As of mid-2026, every major AI platform trains on user data by default. A study released in September 2026 analyzed the privacy policies of frontier AI developers and found that all six leading AI companies now use user inputs to improve their models unless users actively opt out — a shift that has happened gradually and without much announcement.
This is not limited to AI chat services. Photo apps, cloud storage services, note-taking tools, calendar apps, and email clients have all updated their terms to include language about AI training, personalization, or “service improvement” that involves machine learning on your content.
The result is that your personal data — your photos, your documents, your messages, your files — is being used to train or improve AI systems at multiple services you use daily, in most cases without you having made a conscious decision to allow it.
This guide is about finding those clauses, understanding what they mean, and opting out where it is possible.
Why the Standard “Read the Privacy Policy” Advice Does Not Work
Privacy policies are long. The average privacy policy for a major technology service runs between 4,000 and 8,000 words. They are written in legal language optimized to be defensible, not readable. Key disclosures are often embedded in subsections with neutral titles.
Nobody reads privacy policies, and everyone knows this. The policies are structured with this knowledge — the material information is present, but not findable without deliberate effort.
The practical alternative to reading every word is knowing what you are looking for and searching for it.
The Language to Search For
When you open a privacy policy, use your browser’s find function (Ctrl+F / Cmd+F) and search for these terms:
- “train” — the most direct word; catches “train our models,” “train AI,” “training data”
- “improve” — ubiquitous; catches “improve our services,” “improve our AI,” “improve our products”
- “machine learning” — often appears in explanations of how data is used
- “model” — catches “model training,” “language model,” “AI model”
- “develop” — catches “develop new products,” “develop AI features”
- “personalize” — sometimes the euphemism for AI training on personal content
Each of these will return multiple hits; not all are relevant. You are looking for the combination of the term with “your content,” “your data,” “your uploads,” “user data,” or similar phrasing.
Reading the Clause Once You Find It
Once you locate a relevant passage, there are four questions to answer:
1. What data is included?
The clause should specify what categories of data are used for training or improvement. Key distinctions:
-
Usage data vs. content: many services use behavioral data (what features you click, how often you open the app, how long you spend on various screens) for model improvement. This is different from using your actual content (your photos, your documents, your notes). The first is less privacy-sensitive than the second.
-
Aggregate vs. individual data: “aggregate and anonymized data” used to improve recommendations is very different from your specific content being used to train a language model. Aggregated, anonymized statistical data about usage patterns is a much smaller privacy concern than training on your personal files.
-
Future vs. historical data: some clauses apply only to new data going forward; others claim rights over data already in the system.
2. Is consent assumed or requested?
Look for how consent is framed:
- Opt-out default: you are enrolled automatically; you must actively disable the setting to stop data from being used. This is the most common structure.
- Opt-in required: the service must get your affirmative agreement before using data for this purpose. This is less common but provides better protection.
- No opt-out: the policy states that certain uses of data are conditions of using the service that cannot be opted out of. This is legal in most US states; EU/GDPR users generally have stronger rights.
3. Does an opt-out actually stop training, or just future use?
This is where the language gets slippery. Some services say opting out stops your data from being used to train future models, but data already used in training is not removed — the model retains the learning even after the opt-out. Others provide a meaningful forward-only opt-out that they represent as effective. The clause may not make this distinction clear.
4. Is there a difference between “improving the service” and “training AI”?
This is the most important and most obscured distinction. “Using your data to improve our service” is a phrase that can mean:
- Bug detection and performance monitoring (benign)
- A/B testing of features on aggregated user cohorts (low privacy concern)
- Training recommendation algorithms on your usage patterns (moderate concern)
- Training a large language model on your personal content (high concern)
All four can be truthfully described as “using data to improve the service.” The difference between them is enormous from a privacy standpoint.
How to Opt Out on Major Platforms
The settings locations for training opt-outs are not standardized. Each company has chosen a different location, different terminology, and a different level of visibility for the setting. The following are based on policies as of mid-2026.
Google’s AI training settings are distributed across multiple products. For Google Photos specifically, Gemini features have their own controls under Photos Settings → Google AI and Gemini features. For Google’s general AI improvement, the relevant setting is under My Google Activity → Other activity → Google AI improvements.
Google has maintained that personal photo content in Google Photos is not used to directly train its public AI models, but does use interactions to improve product-specific AI features.
OpenAI (ChatGPT)
Under Settings → Data Controls, ChatGPT has a toggle for “Improve the model for everyone.” Turning this off prevents your conversations from being used to train models. Note that this applies to chat content; API usage has separate terms.
If you are in the EU/EEA, you can also submit a data objection under GDPR Article 21 via OpenAI’s privacy request portal.
Meta (Facebook, Instagram, WhatsApp)
Meta’s AI training opt-out is in Account Center → Your information and permissions → Other AI information. Availability of this opt-out varies by region; EU users have stronger access due to GDPR. For photos specifically, this setting affects whether Meta’s AI systems can use your photos and posts for model training.
Apple
Apple has taken a more on-device approach than most competitors. For Siri and Apple Intelligence, Apple processes data in a way that is designed to minimize what leaves your device. Under Settings → Privacy & Security → Analytics & Improvements, you can turn off “Improve Siri & Dictation” and “Share Analytics.” These control whether Apple receives data about your usage for improvement purposes.
Dropbox
Dropbox has updated its terms to include AI features. Under Settings → Privacy, there is an option to manage how your content is used for “AI-powered features and services.” Dropbox has stated it does not use customer files to train foundation models but does use data to improve product-specific features like search and suggestions.
Microsoft (OneDrive, Copilot)
Microsoft’s data use settings for AI are under account.microsoft.com → Privacy → Privacy dashboard → AI. Copilot has separate opt-out settings from other Microsoft services. For personal OneDrive files, Microsoft states files are not used to train external-facing AI models, but Copilot’s personalization uses your activity and content to improve responses.
The Platforms Where Opting Out Is Not Fully Effective
In the US, there is no federal law requiring that opting out of AI training be technically complete. In practice, there are scenarios where opting out does not fully accomplish what you might expect:
- Model unlearning is not standard: if your data was used to train a model before you opted out, the model retains the learned patterns. Opt-out is forward-only.
- Third-party AI features: some apps embed third-party AI services (OpenAI’s API, Google’s Gemini API, Anthropic’s API) for in-product features. Your opt-out with the app does not affect how those third-party APIs process your data.
- Fine-tuning vs. base training: a service may distinguish between training its base model and fine-tuning a product-specific model on your data. Opt-outs may cover one but not the other.
The Strongest Protection Available
The most effective protection against your data being used for AI training is choosing services that do not do it, rather than relying on opt-outs.
When evaluating a service’s position on AI training, look for:
- Explicit language in the privacy policy or terms of service stating that user content is not used to train AI models (not just “we don’t sell your data” — this is a different claim)
- GDPR and CCPA compliance, which provide some enforceable opt-out rights
- Transparency reports or external audits of data practices
- A business model that does not depend on monetizing user data (subscription-only revenue, rather than ads or data sales)
A service that earns revenue through subscriptions has a different incentive structure than one earning revenue through advertising. The advertising model creates pressure to collect more behavioral data; the subscription model does not have the same structural incentive.
Checking What You Have Already Agreed To
Most people have installed dozens of apps and created dozens of accounts over the years, agreeing to terms of service with no real review. A practical audit of your exposure involves:
- Making a list of the apps and services that hold your most sensitive content — photos, personal documents, health information, private communications
- Going to each service’s privacy policy and searching for the training language described above
- Finding the opt-out settings for any service that allows them
- Deciding whether to continue using services whose terms are incompatible with your privacy preferences
This is not a quick process. But it is a concrete one, and it is the difference between being a passive subject of AI training and making an informed decision about it.
Your personal files — the photos, documents, and memories stored across services — represent a detailed record of your life. Who trains AI on that record, under what terms, and with what opt-out rights is worth understanding before the next update to someone’s privacy policy changes it again without announcement.