Gemini Spark can now manage a user's Google Photos library
The assistant will organise, sort and act on a personal image collection. It is the most personal archive most people have, and the agent now has write access to it.
Google has extended Gemini Spark to manage a user's Google Photos library, letting the assistant organise and act on a personal image collection, TechCrunch reported on 4 September.
Photo libraries are among the most sensitive archives a person keeps, and among the least curated. They contain medical images, documents photographed for convenience, screenshots of private conversations, pictures of children, and the record of relationships that have since ended — accumulated over a decade with no expectation that anything would read them systematically.
Why the capability is genuinely wanted
Nobody organises their photos. The library grows to tens of thousands of images, search works if you remember roughly what you are looking for, and the maintenance work — deleting duplicates, sorting an event, finding every picture of a person for an album — is tedious enough that it never happens.
That is a good fit for an agent. It is high-volume, low-stakes, well-specified work of exactly the kind people will not do themselves, and Google already has the classification infrastructure to support it.
The write access is the change
Gemini has been able to search photos for some time. Managing them means acting: moving, grouping, deleting.
Deletion is the part worth pausing on, because photo libraries are the archive people are least able to reconstruct. An agent that removes what it assesses as duplicates or clutter is making judgements about material with sentimental value it cannot evaluate — a blurred photograph of someone who has since died is not a low-quality duplicate.
Google has not published what actions require confirmation, whether deletions are recoverable and for how long, or how a user reviews what the agent did.
The risk this connects to
The fortnight's security reporting makes the shape of the concern concrete rather than abstract.
Simon Willison noted on 30 August that ChatGPT Work closes all three sides of the prompt injection "lethal trifecta": access to private data, ingestion of untrusted content, and a path to act. An agent with write access to a photo library has the first and third. Whether it has the second depends on whether anything it processes can carry instructions — and images can contain text, including text placed there deliberately.
The rest of the month supplies the surrounding evidence: an actively exploited authentication bypass in LiteLLM's MCP handling, infostealers draining Claude accounts by lifting session tokens, and an OpenAI agent that reached production systems at Hugging Face.
The other thing an agent needs
Managing a library well requires understanding its contents, which means classification across the whole archive rather than on the images a user searches for.
Google has not said whether that processing is on-device or server-side, what is retained, or whether any of it informs anything beyond the user's own library. Those questions matter more here than for most features, because the corpus is a decade of one household's private life.
The EU designated ChatGPT a very large online platform on 31 August, with obligations including researcher data access. Google's assistant products are not covered by that designation, and the Commission has not said whether it is considering one.
Google also had a bad week for assistant judgement: hikers had to be rescued after planning a route with Gemini.
Runs the newsroom. Rename this profile in the studio to your own byline.
Related
Every weekday, the AI stories that moved money or shipped code.
No cross-posting, unsubscribe anytime. See all newsletters