Starting Safely With Scheduled Tasks in ChatGPT

This article originally described “GPT-5 automation.” The current product documentation calls the feature Scheduled Tasks in ChatGPT. Do not make the retired ChatGPT model name a prerequisite; check the current Scheduled page and account conditions.

Screen a task with five conditions

Start with “notify” or “prepare a candidate,” not an irreversible action.

Create a bounded starter task

OpenAI currently documents one-off, recurring, and change-based tasks. Supported models, tools, active-task counts, and surfaces can vary, so read the current help and the account's Scheduled page.

Every Monday at 9:00 a.m., remind me to review the following checklist.
Do not sign in to external sites or make purchases.
Label anything you cannot verify as “unverified.”
I will make every decision and perform every external action after the notice.

After creation, check the next run, time zone, notification destination, and reference scope. The help page also notes that a task created in a project with files may not have access to those project files.

Do not automate these by default

• deciding treatment or medication from symptoms;
• placing investment orders, purchases, or contracts;
• automatically evaluating customers or employees;
• publishing unverified news;
• connecting personal data beyond the task's need.

OpenAI policies retain human involvement for sensitive decisions and consequential actions. A notification still requires source verification before action.

Stop conditions

Pause the task when:

1. the same failure occurs twice;
2. its source or assumptions change;
3. credentials or personal data become necessary;
4. the output has no verifiable basis;
5. the notification is no longer useful.

Conclusion

Scheduled Tasks are not an old GPT-5 promise to automate everything. Select candidates by repetition, input, impact, reversibility, and reviewer; begin with a notification or draft; and check current availability in official help and the account interface.