Google Trains Its AI on What You Upload

Introduction

Google says it doesn't train its AI on your Gmail or your Google Photos library. That part is true. Its own help pages say so, in plain language. It's also beside the point, because it isn't the setting Google switched on for most people in June. That one is called Save Media, it defaults to on, and it has been feeding the photos you send through Google Lens, the audio from voice search and Translate, your Search Live recordings, and the files you upload into Google's AI-training pipeline, held for up to four years even if you delete your account. Google didn't announce it with a press release. It sent an email.

Google Split One Setting Into Three

In June, Google took the old "Web & App Activity" control and broke it into three independent settings: Search Services History, Personalized Recommendations, and Save Media. If your Web & App Activity was already switched on, Search Services History inherited that "on" status, and Save Media rode along with it. What Save Media captures is spelled out on Google's own Search Help page: images you send through Lens, audio from voice search and Translate's speaking practice, Search Live recordings, and files you upload. Google's description of what happens to it is direct: the saved media "is also used to develop and improve Google services and technologies, including AI models and safety measures."

Nobody got an opt-in prompt. There was no consent screen, no "yes, use my uploads." The disclosure was a customer email in June, and the setting was already on when it landed. TechCrunch's Sarah Perez surfaced the change publicly on July 6, calling it an "under-the-radar update," roughly a month after Google sent the email most people never opened.

This is where a lot of the coverage blurs two clocks. July 30 is when Google's rewritten Terms of Service take effect, and that document does list "emails you send and receive through Gmail" and "pictures you share with friends through Photos" as examples of "your content" under a broad license. That language is carried over word for word from the terms that took effect in May 2024, and July 30 doesn't hand Google new access to Gmail or Photos. It's still worth treating as a deadline, though: the prompt to go check the settings the June change quietly rearranged.

What the Help Pages Actually Say

Start with the carve-out, because it's real. Google's Photos Help page states that "Google does not train generative AI models outside of Google Photos directly on your imagery and audio from your Google Photos library," and its Workspace documentation says the Gemini features in Gmail and Docs don't train its general models on your content. When Google denies training on your inbox or your photo album, it's citing published policy, not spin. The viral version of this story keeps getting that part wrong. A claim that Google was secretly scanning Android photo libraries went around last December, and Forbes traced it to an evidence-free social post from Proton, a privacy competitor.

The Gmail claim got as far as a courtroom. Thomas Thele alleged in November that Google had quietly turned on Gemini "smart features" across Gmail, Chat, and Meet, letting its AI read private messages without consent. A federal judge dismissed the case on July 7 for failing to show concrete harm, giving the plaintiffs 21 days to amend. That's a dismissal on standing, not a finding on whether the allegation was true.

The pipeline that is actually running sits one settings page over, and Google documents it just as plainly. The Search Help page on Save Media states that media already selected for training can be kept "for up to 4 years," and that window doesn't reset when you delete your account. Switch Save Media off, Google says, and future media "will not be used to train Google's generative AI models, unless you provide feedback." That four-year retention is the part that stuck with me. You don't build a window that long for a feature you consider a casual convenience.

Google's rewritten terms matter less for what changes on July 30 than for the room they leave open. The content license lets Google run "automated systems and algorithms to analyze your content" for purposes including "developing new technologies and services for Google," a phrase broad enough to cover AI uses Google says today it doesn't pursue. The genuinely new July 30 additions sit elsewhere in the document: background internet use, AI-output disclaimers, tighter anti-scraping terms, and data Google derives from your content. The Gmail and Photos carve-outs are product policies, and those can change without asking you to sign anything.

Who Benefits

The beneficiary is Google, specifically the Search and Gemini teams racing OpenAI, Anthropic, and Meta for something that turned scarce: real human data. Web-scraped text, the traditional fuel for these models, is finite and increasingly tangled in copyright fights, and the open pages it came from are filling with AI-generated filler. Data that people create through ordinary product use is the opposite: fresh, and free at the margin once the default is set to capture it. Every photo you run through Lens and every file you upload becomes proprietary training material a competitor can't scrape from outside Google's walls.

The four-year retention window makes this a structured, long-horizon supply line, not a throwaway convenience feature, and the scale it feeds is enormous. At its I/O 2026 keynote, Google said its models were processing 19 billion tokens a minute, seven times the year before. The Terms of Service are the cover. The narrow, true "we don't train on Gmail or Photos" line gives Google a privacy-friendly headline to point at, while the license reserves the standing to widen AI use of your content later, with no fresh consent moment required.

The Same Move, at Every AI Lab

Every major AI lab runs a version of this. Anthropic, the lab that built its brand on being the cautious, privacy-minded one, did the same thing in August 2025. It told consumer Claude users to choose whether their chats would train future models, set a deadline of October 8, and gave the people who opted in five-year data retention against 30 days for those who declined. OpenAI's consumer tier runs on the same logic. Across the industry the default is now consistent: free and consumer accounts train the models unless you opt out, while the enterprise tiers that pay in cash default to training-off. You pay with money or you pay with data.

Whether that default is even legal depends on where you live. Two days after the TechCrunch story, the European Data Protection Board adopted its first pan-EU guidelines on using personal data to train generative AI, giving every national regulator a shared rulebook 22 days before Google's terms took effect. In Europe that's a sharpening enforcement floor; in the US, the equivalent was a customer email. The FTC that would police this here is down to two commissioners after the Trump administration removed both Democrats in March 2025, and its July 7 proposed policy statement aimed Section 5 enforcement at AI "accuracy" and political bias while signaling it would treat conflicting state privacy laws as preempted, not at data-training consent.

None of this is hypothetical for Google in particular. In September 2025 a San Francisco jury hit Google with a $425.7 million verdict for collecting data from nearly 100 million people after they had switched Web & App Activity off. That is the exact control Google has now re-architected into Search Services History, with Save Media riding along, on by default. Texas settled its own case against Google for $1.375 billion over location and biometric data, and Google paid $68 million in January to settle claims that Assistant recorded people after mishearing "Hey Google" and passed the clips to contractors to sharpen its voice AI. Google renamed the setting at the center of that $425.7 million verdict instead of retiring it, then reset it to gather more.

How to Turn Save Media Off

If you want out, the switch is real and it takes about a minute. Open your Search Services History settings at myactivity.google.com, find Save Media, and turn it off. Because Google is rolling the new controls out gradually, some accounts won't see the option yet, so it's worth checking again around July 30. Turning it off stops future uploads from feeding the models, but it doesn't purge what's already in the pipeline: Google's four-year retention applies to media it has already selected, so switching off is forward-looking, not a delete button. While you're there, you can set Search Services History to auto-delete at 3, 18, or 36 months, and review Web & App Activity and Personalized Recommendations, now their own separate controls. If the Gmail question is what sent you looking, that's a different toggle under Gmail Settings, and Google's position is that those smart features don't train Gemini in the first place.

The Bottom Line

The honest version of this is less satisfying than the viral one, and more useful. Google collects the media and files you actively hand to its search tools, not your Gmail inbox, keeps that setting on by default, and holds the right to train on it for four years. July 30 didn't start the training and won't stop it. It's the calendar reminder that makes right now a good time to open a page you've probably never seen.

The fix is a single toggle, tucked under a control Google renamed and switched on for you, disclosed in an email you likely archived without reading. Google Search has more than three billion users. How many of them do you think ever found it?