Survey of Terms for 36 AI Products: Handling of Audio and PDFs Differs

Is it acceptable to input meeting recordings into generative AI? This question is difficult to answer by product name alone. A report by songchong on Qiita presents a survey that reads through the terms of service and information protection policies of 36 AI products by format. Its distinctive feature is that it counts text, audio, images, files, code, and rating operations separately.

The survey covers 36 AI products. It states that the terms, information protection policies, and support documents obtained and saved were checked one by one. 14 products made distinctions by format. Of the 84 cells formed by 6 formats and 14 products, only 21 cells contained descriptions. The remaining 22 products made no distinction by format at all, treating all inputs uniformly. A blank does not mean non-use; it indicates a state with no mention of the format.

The personal edition uses audio and screen sharing for training, while the Workspace edition does not use uploaded files.

This single sentence shows the core of the survey. That is because even for the same product, the conclusion reverses depending on the edition. Gemini has the most format-specific descriptions, with descriptions for 4 of the 6 formats.

Gemini Disclosure Case with Extensive

Format-Specific Descriptions

Gemini mentions 4 of the 6 formats. The personal edition clearly states that it uses audio and screen sharing for improvement including training. The Workspace edition states that submitted files are not even subject to human review. It stipulates that they will not be used for training of the generative AI foundation without permission. It is structured so that the level of protection clearly differs between personal and business use. A large number of format-specific descriptions also means that users face fragmented sources to check. It becomes necessary to verify the permissibility of audio and the permissibility of documents in separate documents.

How to read the 22 products that do not distinguish formats is also shown. For this group, it is natural to interpret that the policy for text input also extends to other formats. For products that use text input for training, transcribed text of recordings may be treated similarly. Conversely, for products that do not use text input, other formats can be read as also not being used. However, as long as it is not expressly stated in the terms, final judgment requires confirmation with each product’s contact point. For products with uniform handling, classification and storage design on the operational side become more important.

Cases Where Handling Reverses Between

Personal and Business Editions

There were 2 cases where the conclusion reversed by edition even under the same product name. The first is GitHub Copilot. For personal Free, Pro, and Pro+, inputs, outputs, and code snippets may be used for training and improvement. It provides a mechanism for users to choose opt-out settings. It stipulates that information from Business and Enterprise will not be used for foundation training. The premise changes between personal use and organizational use.

The second is Gemini. The personal edition uses audio and screen sharing for improvement. The Workspace edition does not use submitted documents for training. Neither can be explained by product name alone. It is necessary to describe them together with the edition name, contract form, and applicable supplementary terms. Including edition specification in confirmation items at procurement is essential.

Handling of Audio and Image Files Differs

by Product

Recordings, photos, PDF documents, and code are not treated the same as text. Policies differ by product. Format-specific examples have been organized from the 14 products that had descriptions.

For audio and recordings, there are three approaches. The personal edition of Gemini specifically names audio as used for improvement including training. Genspark states that it neither stores nor provides the original audio files, and delivers only transcripts and minutes. Manus specifically names audio and audio fragments as inputs and writes that it derives information from them. It does not touch on whether they are used for training. The handling of recordings cannot be answered unless the product is identified.

For images and forms, there are differences in disclosure of protection methods. DX Suite states that before training, it masks names and phone numbers and converts them into an irrecoverable form. On the same page, it also notes that it does not guarantee complete accuracy. Smart OCR states as a function that it trains on character-string images clipped from form images. The accuracy-improvement mechanism itself includes training use. On-site handling of highly sensitive forms requires prior evaluation.

For submitted documents, examples that clearly state non-use stand out. NotebookLM states that it does not use submitted documents for human review or improvement. Submission of feedback by users is an exception. The Workspace edition of Gemini has a similar exception structure. Felo states that for the corporate edition, it will not use data for training except with explicit consent. Operations that leave a record of the consent acquisition process are required.

For code, the specification of input scope is detailed. Cursor specifically names codebases, prompts, edit operations, and snippets. GitHub Copilot specifically names inputs, outputs, snippets, and related context. In development settings, whether the entire repository is covered becomes the issue. Confirming the availability of exclusion settings and their default values is essential.

Incidental records are stipulated in a separate category. LINE WORKS OCR separately states that it automatically creates and stores connection records. IP addresses, browser types, languages, and timestamps are covered. This is information generated without the user’s awareness. It must be read without confusing it with input information.

The move to foreground on-device processing and protection is also covered in Apple to Differentiate with Privacy in Smart Glasses. Design that does not collect and design that restricts by terms are different approaches. A perspective that evaluates both in combination is needed.

Feedback Functions as an Exception to Training Use

Mentions of rating operations were found in 7 products, the most among formats. Even for the same thumb-icon operation, the meaning changes by product. Some products clearly state they will not use it for training. Some products store the conversation that was rated. Some retain data for a certain period even if history is disabled.

Microsoft 365 Copilot clearly states that it does not use feedback submissions for foundation training. Claude stores the related conversation as part of feedback together with thumbs-up or thumbs-down ratings. It stipulates that feedback is used for training even when opt-out is set. Gemini retains data for up to 72 hours for feedback processing even if conversation history is disabled. NotebookLM states that it does not use submitted documents, while making feedback submission an exception. Claude Code describes a method to hide the feedback submission channel itself through organization settings.

The exception settings of Claude and NotebookLM are similar. Although the operating companies and products are different, the same exception exists in the same position. The structure is a principle of non-use with separate handling for feedback submissions. It can be interpreted that exceptions tend to be placed in this position. When explaining to users, the principle and the exception must be shown as a pair. A thumb icon looks like a light operation, but it constitutes an act of providing information.

Need to Distinguish Inputs, Outputs, and Logs

The survey urges reading information in three categories. Information entered by the user, outputs received, and automatically generated information. Confusing them leads to errors in judgment.

The representative of automatically generated information is connection records. They are generated regardless of the user’s intent. Placing them in the same column as input information leads to misreading. Outputs require similar care. A function that creates audio explanations from materials is something the AI creates and narrates. It does not explain the handling of the input recording. In the comparison table, the format of reading training use side by side is adopted. The intended use is to find one’s organization’s products and cross-check edition and format.

The practical points are clear. The permissibility of meeting recordings is determined by product, edition, and format. A uniform permissibility table cannot be created. The place to check is not the general provisions of the terms of service, but the individual descriptions for audio and documents. The handling of feedback submissions must also be confirmed at the same time. Keeping track of four points — retention period, presence of human review, training use, and opt-out means — will give a clear picture.

Editorial Opinion

In the next 3 to 6 months, confirmation of edition specification for corporate use will be incorporated into procurement requirements. Safety judgments based on product name alone will no longer be accepted, and the burden of terms review on information systems departments will increase. Disabling settings for feedback functions may be added to initial deployment procedures.

In 1 to 3 years, format-specific descriptions will become standard disclosure items. Policies that treat audio, images, and work logs uniformly will become difficult to maintain, and separation of retention periods and training use will be required. Users may also shift to operations premised on pre-input classification work.

The remaining issue is whether the setting of exception provisions matches users’ expectations. Retaining data for 72 hours even after disabling history seems within an acceptable range. For designs where feedback submission triggers training use, the balance between convenience and protection is being questioned.

References

• “写真📷録音🎙PDF📄コード💻いいね👍｜私の情報はどこへ漏れていくのか？ AI製品36件の規約とプライバシーポリシーを形式別に調べてみた”, by songchong — Qiita, 2026-09-13T19:32:07.000Z (ARR)
• Source URL: https://qiita.com/songchong/items/e630bf05f4adf0c16ce5?utm_campaign=popular_items&utm_medium=feed&utm_source=popular_items