On September 28, 404 Media reported—based on internal contractor documents—that people evaluating some Microsoft Copilot image-editing results could see the user’s prompt, the uploaded image and generated alternatives. Microsoft’s current documentation for the updated consumer Copilot app separately says prompts, responses and file contents are not used to train foundation models. Those facts are not interchangeable: human evaluation and foundation-model training are different data uses.
A photo upload is not just a button press
Imagine uploading a family portrait because you want to remove a stranger in the background. From the user’s point of view, the workflow is simple: choose a file, type an instruction, download the result. Behind the interface, however, a cloud service may need to receive the original image, process it, store related activity and run quality or safety systems. Depending on the product and workflow, people may also be involved in evaluation.
That last point is what made the September 28 report notable. 404 Media said contractor instructions for certain Copilot image-editing evaluations showed reviewers the original request, the uploaded photograph and two generated edits so they could judge the results. The report describes a specific evaluation workflow; it does not establish that every Copilot image is manually viewed, or that every Microsoft product handles photos the same way.
Ask separately whether the image leaves your device, how long related activity is kept, whether people can review selected interactions, and which exact product or account policy applies.
Training, human review and storage are three separate questions
These terms are often collapsed into one sentence, which is where confusion starts. Training means using data to update or build a model. Human review means a person sees selected material for purposes such as evaluation, product quality or safety. Retention is about how long data or activity remains stored. A service can say one thing about training and something different about review or storage.
Microsoft’s current documentation for the updated Copilot app, available since August 18, states that prompts, responses and file contents in that app are not used to train foundation models. Microsoft also separates the consumer Copilot app from Copilot inside Microsoft 365 and from work or school accounts. That is important because a privacy statement for one surface should not be copied over to another.
The public material we checked does not support a blanket claim that every current Copilot upload is reviewed by a person. The defensible reading is narrower: 404 Media documented a human-evaluation workflow involving uploaded images, while Microsoft’s current product documentation describes separate rules for the updated app and its account types. For users, the practical lesson is to identify the exact service before relying on a short privacy slogan.
| Question | What to look for | Why it matters |
|---|---|---|
| Does the photo leave my device? | Cloud upload or on-device processing | An upload expands the systems that must handle the file. |
| Is it used to train a model? | Current first-party policy for that product/account | Training is a distinct use of submitted data. |
| Can people review it? | Quality, safety, feedback or evaluation language | Human access is a separate privacy consideration. |
| How long is it kept? | History, retention and deletion controls | Storage continues after the visible edit may be finished. |
A photo can reveal more than the subject you meant to edit
An image can expose information that is difficult to notice while you are focused on the edit itself: a child’s face, a client, a delivery label, a school badge, a computer screen, a private interior, an unpublished product, or a document in the background. Metadata can add capture details as well. Removing metadata may reduce some information, but it does not change the fact that an uploaded image has been sent to another service.
This matters especially to photographers and small businesses. A client may have agreed to a portrait session without expecting the original photograph to be sent to a third-party AI editor. A product photo may be commercially sensitive before launch. A family may be comfortable sharing a finished image publicly but not the unedited original. The right question is therefore not “Is AI safe?” but “Does this particular task require this particular file to leave my device?”
A practical five-minute check before uploading
- Name the exact tool and account type. Consumer, business and workplace products can follow different rules.
- Read the current first-party privacy page. Check the date and product scope; old help pages may apply to older versions.
- Separate training from review. A training opt-out should not be assumed to answer every question about quality or safety review.
- Look at the photograph itself. If it contains children, customers, addresses, documents or unpublished work, the cost of an unnecessary upload is higher.
- Ask whether this step needs the cloud at all. Generative replacement may; adding a watermark before sharing may not.
With SIGNA, add a visible watermark and an invisible SIGNA signal on your Android device before sharing. This step does not require uploading the selected photo to SIGNA AI Lab.
After the edit, decide how the sharing copy should leave your device
SIGNA does not replace a generative photo editor. If you need to invent a new background or generate pixels, a cloud AI editor may be the tool you choose. SIGNA addresses a different moment: preparing a supported image before you share it.
On Android, SIGNA can add a visible watermark and an invisible SIGNA signal using on-device processing for the supported photo workflow. Later, you can inspect a supported file in SIGNA to see whether that signal is observable. For that watermark-and-signal step, the photo does not need to be uploaded to SIGNA AI Lab for processing.
That is the natural privacy connection. Local processing does not make other apps private, and it cannot control what happens after you deliberately upload an image elsewhere. It simply removes an unnecessary upload from the SIGNA step. When the job is “prepare this photo before it leaves my device,” reducing the number of services that need the original file is a concrete benefit.
1) Check what the photo reveals. 2) Decide whether the edit really needs a cloud service. 3) If you are preparing the image for sharing, add the visible watermark and invisible SIGNA signal locally. 4) Share the prepared copy rather than treating every tool as one giant workflow.
Privacy becomes part of the photo workflow
That framing is too broad and too sensational. The more useful change is that generative photo editing has made uploading personal images feel routine. As those tools become ordinary, users need to distinguish the visible feature from the data path behind it. A fast edit can still involve an original photograph that is more sensitive than the final result.
For creators, families and small studios, privacy becomes easier to manage when each step has a clear boundary: edit in the cloud when the task genuinely needs it; process locally when it does not; and check the current policy of the exact service you are about to trust with the original.
Primary sources
404 Media — Humans Are Reading Copilot Prompts — And They’re Horrified (Sep. 28, 2026)
Microsoft — Copilot for individuals: data, privacy and responsible AI
Microsoft — Copilot for individuals: activity history
Microsoft — Copilot for individuals: privacy controls and choices
Prepare the photo locally before you share it
Before sharing, use SIGNA to add a visible watermark and an invisible SIGNA signal on your Android device. Later, check a supported file to see whether the signal is observable.
