Google Assistant

• 1. Introduction

  ›

  Google Assistant had spread to more than 1 billion devices by 2020, only four years after its debut. It started small, folded into a messaging app and a speaker. It grew into something that listens in kitchens, cars and pockets across the world. Google had built an assistant before this one, but that older tool could only search, never really talk back. Somewhere in the years that followed, this new one gained an assigned personality and a celebrity voice. It also became a robot caller acting on its own, and the subject of a very public reckoning over what it quietly recorded. How did software built to sound more human end up facing regulators? And why did Google start pulling it back off phones, years after building it into a habit?

• 2. Origins

  ›

  On the 18th of May 2016, Google unveiled this software at its developer conference. The announcement came alongside a new smart speaker called Google Nest and a messaging app named Allo. Sundar Pichai, Google's CEO, described the goal as a conversational, two-way experience, calling it 'an ambient experience that extends across devices.' Later that same month, Google put Doodle leader Ryan Germick in charge of shaping its character. The company also hired former Pixar animator Emma Coats, brought on to give it 'a little more of a personality.' That personality still needed a body, and Google was about to decide exactly which phones would get to run it first.

• 3. Reaching New Devices

  ›

  Outside of Allo and Google Nest, this software was first locked to Google's own Pixel smartphones. In February 2017, Google opened access to other Android phones running versions 6.0.1 or 7.0, starting in a handful of English-speaking markets. Android tablets were left out of that rollout. Wear OS 2.0 picked it up too, with Android TV and Android Auto promised for later versions.

  In October 2017, the Google Pixelbook became the first laptop built with this assistant inside it, followed later by the Google Pixel Buds. By December 2017, Google said it would bring the assistant to phones running Android 5.0 through a Google Play Services update. It also promised support for tablets on Android 6.0 Marshmallow and 7.0 Nougat. Testing began quietly too. In February 2019, Google reportedly started trialing ads inside Assistant results.

  On the 15th of May 2017, Android Police reported that the assistant was coming to Apple's iOS as a standalone app. Google confirmed it two days later at its developer conference. The app reached iPhones that same month.

  At the Consumer Electronics Show in January 2018, Google announced the first Assistant-powered smart displays, screens built into speaker hardware. Lenovo, Sony, JBL and LG all showed models at the event, with the first one reaching buyers in July 2018. These devices could handle Google Duo video calls, play YouTube videos, pull up Google Maps directions, and show a Google Calendar agenda. They could also display footage from smart cameras, on top of everything already built for Google Home. They ran on Android Things paired with Google's own software.

  Google released its own smart display, called Google Nest Hub, in October 2018. A larger version, Google Nest Hub Max, followed later on a different underlying system. By then, this assistant had moved from a feature inside two products to something built into an entire hardware category. Screens were only the beginning of how far it would spread through other people's products.

• 4. Opening the Platform

  ›

  In December 2016, Google launched a developer platform called Actions on Google, giving outside developers a way to build their own apps for the assistant. By March 2017, new tools arrived specifically for building games. Actions on Google had started out limited to the Google Nest speaker. By May 2017, it reached Android and iOS too. Google also introduced a new directory where people could browse compatible products and services.

  First place in a Google-run developer competition won tickets to Google's 2018 developer conference, $10,000, and a tour of Google's campus. Second place and third place won $7,500 and $5,000, respectively, along with a Google Home. The contest was Google's way of pulling outside developers into building for the platform.

  In April 2017, Google released a software development kit, letting outside developers build their own hardware to run the assistant. It ended up wired into Raspberry Pi computers and cars from Audi and Volvo. It also reached home appliances such as fridges, washers and ovens, made by companies including iRobot, LG, General Electric and D-Link. Google updated that kit again in December 2017, adding features once reserved for Google Home speakers and the assistant's own smartphone apps. Outside device makers could now write their own Actions on Google commands and offer text-based interactions in more languages. Users could also set a precise location for sharper local answers.

  On the 2nd of May 2018, Google announced a program to invest in early-stage startups building on top of the assistant. The goal was to help developers create richer experiences for users. Some of those startups built entirely new hardware, others carved out niches in different industries. Not every voice people heard from the assistant, though, came from a startup. Some belonged to actors Google hired directly.

• 5. Finding a Voice

  ›

  Google Assistant launched using the voice of Kiki Baessell for its American female option. She was the same actress who had voiced Google Voice's voicemail system since 2010.

  At Google's I/O conference on the 8th of May 2018, six new voice options arrived, including one recorded by singer John Legend. This was possible because of WaveNet, a voice synthesizer built by DeepMind. WaveNet sharply cut the number of audio samples a voice actor needed to record. Legend's cameo voice did not last. Google discontinued it on the 23rd of March 2020.

  On the 11th of October 2019, Google added Issa Rae as an optional voice, unlocked by saying 'Okay, Google, talk like Issa.' That option did not last either. By April 2022, asking for it brought only a message saying the voice was no longer available, suggesting the listener try a different one instead.

  Voice work also shaped two of the assistant's audio games, Crystal Ball and Lucky Trivia. In April 2019, these games received the biggest voice changes in the application's history. Crystal Ball's British-accented voice slid downward on the word 'probably', and Lucky Trivia's host turned excitable like a game show announcer. A month later, a bug in Google's speech API stripped those studio-quality voices back out, though Google fixed it before May 2019 was over. Voice, it turned out, was only one part of what people expected this software to handle.

• 6. What It Could Do

  ›

  Unlike Google Now, the assistant it replaced, this software could hold an actual back and forth conversation, built on Google's natural language processing. Search results appeared as tappable cards rather than a plain list of links. Users spoke to it by voice, but could also type on a keyboard. They could ask it to schedule events and alarms, adjust device settings, or pull up information from their Google account.

  In February 2017, Google Home owners could shop entirely by voice through Google Express. Launch partners included Whole Foods Market, Costco, Walgreens, PetSmart and Bed Bath & Beyond. More retailers joined in the months that followed. Shopping lists used to live inside Google's notetaking app, GKeep. In April 2017, that feature moved to Google Express and the Google Home app instead, a shift that caused a severe loss of functionality.

  By May 2017, Google said the assistant would gain a keyboard for typed input. It would also identify objects and gather visual information through a device's camera, and support buying products and sending money. That keyboard let users see, edit or delete a history of past queries. The assistant warned against deleting, though, since it used old inputs to sharpen future answers. In November 2017, it gained the ability to name a song simply by listening to it play.

  Vocal shortcut commands let people trigger actions on Android and on the iPad or iPhone, or turn the assistant into a home automation hub. This feature was offered in English, among other languages. In July 2018, the Google Home version of the assistant learned to fire off multiple actions from a single spoken shortcut. The following month, it picked up bilingual support for languages it already handled. Reports also said a third default language could eventually work on Android phones.

  On the 12th of December 2019, Google added an interpreter mode to the assistant's phone apps for real-time translated conversation. That feature had previously been limited to Google Home speakers and displays. That work won Google Assistant the 2020 Webby Award for Best User Experience in the Apps, Mobile and Voice category. Its speech-to-text system could also recognize commas, question marks and periods inside transcription requests. On the 5th of March 2020, it gained the ability to read webpages aloud in 42 languages. By the 15th of October that year, it could also identify a song just from someone humming, whistling or singing it.

  By the time all of this had shipped, Google Assistant was answering more than 500 million people every month. It worked across over 90 countries and more than 30 languages. Not every one of its features, though, involved answering a question at all. Some of them involved the assistant picking up the phone and making a call on someone else's behalf.

• 7. The Duplex Experiment

  ›

  In May 2018, Google revealed Duplex, an extension of the assistant built to carry out full phone calls. It worked by mimicking a human voice, not unlike a robocall. Duplex could call a hair salon to book an appointment, set a restaurant reservation, or ring a business to confirm its holiday hours. None of this required a person on Google's side listening in. When it hit a situation it couldn't handle, it could signal a human operator to step in and finish the job.

  To sound convincing, Duplex used filler words like 'hmm' and 'uh', common phrases like 'mhm' and 'gotcha', and human-like pauses and intonation. Its first limited release came in late 2018, open only to Google Pixel owners. During that release, Pixel users in Atlanta, New York, Phoenix and San Francisco could use Duplex for exactly one thing: booking restaurant reservations. By October 2020, Google had expanded it to businesses across eight countries.

  One line of criticism warned about giving 'the company unprecedented access to human patterns and preferences that are crucial to the next phase of artificial intelligence.' Human operators, critics pointed out, might not even realize they were talking to software rather than a person, which some called deceptive. Privacy advocates also flagged that Duplex's calls were recorded so the system could analyze and refine its own responses. That process fed millions of collected vocal samples back into Google's algorithms.

  Google responded that disclosure would be built into the feature. In its own words, 'we are designing this feature with disclosure built-in, and we'll make sure the system is appropriately identified.' In certain jurisdictions, Google added, the assistant would tell the person on the other end that the call was being recorded. Whether that promise held up against everything users typed, said and asked at home was a separate question entirely, one journalists would soon start asking.

• 8. Reviews, Leaks and a Successor

  ›

  PC World's Mark Hachman gave the assistant a favorable review, writing that it was a 'step up on Cortana and Siri.' Digital Trends went further, calling it 'smarter than Google Now ever was.'

  In July 2019, Belgian public broadcaster VRT NWS reported that contractors paid by Google to transcribe audio clips had listened to sensitive material. That material included names, addresses and private conversations, picked up after the assistant mistakenly triggered on its wake word. Some of those mistaken recordings captured business calls, others captured bedroom conversations. Of more than 1,000 recordings VRT NWS reviewed, 153 had been captured without anyone saying 'OK Google' at all.

  Google acknowledged that 0.2% of recordings were being reviewed by language experts to improve its services. On the 1st of August 2019, Hamburg's Commissioner for Data Protection and Freedom of Information opened a procedure against Google. It aimed to block those reviews for three months, citing GDPR. A Google spokesperson said the company had paused 'language reviews' across Europe while it investigated the leaks.

  In October 2023, at the Pixel 8 event, Google unveiled a mobile version of its Gemini chatbot. It was first shown under the name Assistant with Bard, and then simply Bard. That chatbot launched in the United States on the 8th of February 2024. It was positioned to take over as Android's main assistant, while the original stayed on as an optional extra.

  On the 14th of March 2025, Google announced that the original assistant would stop working on Android 9 and newer devices. It would also stop on iOS and on hardware like Google Nest. Gemini would take over most of its duties. The original would keep running only on lower-end phones, on devices still using Android 8.1 'Oreo' or earlier, and on Android Go, the stripped-down edition Google built for its most limited hardware.