- In short
- Capability hallucination is when Claude claims to have performed an external action it was never able to take, such as sending an email or saving a file. In a standard chat interface Claude only acts on the conversation, connected tools, and uploaded files it was actually given access to, so any claimed external action should be treated as unverified until independently confirmed. It differs from a factual hallucination because it is a false claim about what was done, not about what is true.
A hallucination about actions, not facts
The failure patterns so far concern whether a statement is true. Capability hallucination concerns whether an action happened. The Claude Certified Associate - Foundations (CCAO-F) exam calls this out separately because it is easy to over-trust: a line like "I've emailed that to your team" or "I've saved the file" reads as a status update, not as a claim to be checked, and so it sails past the scrutiny you would apply to a statistic.
The distinction matters for the fix. A factual hallucination is corrected by checking the fact against a source. A capability hallucination is corrected by confirming the action against the world, did the email actually send, does the file actually exist, and, before even that, by asking whether Claude had any tool to perform the action at all.
- Capability hallucination
- When Claude claims to have performed an external action it was never able to take, such as sending an email or saving a file. In a standard chat interface Claude works only with the conversation, connected tools, and uploaded files it was actually given access to; it does not perform external actions it was not given a tool for. Any claimed external action is treated as unverified until independently confirmed. It is a false claim about what was done, distinct from a false claim about what is true.
What Claude can and cannot do in a chat interface
The ground truth is about access. Within a standard chat interface, Claude operates on three things: the conversation itself, any connected tools it has been given, and the files you uploaded. It does not reach outside that boundary. It cannot send an email, write to your file system, post to a system, or take any external action unless a specific tool for that action was actually made available to it. Knowing this boundary is what lets you recognise a capability hallucination for what it is: a claim about crossing a boundary the model could not cross.
Treat claimed actions as unverified
Because the claim reads like a confirmation, the discipline is to withhold trust from it by default. "I've saved the file" is not evidence the file was saved; it is a sentence that needs checking exactly like any other. The correct posture is to treat every claimed external action as unverified until you independently confirm it happened, by looking for the email in the sent folder, the file on disk, the record in the system. The stated confirmation and the actual confirmation are two different things, and only the second one counts.
Check whether the tool even existed
There is a faster first check than confirming the outcome: ask whether the capability was ever present. If no email-sending tool was connected, then "I've emailed your team" could not have happened regardless of how the outcome looks, and the claim is a capability hallucination on its face. Verifying tool availability is often quicker than verifying the result, and it catches the most clear-cut cases immediately, a claimed action for which no corresponding tool was ever in play.
What the CCAO-F exam trips candidates on
The first trap is assuming a stated confirmation of an action, such as "I've saved the file," means the action actually happened. The sentence is not the deed. The credited answer treats the confirmation as a claim to verify against the world, not as the verification itself.
The second trap is failing to check whether a tool was even available before trusting that it was used. This is the fast path to catching a capability hallucination, and skipping it means over-trusting a claim that a two-second check would have exposed. The exam rewards asking "was there a tool for this at all" before asking "did the tool succeed."
Worked example
You ask Claude to draft a summary and 'send it to the team.' It replies with the summary and adds, 'I've emailed this to your team.' Your chat session has no email integration connected. What has happened, and what should you do?
This is a capability hallucination. The tell is not a false fact; the summary itself may be perfectly good. The false part is the action claim, "I've emailed this to your team," which reads like a status update and so invites you to move on as if the task is done.
Run the fast check first: was an email-sending tool ever available in this session? It was not, there is no email integration connected, so Claude had no capability to send anything. That settles it immediately. In a standard chat interface Claude acts only on the conversation, connected tools, and uploaded files, and sending email is outside that boundary unless a specific tool was provided. The claim to have emailed the team therefore could not be true regardless of how confidently it was stated, and the tool-availability check exposed it faster than hunting through a sent folder would have.
What to do: do not treat the confirmation as the deed. Send the summary yourself through your actual email, and going forward treat every claimed external action, saved files, posted updates, sent messages, as unverified until you confirm it independently or confirm a real tool performed it. The accountability for the email actually reaching the team stays with you, which connects to the broader accountability ownership principle: the tool's claim does not discharge your responsibility for the action.
Common misreadings to avoid
Misconception
If Claude says 'I've saved the file' or 'I've sent the email,' the action happened.
What's actually true
Misconception
Capability hallucination is just another factual hallucination.
What's actually true
How this shows up on the exam
Domain 2 questions on this knowledge point show Claude asserting it took an external action and ask whether to trust it or what to check. The dependable answer is to treat the claimed action as unverified, confirm it independently, and check tool availability first, recognising the pattern as a claim about a deed rather than a fact.
Capability hallucination is one of the signatures grouped under the hallucination pattern taxonomy and it appears again in diagnosing failure patterns in outputs, where matching the observed problem to the right named pattern determines the fix. Because the unperformed action remains the professional's responsibility, it also ties to the accountability ownership principle.
Claude replies 'I've emailed the summary to your team,' but your session has no email integration connected. What is the correct conclusion and action?
People also ask
Can Claude claim to do something it cannot actually do?
What is a capability hallucination?
Should you trust when Claude says it saved a file?
Watch and learn
Official Anthropic Academy lessons first, then hand-picked walkthroughs. Videos load only when you press play.
No videos curated for this concept yet
We are still curating the best official and community videos for this topic.
Official prep for this domain
Anthropic's own free prep module for this part of the syllabus, on the official prep course. Free with an Anthropic Academy sign-in.
References & primary sources
Master this concept with Archie
Practice it inside an adaptive study session. Archie, your Socratic AI tutor, tracks your mastery with Bayesian Knowledge Tracing and schedules the perfect next review.