Can An AI Feedback Preview Omit Connector Content That Is Still Submitted?
An AI feedback preview can omit connector-derived content that is still submitted. Compare Microsoft, Google, Anthropic, and OpenAI feedback boundaries.
Yes. An AI feedback preview can omit connector-derived content that is still submitted with your report. Microsoft says it shows preview links when possible but may be unable to preview every collected item. Google says Gemini feedback can include personal content retrieved from Connected Apps. A preview is therefore a review aid, not a complete packet manifest.
Before you click Submit, assume that the feedback boundary may be wider than the visible answer, the source citations, and the fields shown in the feedback pane. Check the provider's current documentation, the consent choices in that exact interface, your account type, and any administrator policy.
The Short Answer
An AI feedback screen can show only part of what a provider may collect. The most important distinction is between:
- Visible answer content — what the assistant displayed.
- Previewed feedback content — what the feedback pane lets you inspect.
- Submitted context — prompts, responses, chat history, connector-derived samples, model prompts, files, and logs the provider says may travel with the report.
- Later use — debugging, human review, safety analysis, product improvement, retention, and in some products model training.
Those sets can overlap without being identical.
Microsoft documents the preview gap directly. Its current Copilot feedback guidance says relevant samples may include files or other content used to formulate an answer even when that content was not shown or referenced in the final response. It also says that when Microsoft cannot preview all collected data, the feedback pane links to more information about what will be collected.
Google's current Gemini feedback instructions say the associated conversation and included content can be collected, including personal content Gemini obtained from Connected Apps. A list of visible citations is not described as a complete feedback inventory.
What Is Confirmed
Current first-party documentation confirms that a feedback preview and a feedback payload are not always the same thing.
| Provider and documented scope | What may be submitted | What the preview does not prove |
|---|---|---|
| Microsoft Copilot with Microsoft 365 apps | Prompt, generated and intermediate responses, chat history, content used for the answer, the actual prompt sent to the model, and additional logs when the user agrees | Microsoft says some collected data may be impossible to preview in the pane |
| Consumer Gemini Apps | Associated conversation, uploads selected for inclusion, and personal content obtained from Connected Apps | Google does not say the visible answer or citation list enumerates every included connector-derived item |
| Commercial Claude | The entire related conversation, content, preferences, styles, and model settings | Anthropic says raw connector content is excluded, but connector data copied into the conversation may be included |
| OpenAI API Playground feedback, when an organization enables it | The conversation up to the feedback point, including inputs, outputs, and uploaded files | OpenAI's documentation describes the shared conversation, not a universal item-by-item connector preview |
The table compares documented scopes, not identical products. Consumer, business, school, API, and managed-workspace accounts can follow different settings and terms.
Microsoft says a preview may be incomplete
Microsoft's disclosure is unusually explicit. The Copilot feedback page says a work or school administrator can control whether content samples are collected. When the interface asks for permission to collect additional data, that data may include:
- the original prompt and system modifications to it;
- the generated answer, intermediate messages, prompt suggestions, and citations;
- Copilot chat history;
- files or other content Copilot used to formulate the answer, including content not shown or referenced in the final answer;
- the actual prompt sent to the large language model, including Microsoft proprietary content; and
- additional logs that may contain file contents, a name or email address, an IP address, or other personal data.
Microsoft says it shows links to additional data when possible. If it cannot preview everything, it provides a link explaining what else may be collected. None of that additional data is collected until the user submits the report.
That final point matters: opening the feedback pane is not the same as submitting the full report. But once the user consents and presses Submit, the absence of a preview for one item does not establish that the item stayed out.
The separate Microsoft 365 feedback overview confirms that ordinary feedback metadata can include the app, submission time, user or tenant identifiers, build and language information, attachment indicators, crash-module details, and optional diagnostics. Depending on policy and user choice, screenshots, attachments, content samples, and logs may also be sent.
For managed organizations, Microsoft says administrators can view, export, or delete feedback according to their roles. The Copilot-specific page adds that an administrator may view collected content samples that do not contain Microsoft proprietary content, although Microsoft may already have started working on a report before an administrator deletes it.
Google says Gemini feedback can include Connected Apps content
Google's Gemini feedback page says that submitting feedback adds the associated conversation, including prompts and Gemini responses. Google also says included content can contain uploaded files and images plus personal content Gemini obtained from Connected Apps.
Connected Apps can expose much more than a citation. The Gemini Apps Privacy Hub lists information from Connected Apps such as emails, files, events, photos, videos, and data from third-party services including MCP server tools. The same hub describes interaction logs, crash and debug information, device and browser data, screen content, page context, URLs, and other service context as Gemini Apps information under the applicable features and settings.
Google's consumer feedback flow does let users choose whether to include uploaded files or images in some cases. That choice is useful, but it should not be reinterpreted as a complete switch for every form of context. The provider separately says personal content obtained from Connected Apps is collected with feedback.
Google also documents a separate lifecycle. Its Privacy Hub says reviewed feedback, associated conversations, and related data may be retained for up to three years after being disconnected from the Google Account. Account type matters: Google's feedback instructions say some work or school users cannot add extra feedback or attachments, and managed Workspace data can follow different terms.
Anthropic documents a narrower raw-connector rule
Anthropic provides a useful counterexample. Its current commercial-product data-use page says feedback submitted through thumbs controls may store the entire related conversation for up to five years. That record can include conversation content, custom styles, preferences, and model settings.
Anthropic also says feedback data does not include raw content from connectors such as Google Drive or remote and local MCP servers. However, connector data may be included if it was copied directly into the conversation.
This boundary is narrower than Microsoft's documented “content used” rule, but it still requires a careful question: did the connector content remain only in the retrieval layer, or did the assistant quote, summarize, transform, or otherwise place it in the conversation being submitted?
Provider differences are why a generic statement such as “thumbs down sends the chat” is not precise enough. Which chat, what hidden context, what logs, what account, and what feedback option all matter.
OpenAI documents conversation and file sharing in an enabled feedback flow
OpenAI's current API feedback-sharing documentation says the Playground feedback feature is disabled by default for organizations. When an owner enables it, a user can share the conversation up to that point with OpenAI's feedback systems, including inputs, outputs, and uploaded files.
OpenAI also warns organizations not to share sensitive, confidential, or proprietary information and says accounts with Zero Data Retention cannot opt into these sharing mechanisms. This is an API organization setting, not proof that every ChatGPT thumbs control or connector behaves the same way.
The durable lesson is consistent: feedback is a separate data path. Normal training defaults, chat-history controls, connector permissions, and feedback submission can each have their own boundary.
What Is Still Unclear
First-party documentation describes possible categories, but it does not expose the exact serialized payload for every live feedback report.
- Microsoft does not publish a universal field-by-field payload for every Microsoft 365 app, tenant policy, device, build, Copilot feature, or experiment.
- “Content used to formulate the response” does not reveal whether the packet contains a full file, an extracted passage, nearby text, a derived representation, an identifier, or a reference.
- Google does not state on its feedback help page how every Connected Apps item is selected, transformed, or represented in a specific report.
- A connector fact can be visible in a summary without revealing how much source material accompanied it.
- A source can be absent from citations while still having influenced the answer or debugging context.
- A provider's statement that feedback is not used to train a foundation model does not answer questions about human review, debugging, safety analysis, product improvement, administrator access, or retention.
- Deleting or disconnecting the original connector after submission may not delete a separate feedback copy governed by another lifecycle.
These unknowns do not prove that a provider secretly sends every accessible document. They show why absence from a preview is not proof of absence from the submitted data.
Use The PREVIEW Test Before Submitting AI Feedback
Use this seven-part PREVIEW check whenever an assistant has touched email, cloud files, calendars, photos, code repositories, CRM records, or another connected source.
P — Pause before pressing Submit
A thumbs selection may record the initial rating immediately, while the full text, samples, files, and logs may wait for a separate Submit action. Read the screen before completing the second step.
R — Read the provider's feedback disclosure
Open the linked “learn more” page in the feedback pane. Look for explicit categories: conversation history, retrieved content, attachments, system prompts, intermediate responses, diagnostics, identifiers, human review, training, and retention.
E — Examine every available preview and toggle
Review all preview links, attachment checkboxes, screenshot choices, and content-sample permissions. Remove anything optional that is not required to report the problem.
V — Verify the account and administrator policy
Personal, business, enterprise, education, and API products can differ. In a managed tenant, ask whether administrators allow content samples or logs, who can inspect feedback, and whether the feedback feature can be disabled for sensitive workflows.
I — Identify every source the assistant used
Do not stop at citations. List the connected apps, files, messages, calendar events, photos, repositories, and previous conversation turns that could have informed the response. Treat unquoted source content as potentially relevant until the provider says otherwise.
E — Exclude unnecessary sensitive context
If the issue can be reproduced with harmless synthetic data, use that reproduction instead of a real customer record, medical detail, contract, private message, credential, or unreleased document. Redact the source before a new test rather than relying on a feedback preview to perform the redaction for you.
W — Write down the submission boundary
For business workflows, record the product, account type, date, toggles, previewed items, documentation URL, and reason for submission. That small audit note is more defensible than a later assumption about what the interface probably sent.
A Harmless Verification Exercise
Do not test with real sensitive data. Create a synthetic file or email containing a unique marker such as FEEDBACK-CANARY-8427 and a clearly fake sentence.
- Connect only the harmless test source.
- Ask the assistant a question that requires retrieving the marker.
- Inspect the answer and citations.
- Open—but do not submit—the feedback pane.
- Inspect every preview, checkbox, and disclosure link.
- Record whether the marker is visible, whether the source is listed, and whether the provider says additional unpreviewed data may be collected.
- Close the pane without submitting unless the test and the submission are authorized.
This exercise can show what the interface reveals. It cannot prove the exact server-side payload or guarantee that a future app version behaves identically. For stronger assurance, an organization needs provider documentation, contractual terms, tenant policy review, data-subject or administrator export where available, and an approved test environment.
What This Does Not Mean
An incomplete preview is not automatically deceptive, insecure, or a breach. Debugging an AI response may require context the final answer did not display. Logs and content samples can help a provider reproduce a failure that a one-line comment cannot explain.
The privacy question is whether the product makes the boundary understandable and gives the user or administrator appropriate control. A defensible feedback flow should clearly distinguish:
- the initial rating from the full submission;
- required metadata from optional content;
- previewed items from unpreviewable categories;
- retrieved connector data from conversation copies;
- human review from foundation-model training; and
- the source record's lifecycle from the feedback copy's lifecycle.
Nor does this article prove that every provider includes raw connector content. Anthropic expressly says its commercial feedback data excludes raw connector and MCP content unless that data is copied into the conversation. That is a materially different documented boundary and should be represented accurately.
Where OpenVeil Fits
OpenVeil is an 18+ privacy-focused hosted AI workspace with browser-local normal chat history and no server-side chat-history record for private chat sessions. It supports chat, web search, files, voice, image and video tools, and browser-local custom personas. OpenVeil does not use prompts, uploaded files, images, audio, selected local-history context, or AI outputs to train foundation models.
That can be a useful fit for adults who want a narrower private AI workspace without granting a broad assistant ongoing access to email, calendars, cloud drives, or enterprise repositories. OpenVeil's normal browser-local history model also avoids turning every ordinary chat into a server-stored account history.
The boundary still matters: OpenVeil is hosted, not fully offline. Active requests may be processed transiently by OpenVeil and necessary AI, search, upload-processing, hosting, routing, security, billing, and infrastructure providers. Uploads, search, voice, images, and video each create their own active-request paths. OpenVeil is not a connector-security product, data-loss-prevention system, feedback-payload auditor, or guarantee against a user voluntarily submitting sensitive material.
Read the OpenVeil privacy policy and the guide to AI feedback and connector data before choosing a workflow.
Frequently Asked Questions
If a file is not cited, can it still be included with feedback?
Yes, depending on the product. Microsoft says Copilot feedback samples may include files or other content used to formulate an answer even when that material was not shown or referenced in the final response.
Does a feedback preview show the exact server payload?
Not necessarily. Microsoft explicitly says it may be unable to preview all collected data. A preview is evidence about what the interface exposes, not a cryptographic or field-level receipt for the full payload.
Does clicking thumbs down immediately send everything?
Provider flows differ. Microsoft says the initial thumbs choice is collected, while the additional data described in its Copilot guidance is not collected until the user submits the feedback report. Read the exact interface and documentation for the product you are using.
Can Gemini feedback include Gmail or Google Drive content?
Google says Gemini feedback can include personal content obtained from Connected Apps. Its Connected Apps documentation includes emails, files, events, photos, videos, and other data, subject to the product, account, settings, and feature used.
Does Claude feedback include raw connector content?
Anthropic says its commercial feedback data excludes raw connector content, including remote and local MCP sources, but connector data may be included if it is copied into the conversation.
Does “not used for foundation-model training” mean the feedback is not retained or reviewed?
No. Training, retention, debugging, safety review, product analysis, administrator access, and human review are separate questions. Microsoft says Copilot feedback is not used to train the foundation models used by Copilot with Microsoft 365 apps, while its documentation still describes collection and approved product-improvement use.
Should I submit feedback about a sensitive connector error?
Submit only if you are authorized and understand the boundary. Prefer a harmless synthetic reproduction, remove optional context, review every disclosure and toggle, and use an approved support or security-reporting channel when ordinary product feedback is not appropriate.
Bottom Line
A visible feedback preview can be incomplete. Microsoft says so directly, and Google says Gemini feedback may include personal content retrieved from Connected Apps. Treat the preview as one layer of evidence—not as proof that every submitted prompt, file, connector sample, log, or identifier is visible.
Before submitting, run the PREVIEW test: pause, read, examine, verify, identify, exclude, and write down the boundary. If the issue involves sensitive data, reproduce it with harmless synthetic content or use an approved security and support process instead.