Source-first safety checklist

AI Companion Boundary Settings Checklist

Boundary controls are practical account features, not a promise that every response will stay within a chosen tone. Labels can change and automated systems can behave unpredictably. This checklist helps you inspect the current settings, state a low-risk preference, test exit controls without sensitive content, and know when to stop using a conversation rather than trying to negotiate with an uncomfortable interaction.

Check the live source before acting. Policies, prices, credit rules, feature limits, labels, and account screens can change. Recheck the provider's current terms, privacy notice, checkout, help pages, and account interface. This site does not claim to have tested paid accounts and does not promise security, compliance, deletion results, fixed costs, or medical outcomes. Page updated 2026-07-26.

Fast evidence screen

What to verify before continuing
AreaUseful current evidenceReason to pause
Preference controlsCurrent settings describe tone or content choices in plain languageA label promises perfect compliance
Conversation exitNew thread, reset, close, or block controls are easy to reachLeaving requires continuing the unwanted exchange
Report routeThe form explains what evidence is sentReporting silently includes more private history than expected
Account-wide controlsMute or content preferences can be revisitedSettings reset without notice or cannot be found later

Locate controls before conversation

Question: Where are reset, new thread, mute, block, report, and support controls?

Action: Open menus and help pages before entering personal context; write down the shortest exit path. This action reduces guesswork; it does not prove anything beyond the current source and account state you actually observed.

Evidence: A control map with labels and locations, dated because the interface may move. Keep the access date and enough context to understand the wording later.

Worked example: A three-step reset path is manageable when known in advance but stressful when discovered mid-conversation.

Pause when: The only apparent exit asks you to provide more content or continue engaging. A pause is a useful result when proceeding would require an assumption.

Set a low-risk preference

Question: Can you express tone and topic boundaries without revealing why they matter?

Action: Use a generic preference statement and avoid personal history during the test. This action reduces guesswork; it does not prove anything beyond the current source and account state you actually observed.

Evidence: The current setting value and whether it persists into a new ordinary conversation. Keep the access date and enough context to understand the wording later.

Worked example: A preference such as avoiding a topic can be tested with neutral content rather than a sensitive story.

Pause when: The system repeatedly pushes against a stated preference or turns the preference into the central subject. A pause is a useful result when proceeding would require an assumption.

Test reset and persistence

Question: Does starting over remove the immediate context, and do account preferences remain?

Action: Use harmless sample text, start a new thread, and observe only what the interface currently shows. This action reduces guesswork; it does not prove anything beyond the current source and account state you actually observed.

Evidence: A brief result record with no private conversation copied into it. Keep the access date and enough context to understand the wording later.

Worked example: A new thread may reset conversation context while account-level preferences remain; verify rather than assume.

Pause when: Old content appears unexpectedly or a reset label does not explain what it changes. A pause is a useful result when proceeding would require an assumption.

Understand reporting scope

Question: What portion of the conversation and account data will a report include?

Action: Read the confirmation dialog and help text before submitting; minimize optional notes and remove unrelated details where possible. This action reduces guesswork; it does not prove anything beyond the current source and account state you actually observed.

Evidence: The report receipt, category, and current explanation of submitted evidence. Keep the access date and enough context to understand the wording later.

Worked example: A report may include nearby messages for context; check the live disclosure before sending.

Pause when: The form requests passwords, financial credentials, or unrelated identity documents. A pause is a useful result when proceeding would require an assumption.

Leave instead of escalating

Question: What is your stop condition?

Action: Decide in advance that repeated boundary failures, pressure, or disturbing output means closing the conversation and reconsidering the service. This action reduces guesswork; it does not prove anything beyond the current source and account state you actually observed.

Evidence: A personal stop rule, not repeated attempts to make the system behave. Keep the access date and enough context to understand the wording later.

Worked example: After one clear reset, a repeated unwanted pattern is enough reason to stop without explaining yourself further.

Pause when: You feel pressured to continue, disclose more, spend credits, or accept behavior you did not choose. A pause is a useful result when proceeding would require an assumption.

Copyable checklist

Complete each line from the provider's current interface or published documents. A blank or unclear answer should remain visible in your notes.

Worked situations

Preference becomes the whole conversation

A setting should help shape interaction, but generated responses may still overfocus on it. Try one neutral reset or new thread. If the pattern repeats, stop rather than adding more personal explanation in an attempt to correct it.

Record what the current source shows, keep personal details out of the worksheet, and mark unresolved questions rather than filling them with an assumption.

Reporting an unwanted response

Before sending, inspect whether the report includes the full thread, selected messages, account identifiers, or uploads. Remove unrelated details if the interface permits. Keep the receipt, then clear local previews according to your own privacy plan.

Record what the current source shows, keep personal details out of the worksheet, and mark unresolved questions rather than filling them with an assumption.

How to record a reliable answer

Use a dated worksheet with four columns: the question, the current source, the exact answer, and what remains unknown. Prefer a live detailed policy or account control over a promotional summary. Preserve enough surrounding wording to avoid changing the meaning, but cover names, addresses, payment values, conversation text, and other personal details. A screenshot records one interface at one moment; it is not proof that the same rule applies later, in another region, or to another account.

When two sources disagree, do not choose the answer you prefer. Record the conflict and ask the provider through a contact route reached from its current account, terms, or privacy page. Keep the case number and exact response. If the answer affects payment, an upload, account access, or sensitive disclosure, wait until the conflict is resolved or choose not to proceed. Unknown is a valid worksheet result.

A calm stop rule

Pause when a required term is missing, the final confirmation differs from the earlier page, a request moves outside the documented account route, or support asks for secrets that normal support should not need. Do not send passwords, one-time codes, full payment credentials, recovery secrets, or unrelated private conversation content. Do not install unknown software or grant remote control. Return through a known official address and verify the account state independently.

Related safety checklists

Frequently asked questions

Do boundary settings guarantee behavior?

No. They are controls to inspect and use, not guarantees about generated output. Keep an exit plan and stop when the interaction repeatedly ignores your limits.

Should I explain personal history so the system understands?

Only disclose what you are comfortable storing and potentially exposing. A short generic preference is usually enough to test whether the control works.

Does reset erase the old conversation?

Do not assume so. Reset, new thread, clear, and deletion may have different meanings. Read the current interface and privacy explanation.

When should I report instead of just leaving?

Use your judgment and the provider's current reporting categories. If you report, check what evidence is included and avoid adding unrelated sensitive details.