An AI saying “done” is a claim, not proof that something happened. A chatbot may have written a draft, attempted an action, or reported success incorrectly. To know whether a task is complete, check the actual result: the saved file, sent message, calendar entry, or other change you requested.
Imagine asking an AI assistant to email your teacher about a school project. It replies, “Done! Your message has been sent.”
Wonderful—until you discover the message sitting in the chat window, not your email account.
The words sounded like a finished job. But where was the email?
That is the myth we need to bust: a confident completion message is not the same as a completed task.
Saying Something and Doing Something Are Different Jobs
Compare these two requests:
- “Write a friendly reminder about tomorrow’s meeting.”
- “Send a friendly reminder to everyone attending tomorrow’s meeting.”
The first asks for words. The second asks for a change outside the conversation.
For an AI system to take an outside action, it needs a suitable tool or connection. Depending on how the system is built, another piece of software may carry out the requested action and return the result. Anthropic’s explanation of AI tool use describes this separation between requesting a tool action, executing it, and receiving its result.
Think of a restaurant. Writing “one cheese pizza” on a notepad is not the same as placing an order. Placing the order is not the same as receiving your pizza.
Each step matters.
For practical purposes, separate AI work into three kinds:
- Creating: producing a draft, plan, picture, or other output.
- Acting: using a connected tool to change something.
- Checking: looking for evidence that the requested result happened.
When you ask AI to do something, decide which of those jobs you actually need.
Why Might AI Say “Done” Too Soon?
There is no single explanation for every incorrect completion message. Two important possibilities are a false claim and an incomplete check.
The completion claim may be made up
A language model—the technology behind many chatbots—generates responses using learned patterns in language. That process can produce convincing statements that are inaccurate or unsupported. Smooth wording does not guarantee that the system has checked what happened.
Yale’s guide to AI hallucinations specifically describes “capability hallucination”: an AI claiming to have performed an action it cannot perform, or saying it completed a task when it did not.
You do not need to decide whether the chatbot was “lying” to spot the problem. The useful question is simpler: What evidence supports its statement?
Some work may be finished—but not enough
AI can also make real progress and stop short of the goal.
In Anthropic’s work on long-running AI agents, researchers described coding agents declaring a project finished too early. They also observed features being marked complete without sufficient testing. More explicit requirements and testing through the app’s interface helped address those problems.
The everyday lesson: progress and completion deserve different labels.
The Four Stages Hidden Inside “Done”
Here is a simple way to describe a task more clearly:
- Planned: The system has described what should happen.
- Attempted: It has tried to carry out an action.
- Reported successful: A tool or service has returned a success message.
- Verified: The resulting state has been checked against your request.
Treat these as useful labels, not a promise that every AI app will display them.
Suppose you ask an assistant to create a calendar event. Your checklist might be:
- Did it only suggest a title and time?
- Did it attempt to create the event?
- Did the calendar service report success?
- Can you find the correct event in the intended calendar?
For the final check, inspect the date, time, time zone, and guest list.
“Something happened” and “the right thing happened” are different conclusions.
What Counts as Good Evidence?
Choose evidence that matches the task. Here are some practical examples:
| Your request | What to check | |---|---| | “Send this email.” | Find the message in the email service’s sent folder; check the recipient and attachment. | | “Save this document.” | Open the file from the intended location and inspect its contents. | | “Add this appointment.” | Open the calendar and check the event details. | | “Update this spreadsheet.” | Inspect the requested cells and confirm unrelated information remains intact. | | “Fix this website button.” | Try the button and check that it produces the intended result. |
Keep your conclusion as narrow as your evidence. Finding a sent email, for example, is not the same as knowing the recipient has read it.
For document edits, preserve the original and compare it with the new version. Our guide to using AI to compare two versions of a document explains a useful approach: ask for specific changes, then check those changes in the files themselves.
Ask for a Checkable Report, Not Reassurance
Instead of ending with “Are you sure?”, ask for details you can inspect.
Try this instruction:
Before acting, tell me whether you have the necessary tool and permission. Afterward, separate what you attempted, what the tool reported, and what you verified. Give me a way to check the result. If anything is unfinished or uncertain, say so clearly.
For a document task, you might add:
Give me the saved file’s name and location, and tell me whether you reopened it to check the changes.
This is a way to request a clearer report—not a guarantee of accuracy. The report still needs checking, because an AI can generate unsupported details as well as unsupported claims of success.
That includes convincing-looking references. Our explanation of why AI citations can look real even when they are wrong explores the same distinction: looking official is not proof.
What Should You Do When the Result Is Missing?
First, pause. Do not immediately ask the assistant to repeat an action that could create an unwanted duplicate.
Imagine a party invitation that might already have been sent. A second attempt could leave your friends with two invitations. Before retrying, check the relevant app.
Use this recovery routine:
- Inspect the destination. Look in the calendar, folder, spreadsheet, or message service.
- Check for partial completion. Identify what changed and what is still missing.
- Describe the gap precisely. Say, “The file exists, but the final section is unchanged,” rather than simply, “It didn’t work.”
- Choose the next step. Ask for a targeted correction, retry after checking, or finish manually.
For important actions, consider requesting a preview before allowing the change. “Show me the message before sending” gives you a review point.
If you are helping a child use AI, keep an adult involved before sharing personal details, spending money, or sending messages to other people.
Better Trust Starts With Better Questions
The lesson is not “AI can never do anything.” Tool-using AI systems can perform actions, and developers can build checks around those actions. The lesson is that an assistant’s final sentence should not replace evidence of the result.
You do not need to inspect every playful poem with a magnifying glass. Match your checking effort to what could go wrong.
For a birthday rhyme, read the rhyme. For a saved assignment, open the file. For an important message, inspect the actual message service.
Keep this rule close:
Use AI’s answer to understand what it claims happened. Use the result to check what actually happened.
“Done” can begin that check. It should not end it.


