Is AI Safe for Co-Parenting? A Privacy Checklist

AI can be used safely in a co-parenting workflow only when the parent understands what information leaves the device, why it is needed, who processes it, and what happens afterward. “Powered by AI” does not answer any of those questions.

Family records can contain children's names, school and medical details, custody schedules, home addresses, finances, allegations, and court documents. The correct default is data minimization: provide only the information needed for the specific task.

Start by mapping the data flow

For each AI feature, ask:

1. What action triggers it?
2. What exact content is sent?
3. Does it include only the current draft or broader family history?
4. Which company receives it directly?
5. Which infrastructure or model providers process it?
6. Is the input or output stored, logged, or reviewed?
7. Is it used to train or improve any model?
8. How can the user delete it or opt out?

Different features inside one app may have different answers. A tone check may send one unsent draft. A family assistant may retrieve selected records. A weekly summary may use aggregated counts. A document extractor may process the uploaded file. Evaluate each flow rather than assigning one vague privacy label to the entire app.

The 10-point AI co-parenting privacy checklist

1. Purpose limitation

The app should explain what the AI is doing: tone review, summarization, extraction, categorization, or question answering. Avoid open-ended access to the full family account when one draft would be sufficient.

2. Minimum necessary context

Remove information the feature does not need. A tone rewrite usually does not require a child's full name, date of birth, school, address, case number, or medical history.

3. Training policy

Look for a direct statement about whether inputs and outputs train or improve models. Amazon states that Bedrock customer content is not used to train base models or shared with model providers (AWS Bedrock privacy). That statement applies to Bedrock's service; it does not automatically describe every app, proxy, log, or database around it.

4. Retention and logs

“Not used for training” does not necessarily mean “not retained anywhere.” Check application logs, infrastructure logs, abuse monitoring, support access, backups, and output history. Ask how long each copy remains.

5. Human access

Determine whether employees, contractors, support agents, or model providers can review content, and under what conditions. Policies should distinguish routine processing from exceptional security or support access.

6. Children and third parties

A parent may hold information about a child, co-parent, doctor, teacher, therapist, or lawyer who did not choose the AI tool. Share only what is necessary and permitted. Do not upload confidential third-party records merely because the account allows files.

7. Account security

Privacy fails if the account is compromised. Use unique passwords, multi-factor authentication where available, protected recovery methods, current devices, and careful control of shared email or app-store accounts.

8. Permissions and visibility

Confirm who can see the input, output, and final saved item: the user alone, the linked co-parent, children, invited relatives, or professionals. A private draft and a shared family message should not be confused.

9. Deletion and export

Understand whether deleting an AI conversation removes the underlying family record, whether closing the account deletes data, and which exports remain available. Preserve required records before deleting anything subject to an order or legal hold.

10. High-risk exclusions

Do not send passwords, government identifiers, banking credentials, precise live location, intimate media, privileged legal advice, or unrelated medical details through an AI feature unless there is a clearly justified and protected process.

What research and standards say

NIST's Generative AI Profile identifies privacy, confabulation, information security, and human oversight among the risks organizations should manage (NIST AI RMF). The framework is aimed at organizations, but its questions are useful for families: what is the intended task, what can go wrong, and who reviews the result?

Research on AI-mediated interaction also suggests that transparency affects trust. In nine studies with 6,282 participants, perceived AI involvement changed how people rated empathy and support, even when the response text was held constant (Nature Human Behaviour). Privacy and transparency are not merely technical details; they affect how communication is understood.

General chatbot versus integrated feature

A general chatbot may be convenient, but the user must supply context manually and investigate its settings. An integrated feature can limit the input, connect outputs to source records, and provide a product-specific privacy explanation. Neither category is automatically safe.

Prefer a service that clearly documents:

• the feature-specific input;
• optional versus automatic processing;
• model and infrastructure providers;
• training and retention rules;
• user control and human review;
• source citations for summaries or answers.

OnePage's published approach

OnePage documents its optional AI features separately in the privacy policy. The policy explains that features run only when the user invokes them, describes what each feature sends, and states that AI content is not used to train models. AWS separately states that Bedrock inputs and outputs are not used to train foundation models (AWS FAQ).

Read the current policy rather than relying on this summary. Privacy practices, vendors, and functionality can change, and the published policy should remain the source of truth.

For product differences, read co-parenting apps with AI. For the editing workflow itself, see can AI help co-parents communicate better?.