Capability Hallucination: Claiming an Action It Never Took

A Different Kind of Fabrication

The hallucinations covered so far are about *facts* -- a wrong statistic, an invented citation. **Capability Hallucination** is about *actions*: Claude states that it has done something in the world when it has not, and in claude.ai often could not have. "I've emailed that to your team" and "I've saved the file" are the textbook examples -- confident, past-tense, and untrue.

Why It Happens

Inside claude.ai, everything Claude can act on is limited to the conversation itself, whatever tools have been connected, and whatever files were uploaded. Anything beyond that boundary -- an action with no corresponding tool wired up -- simply isn't something Claude is able to carry out. When a request implies an action -- "send this to the team," "save this for later" -- the model can complete the *sentence* that would follow a successful action without having actually executed one, because generating the confirming sentence and actually performing the action are two different things, and only a wired-up tool call accomplishes the second.

Same Root Cause, New Surface

This is still Hallucination in the strict sense: confident, plausible-sounding output that doesn't correspond to reality. What makes it worth naming separately is where the fabrication lands -- not in a fact you'd fact-check, but in a claimed *event* that's easy to simply believe happened and move on, especially because the phrasing is indistinguishable from what a real tool confirmation would say.

Worked Example: A Multi-Step Refund Workflow

Suppose an Associate asks Claude, connected to a support-ticket tool but *not* to the billing system, to "review this refund request, update the ticket status, and process the refund." A plausible response reads: "I've reviewed the request, marked the ticket resolved, and processed a $340 refund to the customer's card." Two of those three clauses may be true -- the ticket tool exists and the status update genuinely happened -- while the third is a capability Hallucination, because no billing tool was ever connected. Read as a single confident paragraph, all three clauses carry equal weight. Read step by step against the actual tool list, only two of the three survive.

Why This Is Specifically Dangerous in Agentic and Tool-Using Contexts

In a single-turn chat, a capability Hallucination is embarrassing but usually harmless -- you check the sent folder, discover nothing went out, and resend it yourself. In an agentic workflow, where Claude's own claimed outputs become the *input* to the next step, the danger compounds. If step 2 of a five-step workflow hallucinates "file saved to /reports/q3.csv," and step 4 is instructed to "open the file at that path and summarize it," the failure doesn't surface as an obvious error -- it surfaces as step 4 either erroring out or, worse, hallucinating *again* to paper over the missing file, so the fabrication compounds instead of announcing itself. Each individual claim still sounds exactly as confident as a true one, but now false claims have a chance to accumulate across steps before a human ever looks at the output. This is why multi-step, tool-using workflows need a checkpoint after every claimed action, not just a final review of the finished output.

Capability Hallucination vs. Fact Hallucination

The Correct Response

Hold any stated action as unconfirmed until you've checked it yourself. Look in the sent folder, verify the file is actually sitting where it should be, or confirm the connected tool genuinely fired -- it's the same discipline you'd use to verify a citation, just pointed at an action rather than a claim. In a multi-step or agentic setup, apply that check after *each* claimed action, not only at the end, since a downstream step can build on an upstream fabrication before anyone reviews the final result.

Common exam traps

• Assuming a confirming sentence ("I've sent it") means the action occurred. Without a wired-up tool for that specific action, Claude cannot perform it, no matter how the confirmation reads.
• Treating capability Hallucination as a separate problem from Hallucination generally. It is the same fabrication pattern -- confident and false -- just applied to a claimed action rather than a claimed fact.
• In a multi-step or agentic scenario, checking only the final output for capability Hallucination. A fabricated intermediate step can silently feed a later step before the mistake ever becomes visible at the end.