In April 2026, Google announced a global rollout of Personal Intelligence, a feature allowing its Gemini AI assistant to access users’ Google Photos libraries for personalized image generation. The feature reached India on April 14, expanded to nearly 200 countries days later, and was formalized at Google I/O in May. It remains blocked in the European Economic Area, Switzerland, the United Kingdom, and Japan. The expansion marked the point at which a longstanding technical capability became an explicit product feature with user-facing opt-in prompts.
In this article
What Google Announced and When
On April 16, 2026, Google announced that Gemini could access users’ Photos libraries to generate personalized images. Google announced on April 14, 2026, that it was bringing Gemini’s Personal Intelligence feature to users in India. Days later, Google brought Personal Intelligence to Gemini users around the world, excluding Europe. At Google I/O 2026 in May, the company expanded Personal Intelligence in AI Mode to nearly 200 countries and territories across 98 languages with no subscription required.
The update, part of Google’s new Personal Intelligence initiative, allows the system to analyze images stored in Google Photos to generate highly personalized content. The company confirmed its latest update can scan all user photos to use actual images of people and their loved ones in AI image generation. A new AI model known as Nano Banana 2 enables Gemini to draw on users’ photo libraries to identify people, preferences, and past experiences, incorporating real images of friends and family into AI-generated visuals.
The announcement was not the introduction of photo scanning itself. Google Photos AI scanning has existed in basic form since 2015. What changed in April 2026 was the integration of that scanning into a cross-service AI assistant with image generation capabilities and explicit user prompts to connect accounts.
The Distinction Between Scanning and Access
Google Photos has always scanned uploaded images. The service has offered facial recognition, object identification, and searchable text extraction for years. Google scans photos for child sexual abuse material with trained specialist teams and automated technology. The 2015 implementation provided automatic organization and search functionality. None of this was new in April 2026.
What changed was the relationship between that scanning and other Google services. Personal Intelligence can access text, photos, and videos from Gmail, Calendar, Drive, Workspace apps, Google Photos, YouTube, Search, Maps, and more first-party apps. Personal Intelligence automatically pulls from Google Photos, Calendar, and Gmail to understand context about users’ inner circles. The feature does not introduce a new scanning capability. It connects an existing scanning infrastructure to a generative AI system that produces outputs based on that scan data.
Connecting Google Photos to Gemini’s Personal Intelligence is opt-in. Personal Intelligence is explicitly an opt-in feature, and users can choose which apps Gemini has access to at any time. Personal Intelligence is off by default, as users have the option to choose if and when they want to connect their Google apps to these services. The architecture presents a consent prompt to users who have not yet enabled the feature. Whether that prompt is surfaced at an appropriate moment, with sufficient clarity, in contexts where refusal is practical, is a different question.
What Google Says It Does Not Do
Google’s statements about Personal Intelligence distinguish between access for inference and use for training. Google says it does not train any generative AI models outside of Photos with users’ personal data, which includes other Gemini models and products. Gemini does not train directly on Gmail or Photos data; it references them to answer questions, but keeps the underlying personal content separate from model training. Google claims that while it does not train AI models directly on private photo libraries, it does train on limited information, such as particular prompts for Gemini and the AI response to them.
This is a technical distinction that matters. Training on user photos would mean incorporating those images into model weights, making them part of the system’s learned representations. Referencing photos for inference means the model processes them at query time to generate a response, without updating its parameters. The former creates a risk of data leakage through model outputs or memorization. The latter creates risks around access control, retention, and secondary use of inference logs.
Google says personal data in Photos is never used for ads, Ask Photos responses are not normally reviewed by humans, and generative AI models outside Google Photos are not trained on personal data in Photos. Gmail, Drive, and Photos content is not used for targeted ads by default, and Google’s privacy policy still says it does not use content from Gmail, Drive, or Photos to personalize ads by default. However, Google updated its privacy policy to clarify that it could use information users choose to share with certain Google services to personalize ads. The conditional matters: the policy creates a pathway, not a current practice applied to all users. The pathway depends on which services are connected and under what terms.
Geographic Exclusions as Signal
The feature remains disabled by default across the EU, UK, and Japan due to stricter privacy regulations. The global expansion excludes the European Economic Area, Switzerland, and the UK. The global rollout is live for Google AI Plus, Pro, and Ultra subscribers everywhere except the European Economic Area, Switzerland, and the UK.
This is not a technical limitation. Personal Intelligence works in the same languages and on the same infrastructure in jurisdictions where it is unavailable. The exclusion is regulatory. The General Data Protection Regulation in the EU, the UK GDPR, and Japan’s Act on the Protection of Personal Information impose constraints on automated processing of personal data, on the combination of datasets for purposes beyond the original collection, and on the legal bases required for such processing.
Google has chosen not to offer Personal Intelligence in those jurisdictions. That choice is not evidence that the feature violates those laws. It is evidence that Google has determined the compliance cost, litigation risk, or both are not worth the deployment. The decision reveals something about the feature’s privacy profile that statements about encryption and user control do not.
Email as Persistent Surveillance Infrastructure
The reaction to Personal Intelligence focused on photos because photos are emotionally salient and because the feature produces visible AI-generated outputs using recognizable faces. But the architecture applies equally to email. Personal Intelligence automatically pulls from Google Photos, Calendar, and Gmail. Gmail is not a photo storage service. It is a communication platform, and most of the information in a Gmail account was written by other people.
When a user enables Personal Intelligence for Gmail, they are granting an AI system access not only to their own correspondence but to everything anyone has ever sent them. The sender had no opportunity to consent to that access and no notification that it occurred. This is not unique to Personal Intelligence. Webmail providers have always had technical access to message content, and Gmail has scanned messages for various purposes since its launch. But the integration with a generative model that produces outputs based on multi-service correlation changes the threat model.
Consider a journalist who uses Gmail and who has received messages from confidential sources. If that journalist enables Personal Intelligence, Gemini can now correlate those messages with the journalist’s location history, calendar appointments, and photo metadata. It can infer relationships the journalist has not documented anywhere else. It can surface connections the journalist would prefer not to surface. The system is not designed to distinguish between convenient inferences and dangerous ones.
Email providers are infrastructural. For most professionals, changing email providers is not practical. Vendor lock-in is not just about technical migration cost. It is about the fact that an email address serves as professional identity, as the method of account recovery for every other service, and as the coordination point for everyone in a professional network. In that context, an opt-in prompt is not a free choice. It is a negotiation with an unequal party who controls access to an individual’s own archive.
Architectural Alternatives: Onion Mail and PQCServer
Email infrastructure that enforces encryption at the client layer, rather than relying on server-side policy commitments, operates on a different principle: the server cannot read mail because the server does not hold decryption keys. Communication over Tor, payment in cryptocurrency, and PGP encryption enforced at the client mean that access control is a matter of cryptographic capability, not policy. The server cannot scan what it cannot decrypt. There is no AI assistant to enable or disable because there is no plaintext on the server for an assistant to read.
This is not a feature comparison. Systems built on client-side encryption do not offer automatic photo organization, calendar integration, or contextual search because those features require server-side access to plaintext. The architectural choice is a trade-off: convenience of search and AI inference is exchanged for the assurance that no third party, including the service provider, can access messages or correlate them with other data.
Open-source post-quantum cryptography platforms such as PQCServer, released under AGPL-3.0, extend that principle to developers building their own systems. The code is auditable. The encryption is resistant to quantum attacks. The license requires that modifications remain open. These are tools for building infrastructure that does not depend on the service provider’s promises about what it will or will not do with access. The provider does not have access.
Trajectory
Personal Intelligence is not an isolated feature. It is part of a pattern in which AI capabilities are integrated into existing services, access to user data is consolidated across product boundaries, and consent is managed through interface design rather than law. The feature will expand. The model will improve. The geographic exclusions may narrow if Google determines that the regulatory environment has changed or that the revenue justifies the compliance cost.
The underlying question is not whether Google is scanning photos now. The question is what happens when the infrastructure for scanning, correlating, and generating outputs from everything an individual has ever written, photographed, or searched for becomes a default expectation. The infrastructure is already built. April 2026 was the month Google asked users to turn it on.