A reference for AmeriCorps program leaders

AI privacy · practical guidance

Keep your program’s data
out of AI training.

Before you paste a prompt or upload a file, check the tool and the account you’re using. Here’s how to limit training use across eight major AI services.

Official guidance checked October 7, 2026 · U.S. audience · Settings and terms can change.

01

OpenAI

ChatGPT

Turn off model improvement

Personal accounts, including paid personal plans

What to do

  1. Open your account menu and select Settings.
  2. Select Data Controls.
  3. Turn off “Improve the model for everyone.”
What this changes

New conversations are excluded from training. History can remain; the setting follows your signed-in account across devices.

What to watch for

Temporary Chat is also excluded from training. Feedback (thumbs up/down) can allow the associated conversation to be used for training even after opt-out. Turning off memory is a separate setting.

Using a work account?

ChatGPT Business, Enterprise, Edu and the API exclude inputs and outputs from training by default. Confirm you are in the approved work workspace.

Official guidanceData controlsTraining & feedback
02

Google

Gemini

Turn off Keep Activity

Personal Google accounts, including Google AI subscriptions

What to do

  1. Open Gemini, then Settings & help → Activity (or Gemini Apps Activity in the account menu).
  2. At the top of your activity page, select On.
  3. Choose Turn off, or Turn off and delete activity. Confirm Keep Activity is off.
What this changes

Future chats are excluded from model improvement if you do not submit feedback. Google still keeps them for up to 72 hours.

What to watch for

Temporary chats are not used to train Google’s AI models. Feedback with Keep Activity off can include the last 24 hours of chats and their uploads. Previously reviewed data can be retained separately; switching off is not a reversal of past training.

Using a work account?

Qualifying Workspace core-service editions provide enterprise data protections. Some work accounts use Gemini as an additional service with different rules. Ask your administrator to confirm your edition and protections.

Official guidanceActivity settingsPrivacy & feedbackWork-account protections
03

Anthropic

Claude

Turn off model improvement

Claude Free, Pro and Max

What to do

  1. Open your account/name menu and select Settings.
  2. Select Privacy.
  3. Turn off “Help Improve our AI models.”
What this changes

New and previously stored chats and coding sessions are excluded from future training runs. Training already started and existing models are not undone.

What to watch for

Incognito chats are excluded from model improvement. Safety-flagged conversations may still be used for trust and safety models or safety research. Review what you agree to share before submitting feedback.

Using a work account?

Claude for Work and Anthropic API inputs and outputs are not used for training by default. Explicit feedback or other opt-in sharing can create exceptions.

Official guidanceChange the settingIncognito & trainingCommercial services
04

Google · official help now uses “Gemini Notebook”

NotebookLM

No training toggle needed

Personal accounts and qualifying Workspace accounts

What to do

  1. Check which Google account is signed in before uploading sources.
  2. For a personal account, avoid submitting feedback on interactions containing confidential data.
  3. For program work, confirm Workspace protections with your administrator and check notebook sharing permissions.
What this changes

Google says personal-account data is not used to train the service unless you provide feedback. There is no separate training opt-out switch to find.

What to watch for

Personal-account feedback can expose the full interaction context: uploads, questions and responses. Gemini’s Keep Activity switch is not a substitute for these separate NotebookLM rules.

Using a work account?

For Workspace and Workspace for Education users, uploads, queries and responses are not used to train AI models or reviewed by human reviewers, including when rating responses.

Official guidanceData protectionsFeedback rules
05

Microsoft

Microsoft Copilot

Turn off both training controls

Consumer Copilot app and copilot.com

What to do

  1. Sign in at copilot.com and select your profile icon, then your profile name.
  2. Open Privacy.
  3. Turn off “Training on conversation activity” and “Training on voice conversations.”
What this changes

You can control training on conversation activity and voice separately. Check both controls.

What to watch for

In the desktop app: Settings → Privacy. On mobile: account menu → Account → Privacy. Memory, personalization and deleting history are separate controls.

Using a work account?

Microsoft 365 Copilot and eligible work/school Copilot Chat use enterprise protections; prompts and responses are not used to train foundation models. Confirm the exact product and account with IT.

Official guidanceConsumer privacy controlsWork/school data protection
06

Perplexity

Perplexity

Turn off AI data usage

Personal accounts, including Pro

What to do

  1. Open your profile and Settings.
  2. Find the AI Data Usage training control and turn it off.
  3. If your interface instead labels it AI data retention, check its description and confirm that training is disabled.
What this changes

Perplexity’s official FAQ documents an AI Data Usage opt-out for using search data to improve its models.

What to watch for

The official FAQ available for this review is older than the other guides. Labels may differ; confirm the current description in your account. Do not treat a paid personal plan as an enterprise privacy agreement.

Using a work account?

The official plan comparison lists Enterprise plans as excluding data from model training.

Official guidanceOfficial data FAQPlan privacy comparison
07

xAI · standalone app and X

Grok

Check each service you use

Grok.com, Grok mobile app and Grok on X

What to do

  1. Grok.com: Settings → Data → turn off “Improve the Model.”
  2. Grok mobile: Settings → Data Controls → turn off “Improve the model.”
  3. On X: Settings → Privacy & Safety → Data sharing and personalization → Grok & Third-party Collaborators. Uncheck the option allowing public data and Grok interactions to be used for training and fine-tuning.
What this changes

The standalone opt-out applies to new conversations. Grok on X has its own control; check it separately.

What to watch for

Private Chat is excluded from training on standalone Grok. Feedback can still be used; X feedback can include the related conversation. Signed-out users in some regions cannot opt out. Personalization is a separate control.

Using a work account?

Business and enterprise customer content is excluded from model improvement according to the consumer FAQ. Confirm your organization’s service and contract.

Official guidanceStandalone controlsGrok on X
08

Meta · Facebook, Instagram, Messenger and WhatsApp

Meta AI

No universal U.S. switch verified

Consumer Meta AI services

What to do

  1. Open Meta’s Privacy Center and review the AI information and rights available for your location.
  2. If your region offers an objection process, follow that process and keep its confirmation.
  3. For U.S. program work, do not enter confidential data into Meta AI when you cannot verify a suitable no-training protection.
What this changes

A broadly available U.S. training-off control could not be verified from official guidance. Do not assume an ad preference or deletion request is a training opt-out.

What to watch for

Meta describes using interactions with its AI and adult public content for training. EU and UK objection processes are region-specific; they are not a general U.S. opt-out.

Using a work account?

Use an organization-approved service with documented protections for confidential program information. Public social posts can be a separate source of training data.

Official guidanceMeta AI privacyEU objection informationUK objection information

A practical starting point

Make this a team habit.

  • Set an approved account and workspace. Confirm training terms, retention, access and connected apps with the person responsible for your program’s technology.

  • Use the least data needed. Draft with placeholders. Remove names, contact details, member IDs and identifying narrative details before sharing examples.

  • Keep sensitive records in approved systems. This includes background checks, Social Security numbers, health or accommodation information, eligibility documents, bank details and identifiable beneficiary case notes.

  • Check uploads as carefully as prompts. A timesheet export, member file or spreadsheet can contain hidden columns, comments or identifying details. Review the full file before using an approved AI service.

  • Pause before rating a sensitive response. Several providers can use feedback and the related conversation even when training is off. Check the platform-specific exception above.

  • Recheck when accounts or products change. Turning training off does not erase earlier training. Deleting chats, disabling memory and preventing public sharing are separate decisions.