Data Sharing Policy
Help improve AI. Choose what you contribute.
The full run sharing terms are a draft, not an effective policy. Basic run analytics is already implemented and is described separately below. On-device personal-data exclusion or anonymization must be implemented and verified before sharing is enabled. Retention, recipient disclosures, international processing arrangements, and withdrawal procedures must also be finalized.
01 Basic & full collection
Basic run collection is part of using LocalAI Tools, including when you are signed out. Instrumented runs, model loading and downloads record operational metadata: model/runtime identifiers, requested precision/device, timestamps, measured timing, success/failure/cancellation status, coarse error categories, streaming measurements, run/project/workflow correlation IDs, random installation/session IDs, browser/OS categories, available device information, language and approximate connection country/region/timezone where available. We use this metadata to understand usage, reliability, compatibility and speed.
Basic records exclude prompts, messages, input files/media, generated content, full configuration, filenames, raw exception messages, account IDs, raw user agent strings and IP addresses. Random identifiers and technical context are not a guarantee of anonymity. Full sharing uses a separate browser preference.
Full run collection is optional and off until you enable it. When available for a model, the setting covers future eligible direct model runs in this browser and can include inputs, files, settings, and final results. Opting in under the notice that includes Structured Data also covers individual browser model calls in that app, including source content and extracted results. Earlier choices cover direct model runs only. Other app workflows, comparison runs, benchmarks, and model-page batch runs are excluded.
Use Privacy settings to turn optional full sharing on or off in this browser, with or without an account. Turning it off clears pending full uploads. Basic logging continues. For data already received, contact us to request withdrawal or deletion.
Errors and interruptions: basic records identify a failed or cancelled outcome and elapsed time, with a coarse error category for failures when available. Explicit interruption or termination is recorded as cancelled when reported by the runtime. Raw error messages remain in local run history. With full capture authorized, a failed or cancelled run may still upload its input and configuration; output is null when no final result was returned. Stream chunk counts do not mean partial generated content was captured.
A full record’s “complete” capture status describes serialization, not successful inference; the linked basic outcome describes the run. Omissions are marked partial. A crash, abrupt tab closure, storage failure or interrupted delivery can leave a start without a finish or no full record. A missing finish has an unknown outcome.
02 Your choice
You can contribute a run to help improve AI models and related tools. Sharing is optional. You can run models without agreeing.
Allowing full sharing authorizes LocalAI Tools to collect eligible future runs in this browser for the purposes described here. Previous one-run choices do not carry over to this setting.
04 Personal data protection
Identifiable personal data must not be collected through optional run sharing. Data must either contain no information that identifies a person or be effectively anonymized on your device before it leaves the device. Anonymization after upload does not satisfy this requirement.
This applies to inputs, outputs, files, metadata, filenames, embedded document properties, and identifiers. Examples include names, email addresses, phone numbers, addresses, account details, faces, voices, precise locations, and combinations of information that could identify someone.
Removing a name or replacing an identifier with a hash is not enough if a person can still reasonably be identified. The check must address indirect identification and links between records, as well as obvious identifiers.
If text, images, audio, video, documents, or technical metadata cannot be reliably anonymized, that content must be excluded from capture. Your agreement to sharing does not override this restriction.
Implementation status: the current full-capture system does not guarantee these safeguards. Full sharing must remain disabled until the required on-device controls and verification are in place. This draft is a requirement for that implementation, not a claim that anonymization already works.
This section covers optional full shared run data; it does not claim that basic operational metadata is anonymous. Account details and information you intentionally send through a support request are separate and are governed by the applicable privacy information.
05 How it may be used
- Diagnose errors and understand model behavior.
- Evaluate accuracy, reliability, safety, compatibility, and speed.
- Create evaluation datasets and improve testing methods.
- Label, annotate, correct, or prepare examples for research.
- Train, fine-tune, optimize, or improve AI models and supporting software.
These activities may support commercial products as well as research. Human reviewers may examine shared content.
06 Who may receive it
Under this draft, eligible run content could be reviewed by authorized LocalAI Tools staff and shared privately with people and organizations working on model research, testing, annotation, or improvement. These may include the original model provider, other model developers, independent researchers, evaluation teams, contractors, and collaborating organizations. Service providers may handle the data to provide services such as hosting and storage.
Authorized people may see eligible anonymized examples, including inputs, files, outputs, and related run details. Access would be limited to the purposes described above. All run content must pass the on-device personal-data checks before upload.
Before you opt in, the notice would explain who may receive the data and why. Where applicable, it would name independent organizations that rely on your consent. Public posting, sharing with other users through a link, and release in a public dataset are outside this permission and would require a separate opt-in for eligible content.
07 Your responsibility
Only share content you are entitled to contribute for these purposes. Avoid submitting passwords, access keys, confidential business information, or sensitive information about yourself or others. Your agreement does not replace any permission required from other people represented in the content.
You retain your ownership rights. You authorize the copying, processing, annotation, and sharing reasonably needed for the purposes described here, subject to applicable law.
08 Storage & retention
Shared records and files are stored online and may be processed internationally.
Proposed retention: eligible anonymized run records and uploaded files would be retained for up to 90 days, with deleted copies removed from backups within an additional 30 days. These periods must be confirmed and implemented before this policy takes effect. Engineering proposals currently distinguish 90 days for basic metadata and 30 days for full records/assets. Automatic retention and deletion are not implemented; these are not current deletion guarantees.
09 Changing your mind
Turn off full sharing in Privacy settings to stop future full uploads. After submission, contact us to request withdrawal or deletion. We may need information identifying the relevant run.
Withdrawal stops future consent-based use where applicable; it does not undo processing that lawfully occurred beforehand. Deleting source records may not remove their influence from models already trained. We must explain any applicable limits rather than promise complete reversal.
10 Privacy & your rights
This policy covers the optional contribution of full run data. Our Privacy Policy separately explains account information, ordinary service analytics, and applicable privacy rights. Agreeing here does not waive those rights.
Broader sharing requires a new consent version. An earlier agreement limited to the original model provider does not authorize this expanded sharing.