integration into so many aspects of our daily existence feels like such an exasperating intrusion, even considering the growing array of genuinely helpful capabilities Gemini now brings to the Android ecosystem and its legitimate utility in more circumscribed, specific contexts.

With Google’s annual I/O conference now rapidly approaching, promising another deluge of Gemini-related pronouncements, it’s an opportune moment to pause and consider the fundamental shortcomings in Gemini’s current evolution and the formidable challenges Google faces in addressing them.

[Receive insightful, no-nonsense perspectives directly in your inbox with my complimentary Android Intelligence newsletter. Discover three fresh tips every Friday!]

Gemini’s Fundamental Difficulties

Let’s begin with the obvious: Gemini’s most significant drawback today is, in fact, the very same issue we highlighted two years ago — its persistent failure to be a truly effective assistant for the straightforward types of help most everyday Android users (and other Google-connected device owners) genuinely desire and require in their daily routines.

Part of me had hoped that Google’s extensive development efforts with Gemini over the past few years would have brought about more substantial changes. Yet, reflecting on conversations with fellow Android enthusiasts over recent months and considering Gemini’s numerous deficiencies in my own experience, I’ve come to realize that this core inadequacy remains as pronounced as ever.

And the real-world illustrations, drawn from both my personal encounters and the frustrations shared by others, are (a) seemingly endless — and (b) primarily converge on the recurring theme that the service is both harder to use and less effective than its predecessor.

For example:

• It wasn’t long ago that Google made a considerable fuss about “continued conversation” functionality with Assistant — available on compatible Android devices as well as various Google/Nest smart speakers and displays. This allowed you to initiate a query like “Hey Google, what’s the weather?” and then, after a response, simply follow up with “How about this weekend?” Gemini currently lacks this hands-free capability, and moreover, it fails to retain the context of your previous exchange, meaning it won’t comprehend such a follow-up question even if you explicitly say “Hey Google” again to reactivate it.
• Regarding weather inquiries, attempting to get Gemini to provide straightforward yet specific weather information — such as “When will it rain today?” — often proves to be an exercise in futility. I’ve lost track of how many times someone in my household has tried to extract such data from a Gemini-enabled device, only to eventually give up and manually check the forecast. Given that most ordinary individuals primarily rely on services like this for weather updates, this represents a significant regression.
• Gemini’s handling of connected smart devices — like lights, thermostats, and similar gadgets — is notoriously unreliable and frustrating to the extent that many users I’ve spoken with have simply abandoned trying to use it for such purposes altogether.
• Do you wish to activate a simple stopwatch on your supposedly smart Google-connected display — the sort of basic function you might frequently use in an office or kitchen? Good luck: Gemini, for inexplicable reasons, cannot perform this task.
• While Gemini, much like the Assistant before it, can commit most facts you tell it to memory — a genuinely useful feature — the once-convenient ability to instruct your Android phone to recall your parking spot and then later navigate back to it is bafflingly absent in our new Gemini reality. (Gemini will merely save what you tell it as plain text, lacking the formerly helpful Maps integration and instant navigation option.)
• Similarly, the previously valuable function of setting reminders tied to a specific location — so the alert triggers when you arrive at that place? Unfortunately, that’s largely gone now, with Gemini in charge.
• Need to perform any task without an active internet connection? That’s impossible: Gemini only functions when it’s online, rendering it utterly useless in areas without a signal.

And that, truly, is just the beginning.

Beyond the Basics: Gemini’s Broader Issues

Beyond countless specific examples of shortcomings, there’s the undeniable reality that Gemini often proves to be slow, inconsistent, and unreliable as a fundamental virtual assistant when compared to the system it superseded. So, while it can certainly perform impressive party tricks like generating fabricated images and videos, creating 30-second “songs” on demand, and serving as a verbose conversational partner for those moments you’re yearning to engage in lengthy discussions with a computer, it more often than not represents a step backward for the tasks that truly matter — the straightforward requests most of us actually expect a virtual assistant to handle.

And, once more, this doesn’t even account for the deeply troubling matter of accuracy and the capacity to trust any information Gemini provides as an on-demand assistant. Whether you’re inquiring about your calendar or seeking broader general knowledge, engage with Gemini or any other large language model chatbot sufficiently — and pay close attention to its responses — and you will encounter instances where it simply invents facts and presents these inaccuracies as truth.

As I noted early in the AI chatbot phenomenon, for online interactions, getting something right even 90% of the time is simply not adequate. Precision and thoroughness are paramount. A concise answer is acceptable — but only when it’s correct. And when you can absolutely trust its correctness.

If even one out of every ten attempts at using a service yields a flawed or otherwise unsatisfactory outcome, users tend to rapidly lose confidence in it. Consequently, they then resort to alternative tools for the same objective. More than just a few isolated errors, and the effort simply isn’t worthwhile for the majority of ordinary users.

Studies on AI chatbots’ accuracy rates vary — which in itself likely highlights the sheer inconsistency of these systems, even when asked the identical question multiple times — but if you investigate, you’ll struggle to find many reports suggesting these systems achieve correctness even 90% of the time. Most estimates indicate they perform far, far worse.

And it only takes a single unnoticed inaccuracy or mistaken action to lead you astray — personally or professionally. Consider the Meta director of AI safety (!) who inadvertently allowed an AI agent to erase her inbox recently. Or recall the organizers of an AI conference (!!) that discovered dozens of “hallucinated citations” in accepted papers. Or ask the National Weather Service, whose AI-generated forecasts recently included entirely fabricated town names like “Orangeotild” and “Whata Bod” (and as unbelievable as it sounds, I assure you I’m not inventing that).

The litany of examples just keeps growing. And growing. And growing.

So, while Google now expects us to entrust Gemini with even more intricate functions — such as requesting a ride or facilitating online purchases, both of which debuted this week as limited beta features for specific devices — it’s difficult to envision how it can be consistently and dependably reliable enough to be truly advisable, or even merely acceptable, for use in the real world. (Have you observed how this comparable functionality has fared with Chrome’s recently introduced in-browser equivalent?)

And, more critically, the much simpler, foundational ways in which most of us genuinely want an assistant to help us remain frustratingly incomplete and unpredictable, three full years into Gemini’s existence.

I engage with a great many Android phone owners and individuals deeply integrated into the Google ecosystem, and I’m not sure I’ve encountered anyone who genuinely believes that the current state with Gemini is superior to the previous Assistant experience. Quite the opposite, almost all my discussions in this area revolve around complaints regarding how significantly worse the virtual assistant experience has become in the most crucial aspects, and how Gemini’s new offerings tend to range from “briefly entertaining” to “more alarming than promising, concerning their long-term implications.”

After three years and an immense amount of excessive hype, the most significant questions occupying my thoughts are simply: Was this truly what anyone asked for? And, more importantly: Is this the future we envisioned?

Beyond the confines of Google’s headquarters and the broader tech industry, I suspect you’d struggle to find many definitive “yes” answers.

Receive straightforward insights into Google’s world directly in your inbox with my Android Intelligence newsletter — delivered free, every Friday.