GAFAM NewsEssay

Can Google make an AI comeback? Reading Gemini through Google+ and other past failures

Where Google stands in AI, and what Google+, Glass, Nexus Q and Google Home teach about turning impressive demos into products people keep using.

First published in Japanese: October 26, 2026
English edition: November 2, 2026

Read the Japanese original on note「GoogleはAIで巻き返せるか？　Google+など過去の失敗から読むGeminiの現在地」

Good evening. The Owl here.

Will Google make a comeback in AI?

There are plenty of reasons to want to answer “yes.”

Google has deep strength in basic AI research and product development, enormous money and talent, and its own data centers. It also has strong touchpoints with users: Search, Gmail, Android and YouTube.

At the same time, there is no shortage of reasons to want to answer “no.”

Google is a company with a long list of businesses and products it announced with grand ambitions, failed to bring up to expectations, and quietly wound down. Behind them you can glimpse the weaknesses of a big company: an organization split into silos, and caution and impatience living side by side.

I’m just an owl in a tree, not a prophet or an expert, so I won’t try to answer that question with a “yes” or “no” today.

But if we sort out where Google stands in AI, what it has won back and where it is stumbling, and look at what futures past bets like Google+, Google Glass, Nexus Q and Google Home promised and what reality they actually met, we may be able to see what is really being tested in today’s AI race.

That’s what was on my mind as I sat down to write this piece.

Note: the description of the current situation in this article is based on information as of October 26, 2026.

What Google+ showed about the limits of lock-in

If we’re looking back at Google’s past failures, Google+ is the place to start.

Google+ was Google’s attempt to take the center of social networking, backed by the huge user base of Search, Gmail, YouTube and Android. At its developer conference, Google I/O, in 2012, Google said Google+ had passed 250 million sign-ups and showed off a tablet app and an events feature.[1]

Google+ was never meant to be just a standalone social network. Google wanted its profiles to become each user’s “face,” shared across the company’s many services.

The YouTube integration is the clearest example. In 2013, to comment on YouTube, you were required to connect your channel to a Google+ profile or page.[2] Even people who only wanted to leave a reaction under a video were pushed into dealing with Google+. It was a lock-in move, pulling users of existing services into Google+.

But in 2015, YouTube reversed course and stopped requiring a Google+ profile to upload videos, comment or create a channel.[3]

Google+ carried on after that, but in 2018 Google itself acknowledged that it had not achieved broad adoption among consumers or developers. Ninety percent of “sessions,” a single stretch of use, lasted less than five seconds.[4] The consumer service shut down in April 2019.[5]

You can’t blame the shutdown on the YouTube integration alone. Still, the story shows the limits of Google’s strengths. Google could build mechanisms that nudged people to sign up, but it couldn’t create a reason for people to want to open the app on their own.

What a social network needs is relationships, conversation, a sense of comfort, and daily habit. Tightening links to existing services doesn’t produce any of those by itself.

Something similar could happen with AI. Google can build its own AI, Gemini, into Search, Gmail and Android. But lots of chances to use it don’t automatically turn into the trust that makes people want to hand things over to Gemini.

A stumbling start, nearly recovered through multimodal AI

So how has Google moved in the generative AI race?

ChatGPT was released to the public in November 2022.[6] Google was the company that created the Transformer, the architecture underlying much of generative AI, and it had already built a conversational AI called LaMDA.[7][8] Even so, it was OpenAI’s ChatGPT that suddenly brought conversational generative AI into everyday life.

In February 2023, Google put its conversational AI, Bard, front and center. But a factual error about astronomy turned up in a promotional answer, and it stumbled out of the gate.[9] The Gemini introduction video that December gave the impression that the model was seeing, hearing and answering in real time, but Google’s own making-of explainer shows the interaction was done with still images and text prompts.[10] There was a gap between the expectations the presentation created and how the model was actually used.

In 2024, Google paused image generation of people. Adjustments meant in part to account for diversity among the people depicted didn’t work as intended, and the model produced images that were historically inaccurate, among other problems.[11] Having the technology is one thing; shaping it into something people can use with confidence is hard. That problem surfaced early.

What changed the momentum was progress in AI that handles images and video.

Nano Banana, an image generation and editing model released in August 2025, raised Google’s profile.[12] In November that year Google launched Gemini 3, and at I/O 2026 it announced Gemini Omni, which includes video generation and editing.[13][14]

The ability to handle different kinds of information, such as text, images, audio and video, is called “multimodal.” Google’s progress in this area made people think, “Google really is strong after all.”

That momentum, though, can’t really be said to have become a runaway lead that lasts to this day.

On the Artificial Analysis image-editing leaderboard I checked on October 26, Nano Banana 2.1 is in the top group but not in first place.[15] In the video-generation rankings too, models from China’s Alibaba and ByteDance sit above Gemini Omni Flash 1.1.[16]

Rankings shift with evaluation methods and update dates, and they can’t settle who wins for every use. But this doesn’t look like a situation where we can relax and say “Google fully came back with Nano Banana” either.

The next contest: can you hand AI the work?

After multimodal, the areas that matter most in today’s AI race are coding and agents.

Coding means writing the programs that make apps and other software run. An AI agent is a system that takes a person’s instructions, works out the necessary steps, and carries out the work while using tools.

Put the two together and AI is getting to the point where you can hand it a whole bundle of work, writing code, fixing mistakes and checking that it runs, at a level that rivals a skilled software engineer.

In this area, Google still has work to do on adoption. In Stack Overflow’s 2026 survey, 65.5% of respondents named Claude Code as an AI coding tool they had used in the past year, 29.5% named OpenAI Codex, and 16.0% named Google Antigravity.[17]

While Google lags as a tool of choice among developers, there are also reports that teams inside Google, including DeepMind, Cloud and Android, are competing with one another to build in this space.[18]

The relationship between Antigravity and Android Studio shows that difficulty well.

Android Studio is not just a code editor for Android. It is expanding into a development environment where AI agents can take on design, implementation, testing and verification on the emulator. Antigravity, meanwhile, is an environment for handing a wider range of software development to AI.

The two aren’t direct rivals; you can, for example, run Antigravity inside Android Studio.

But the preview release of Android Studio now lets you choose Claude Agent and Codex as well as Antigravity.[19] From a developer’s point of view, “building with Google’s AI” still means deciding for yourself which product to start from and how much to hand over. Google’s lineup is rich, but it’s hard to say the division of roles is clear at a glance yet.

Agents that handle everyday tasks still have problems too.

In WIRED’s May 2026 hands-on with Gemini Spark, the agent got credit for planning a party from email and calendar data, but it couldn’t complete a restaurant reservation.[20] It’s a single trial, but it shows the gap between drafting a plan and actually getting the errand done.

Google’s user base remains huge. According to its April–June 2026 earnings call, the Gemini app has 950 million monthly users, and AI Mode, which lets you have a conversation with AI in Search, has passed 1 billion.[21]

Those numbers alone don’t tell us how many people are comfortable handing complex work to it, though. Remembering Google+, I want to look at both how many people use it and how they use it.

The futures shown at I/O don’t always reach everyday life

Google is good at showing off new technology in impressive ways, and its annual Google I/O event is the symbol of that.

But some of the ideas that brought the house down there didn’t take root in daily life as much as people expected.

Android@Home, unveiled in 2011, laid out a vision of connecting devices around the home, but it never caught on widely.[22]

In 2012, skydivers wearing the glasses-style device Google Glass streamed video from the sky to the venue.[23] It went on sale as the Explorer Edition, but sales of that early version ended in 2015.[24] The enterprise product that followed also stopped selling in 2023.[25]

At the same 2012 I/O, Google also announced a spherical device called the Nexus Q. It connected to TVs and speakers to play music and video, priced at $299, and preorders opened.[26] But shipping was delayed indefinitely, and it never reached the planned general release.[27]

Project Ara, which aimed to build a smartphone with swappable parts, showed plans for a developer device at I/O 2016 and was canceled later that same year.[28]

Of course, Google has also produced many successful products. Still, the power to wow an audience at a launch event doesn’t automatically become the power to build products people keep using.

Google Home and Nest: the test of fitting into daily life

There are also products Google has worked on for a long time with the aim of being used day after day, yet which still can’t be said to have reached that goal. Google Home and Nest, aimed at the smart home market, are among them.

In 2014, Google announced it would buy Nest Labs, maker of thermostats and smoke detectors, for $3.2 billion in cash.[29] By bringing in a company that made devices for the home, it tried to widen its touchpoints with daily life.

Two years later, at I/O 2016, it announced Google Home: a device for using Google’s voice assistant at home, where you could play music, set timers and control compatible lights and other devices just by speaking.[30]

But bringing the products and services together took time. In 2018 Google merged the Nest and Google Home teams, and in 2019 it folded the Google Home lineup into the Nest brand.[31][32]

This wasn’t just a change to the internal org chart. In the 2019 migration to Google Accounts, some of the ways Nest connected with other companies’ products could stop working, and Google itself advised affected users to hold off on migrating.[33]

For the company, it was a change to consolidate services. For the people using them, it became a question of whether what worked yesterday would still work tomorrow. Home equipment isn’t something you replace after every launch event.

And now Google is trying to rebuild its home voice AI with Gemini for Home. Where you used to have to give commands one at a time in fixed phrasing, the goal now is to follow the flow of a conversation and handle more complex requests.[34]

Google Home and Nest have products that have genuinely been useful, and people who use them. What their history shows is that problems remain to be solved even after a product reaches the home.

Add useful features while keeping what already works working. Protect the family’s information, and make sure people can take back control when something goes wrong. That kind of steady groundwork is what makes a tool people rely on at home.

Even with LLMs, turning research into products takes time

If conversation becomes natural, will these problems solve themselves?

One of the technologies behind ChatGPT and Gemini is the large language model, or LLM, which learns from huge amounts of text and other data and produces answers in words. Being able to grasp meaning even when the wording varies a bit, and to keep a conversation going, is a big step up from earlier voice assistants.

But replying naturally doesn’t mean it can operate appliances or a calendar correctly.

Say someone asks, “I have an early morning tomorrow, so wake me up earlier than usual.” What time should the alarm go off? Does the AI have permission to read the calendar? Is it confusing this person’s schedule with another family member’s? If the time is ambiguous, it needs to ask; if setting the alarm fails, it needs to say so instead of pretending it worked.

And if you’re handing over reservations or shopping, it needs to check with you before any charge is made. A good reply from the AI doesn’t finish the job. You have to combine it with a system that actually carries out the actions, one that catches mistakes, and one that lets a person stop it.

Google already runs services many people depend on, such as Search, email and home devices. Bringing new AI into them means keeping what already works and protecting personal information. These conditions help explain why turning research into products that many people can use is so hard.

That said, Google’s late start can’t be explained by caution alone. As we saw earlier, it also faces the challenge of pulling its own products and teams together. Beyond the ability to build the technology, what’s being tested is judgment: where to launch first, and how to turn user feedback into improvements.

While Google takes its time, users can get used to other companies’ AI and build their workflows around it. Even if Google catches up on performance, it will need further reasons to get people to change those habits.

Do Google’s strengths pay off for the people using it?

Even so, Google has strengths it can use for a comeback.

If AI helps inside tools people already use, such as Gmail, Maps, YouTube and Android, users are spared the effort of learning something from scratch. For example, if, within the scope you allow, it could sort out a travel itinerary from your email and check travel times in Maps, then the links between services would have real value.

If anything is going to make the difference from the Google+ era, it’s this: integration that actually saves users effort. If people can feel that, the touchpoints with existing services become a powerful weapon.

Owning the infrastructure that runs AI also gives Google room to lower costs. Whether that advantage shows up in pricing or ease of use is a separate question we’ll need to keep watching, though. Safety isn’t determined by the size of the company, either. What gives users peace of mind is being able to see what the AI read and what it did, and being able to stop it when needed.

Google DeepMind is also working on technology further out. A “world model” is AI that learns about space, how objects move, and what happens when an action is taken, and recreates that in a virtual environment. With Genie 3, announced in 2025, it showed research that generates virtual worlds a person can move around in while controlling them.[35]

Research like this could help with robot training, games and more in the future. But excellent research doesn’t turn straight into a useful product. Here too, the question is the distance between the I/O stage and everyday life.

A comeback will be decided after the announcements

When I look at Google’s AI, there are three things I want to check from here.

First, whether users keep coming back on their own. Not just trying a feature because it was added to an app, but returning because they feel “this really helps me.”

Second, whether the work you hand over gets finished. Not just suggesting a reservation, but confirming the conditions and completing it. And when it can’t, whether it can explain why and hand off to a person.

Third, whether it can be used with confidence over the long run. Whether, even as product names and systems change, the way people use it today is preserved and they can keep using it without confusion.

Google+ showed that tying something to existing services doesn’t create a habit. Glass and Nexus Q showed that drawing attention doesn’t guarantee a product will take root. And Home and Nest teach us that challenges with integration and continuity remain even after a product reaches people’s homes.

Google has the technology, the talent and the money. Past failures don’t decide the next one. But it is worth looking back at the places Google failed to reach despite all those strengths.

Beyond how impressive the next Gemini’s performance turns out to be, I want to see what errands the people using it actually got done. Only after watching those add up do I want to judge whether Google has made its comeback.

🦉 A young owl picks the wrong branch, flies off again, and slowly learns where home is. Maybe what Google needs, too, is not one flashy leap but a place people want to come back to every day.