Every major cloud service now has an AI training clause in its terms of service. Most also have an opt-out. The opt-out is usually buried in settings, takes multiple clicks to find, and is presented with enough ambiguity that a significant percentage of users who believe they’ve opted out have not, in fact, done so effectively.
Understanding why requires understanding what the training pipeline actually looks like — not at the policy level, but at the technical level. Once you see the mechanics, the gap between “opt-out of AI training” and “your data is not used for AI training” becomes much clearer.
The AI Training Pipeline in Simple Terms
Building a modern AI model — the kind that powers “search your photos by what’s in them” or “summarize this document” — involves three broad stages:
1. Data collection: Gathering a large corpus of examples that represent the task the model should learn. For a photo search model, this means large collections of photos with associated labels or descriptions. For a document summary model, this means collections of documents with paired summaries.
2. Training: Using a process called stochastic gradient descent (or similar optimization algorithms), the model learns patterns in the data by repeatedly adjusting its internal parameters to minimize prediction errors. This is computationally expensive and happens on specialized hardware (typically large GPU clusters). Each training run consumes the dataset repeatedly across many “epochs.”
3. Inference: The trained model is deployed to answer real-time queries: “what’s in this photo?” or “summarize this document.” Inference is much cheaper computationally than training, and it’s what runs when you use AI features in a cloud service.
Your personal files are relevant primarily to stages 1 and 2, though stage 3 has its own privacy implications.
How Personal Files Enter the Training Pipeline
When a cloud service says it “uses your data to improve the service,” this typically means one or more of the following:
Direct inclusion in training sets: Files are selected from the corpus of user-uploaded content, labeled (either by human reviewers or by another AI system), and included in a training dataset. This is the most direct form of use.
Implicit feedback signals: When you search for “beach vacation 2022” in your photo library and click on a photo that’s actually from a 2021 beach trip, that interaction is a training signal. You’ve told the system that this photo matches (or doesn’t match) a search query. These implicit signals train “relevance models” that improve search quality. They operate even when you haven’t explicitly been shown a training opt-in.
Fine-tuning on user populations: A base model trained on a large public dataset is subsequently fine-tuned on the user base’s data to improve performance for the specific patterns of that user population — the types of photos people typically store, the kinds of documents they scan. This fine-tuning uses aggregate patterns from user data without necessarily processing individual files as clearly labeled training examples.
Federated learning: Some services use federated learning, where a model is trained on a device using local data, and only the resulting model update (not the raw data) is uploaded to central servers. This preserves more privacy than direct data collection, but still uses your personal files as training inputs on your device. Apple has used this approach for some features. “Your data doesn’t leave your device” is technically true in federated learning, but your files are still used for model training.
What Opt-Out Actually Prevents
When you find the opt-out setting for AI training in Google Photos, Dropbox, or another major service, what you’re typically preventing is:
Direct inclusion in labeled training sets: The clearest form of use. Your files won’t be manually reviewed, labeled, and added to a training dataset. This is meaningful.
Some forms of explicit feedback collection: Your click patterns and interaction signals may be excluded from certain training pipelines.
What opt-out typically does not prevent:
Implicit learning from access patterns: The fact that you store a lot of beach photos, that you search for certain types of content, that you tend to organize files in a particular way — these behavioral signals are often processed separately from the content of your files and may fall outside the scope of a content-opt-out.
Federated learning on device: Many opt-outs cover cloud-based training but not on-device learning. If a model update happens locally using your photos as training inputs, and the setting description covers “your photos being sent to our servers for training” — you’ve technically opted out of what was described, while on-device training continues.
Processing for “service improvement” broadly defined: “Improving the service” is broader than “training AI models.” Processing your files to improve caching algorithms, security monitoring, compression techniques, or content delivery isn’t AI training in a strict sense — but it involves analyzing your content.
Historical processing before opt-out: If processing for training happened before you opted out, that data doesn’t get removed from training sets retroactively. The opt-out prevents future use, not remediation of past use.
The Difference Between “Doesn’t” and “Won’t”
There’s an important linguistic distinction in how AI training commitments are stated.
“We don’t use your data for AI training” is a present-tense factual claim. If the terms say this without qualification, it means the current system architecture doesn’t include your data in training pipelines.
“We won’t use your data for AI training without your consent” is a policy commitment about future behavior. It’s less absolute — it means the current behavior is opt-in, but the architecture could change if you consent (or if the opt-in mechanism is changed).
“We may use your data to improve our services” is a permissive clause that covers AI training without naming it explicitly. This language is present in almost every major cloud service’s terms.
“Your data is never used to train AI models” — note the universal negative — is the strongest available commitment and is rare. Very few commercial cloud services make this claim.
When evaluating a cloud service’s AI training practices, the specific language in the terms matters more than the marketing claims on the homepage.
What the Technical Architecture Looks Like From the Outside
There are a few indirect signals that tell you something about how a service’s AI training pipeline is structured:
On-device vs. cloud processing: A service that does AI processing on-device (like Apple’s on-device ML for some Photos features) has a different training architecture than one where all processing happens in the cloud. On-device processing doesn’t necessarily mean no training use, but it changes the data flow.
The business model: A service funded by advertising has a stronger commercial incentive to use user data for model training (better ad targeting, better user engagement) than a subscription service that gets paid directly by users. The incentive structure doesn’t determine what the service does, but it makes it easier to evaluate stated commitments.
Transparency reports and independent audits: Services that publish detailed transparency reports about data use, that have undergone independent privacy audits, and that publish verifiable technical documentation about their data pipelines are making it easier to evaluate their commitments. Vague “we care about privacy” language on a website is not the same as auditable documentation.
The scope of the training opt-out: If a service’s opt-out is very specific (“don’t include my photos in labeled datasets for visual search training”) and leaves significant carve-outs (“we may still use aggregate patterns”), that specificity reveals the shape of the training pipeline. A more absolute opt-out (“none of your data is used for any model training”) reflects a different architecture.
Why a Genuine No-Training Commitment Is Structurally Different
For a cloud service to genuinely not use user files for AI training — not just permit opt-out, but maintain an architecture where user files never enter training pipelines — requires deliberate design choices:
- No labeled dataset collection from user content
- No implicit feedback loops that feed back into model training
- No federated learning component that uses user devices for model updates
- No “model improvement” sharing with third-party AI providers that process file content
This isn’t impossible. It’s a business model choice. Services that are subscription-funded, that don’t operate advertising businesses, and that make privacy a core differentiator have structural incentives to maintain these choices — because their revenue depends on trust, not on data.
daftei is built around these principles. Files stored on daftei — photos, documents, voice recordings, anything — are not used to train AI models. Not with opt-out: never, by design. The technical architecture doesn’t include a training pipeline, and the business model (subscription-funded, no advertising, no data licensing) removes the incentive to build one.
Reading the Terms Yourself
No blog post is a substitute for reading the specific terms of a service you’re currently using. Key things to look for:
-
Search for “train” and “improve” in the privacy policy and terms of service. What activities are covered by each term?
-
Look for the opt-out mechanism and read it precisely. What does it cover, and what are the carve-outs?
-
Read the effective date and change notice policy. Terms can change. A strong commitment made today can be narrowed tomorrow with 30 days’ notice.
-
Check whether the policy distinguishes between content and metadata. A policy that protects photo content might still permit training on access patterns, search queries, and behavioral signals derived from how you use the service.
-
Look for third-party AI provider mentions. Some services outsource AI features to third-party providers (OpenAI, Anthropic, Google, AWS). Data sent to those providers may be governed by their terms, not the cloud storage service’s terms.
The pipeline is more complex than the checkbox suggests. Understanding it helps you make a real decision about where your most personal files live.