Illustration by Matteo Giuseppe Pani / The Atlantic.
In May, I asked Google’s chatbot, Gemini, to write a birthday letter to my best friend. Within seconds, it spat out the most impressive piece of AI writing I have ever encountered. Instead of reading as soulless, machine-generated text, the letter felt unnervingly like something I might’ve actually written. “You’re probably rolling your eyes,” the letter read, after a sentence that my friend would most definitely have rolled his eyes at. All I had typed into the chatbot was a nine-word prompt containing my friend’s first name and the age he was turning. But the letter referenced real moments from our friendship. One paragraph recounted a conversation we had shared on the eve of college graduation; another reflected on a challenging period we had navigated together. Gemini had even included his correct birth date.
I hadn’t planned to let AI write the birthday letter for me. When I opened Google Drive to type it up myself, Gemini popped up and volunteered to help out. Since the spring, when I first signed up for a free trial of Google’s AI Pro subscription—normally $20 a month—Gemini has followed me around the Googleverse. The tool is akin to a souped-up version of Microsoft Clippy: In Gmail, it offers to summarize long threads and draft entire messages. In Sheets, it volunteers to assist with data analysis, generating colorful bar graphs at the click of a button. But Gemini has proved most alluring in Drive, where the chatbot can automatically find and consult relevant files before generating text. That’s how Gemini was able to whip up such a good birthday letter: It already knew a lot about me (and, by association, my friend).
Of all the things that chatbots excel at, they have generally not been very reliable for individualized tasks. Ask an AI tool to write an essay on, say, the history of popcorn, and you will likely get a decent response. But ask it to write a speech for your sister’s wedding, and the result will probably be quite poor. You might get a better speech if you feed the chatbot a decade of your texts and emails, her wedding website, and previous toasts you’ve given for other loved ones. But that process takes time and effort, which most people don’t put in.
Tech executives dream instead of hyper-personalized chatbots that automatically have access to all of the information they might ever need. After sucking up the web to build models capable of generating coherent text, AI companies are now mining our personal troves of data to teach chatbots everything there is to know about us. Google, with its colossal data empire in tow, is particularly well positioned to lead the way. If OpenAI introduced us to the Hallmark-card version of AI writing, Google is ushering in a new chapter where chatbots are capable of drafting the sort of intimate letters you might write to your best friend.
The birthday letter was just the start. Not only could Gemini write fairly convincingly in my voice; the chatbot, as I quickly learned, was teeming with my personal information. When asked, it accurately described my financial goals, my vaccination history, and my parents’ physical appearances. To test the limits of how much Google knew about me, I told the chatbot to make a CIA dossier. The first section (“IDENTIFYING INFORMATION”) listed my full name, email address, and current location. Not too crazy. Section two (“RELATIONSHIPS & PERSONAL HISTORY”) accurately described the details of both a long-term romantic relationship and a brief high-school fling. By section three (“PSYCHOLOGICAL PROFILE”), the chatbot was dissecting my communication style and emotional intelligence. And in section four (“POTENTIAL VULNERABILITIES”), Gemini had outlined my travel history, citing the time I had spent abroad as an exchange student, and diagnosed me as an overthinker.
Not everything in the dossier was accurate. Gemini struggled to disentangle fact from fiction, occasionally confusing details from short stories I’ve written with real-life anecdotes. When I later asked the chatbot if it knew my birthday, it told me I was born in 2010 (wrong, though it got the date right on a second try). Even though the birthday letter was startlingly good, Gemini occasionally slipped into a more generic chatbot register—at one point, it described the future as “everything shimmering in the distance.”
Still, Gemini knows me much better than other chatbots do. When I asked ChatGPT to create a CIA dossier, it failed miserably: The bot overinterpreted my prompt, explaining that a key part of my personality was my “taste for espionage tropes.” The other details it added were vague and unimpressive. There’s a clear reason for the discrepancy. Unlike Google, OpenAI doesn’t have half my lifetime’s worth of my data stored up. In Gmail, I have more than 200,000 emails, amounting to 30 gigabytes, some of which date back to elementary school. My Drive contains another 45 gigabytes of files, such as chemistry study guides and travel itineraries, half-written poems and unsent love letters, budgeting spreadsheets and New Year’s resolutions, insurance appeals and symptom trackers.
Even if you don’t spend your free time soliloquizing in Google Docs like I sometimes do, the search giant likely knows enough about you to train your own custom chatbot. Our emails, files, and browsing histories are all already at the company’s fingertips. Chrome is the most popular browser in the world; almost one-third of the planet’s emails are sent with Gmail; and Google’s productivity apps have billions of users who store files across Drive, Docs, Sheets, and Slides. That’s to say nothing of Maps, YouTube, or the entire Android ecosystem.
Google knows it’s sitting on a gold mine. In May, at the company’s annual software conference, the Gemini team lead Josh Woodward said Google’s goal is to make the chatbot the most “personal” and “proactive” AI assistant around. He offered education as an example. College students are flocking to ChatGPT, but those same students do much of their work using Google software such as Docs and Slides. “Imagine you’re a student; you’ve got a big physics exam looming,” Woodward said. Gemini might see the test on your calendar a week out and send you “personalized quizzes” based on the readings and lecture notes you’ve already stored in Google Drive. There are countless other ways you might use such personalized AI. When I asked Gemini to write me a cover letter, it automatically consulted several I had previously written. When I prompted Gemini to make me a summer-reading list, it first combed through email exchanges with high-school and college instructors, a list of my favorite books, and two editions of a weekly newsletter I subscribe to.
Google is not the only company pushing forward with bespoke AI. Sam Altman recently described the “platonic ideal state” for ChatGPT as a model with access to “your whole life.” This chatbot would ingest every piece of information you had ever produced or encountered—including the books you had read, emails you had sent and received, and even conversations you’d had with your friends and family. With the explicit goal of making ChatGPT more personalized, OpenAI recently upgraded the chatbot’s “memory” feature, such that the bot is now able to reference all of a user’s past conversations.
But building up that data will take time. Legacy tech firms such as Apple and Microsoft do already have plenty of data to draw on, but Google is further ahead in its consumer AI efforts. Then there’s Meta: The company’s stand-alone AI app, which launched this spring, encourages users to link the assistant to their Facebook and Instagram accounts for “an even stronger personalized experience.” Facebook comments and Instagram DMs, however, are simply less meaty than email exchanges and PDF documents.
Google has faced a bumpy road since generative AI exploded a few years ago. The technology has presented the biggest threat yet to Google’s search business, and the company’s share of the market recently dropped to its lowest in a decade. At the same time, usage of Google’s AI tools has skyrocketed over the past year, and the company recently rolled out a new AI search mode in an attempt to steal search queries back from the likes of ChatGPT. Now, with the company’s personalization advantage, Google could surge ahead.
Whether Google or another company gets there first, this new era of AI is coming. For years, we have been shedding information online through clicks and likes, photographs and files, emails and search queries. That digital exhaust is now getting a second life. Already, it can be difficult to figure out whether text that you encounter online is generated by AI. Soon, while looking back on old emails, you might even feel that way about your own writing.