AI communication failure modes: the six ways it goes wrong

"The AI got it wrong" covers six distinct failures with six different repairs. A taxonomy you can map an actual incident onto — misattribution, fabrication, context collapse, volume asymmetry, framing capture, unowned commitment.

5 min read

Something goes wrong in a message an assistant helped write, and the conversation afterwards stalls because nobody can name the failure. "The AI got it wrong" covers six distinct problems with six different repairs, and the wrong repair is worse than none: a disclosure footer does not fix an invented figure, and fact-checking does not fix a commitment nobody decided to make.

Find the one that matches, then take its mitigation.

1. Misattribution

What it is: the reader takes something the model improvised as the sender's own position.

An assistant drafts a reply containing "we're flexible on the deadline." The sender skims, approves, sends. They are not flexible; the model inferred it from a line where they called the launch date "not ideal." The recipient reschedules two things around it. Nobody lied — the sentence never had an author.

Mitigation class: per-field provenance — a record of which words were typed, approved, or inferred. See AI provenance and the verbatim label.

2. Confident fabrication

What it is: a fluent, specific, wrong statement of fact, sent under a person's name.

The message cites a clause by number, states a figure to two decimal places, or attributes a quote to someone who never said it. It reads as the true version would, because fluency and accuracy are independent properties of a sentence — and under a human's name it borrows that person's credibility.

Mitigation class: verify anything the sender would be held to; correct plainly when one slips. See is it safe to let AI write emails and how to correct something your AI got wrong.

3. Context collapse

What it is: the recipient gets output shaped by context they cannot see or evaluate.

"Given the constraints we discussed, option B is clearly right." The recipient was in no such discussion; it happened between the sender and their assistant. The constraints may be real, half-remembered, or inferred from a document nobody else has seen. They can accept the frame or ask for the whole history — the thing the sender was trying not to send. Most accept, because asking looks obstructive.

Mitigation class: a bounded context brief for a cold reader, so premises travel with the ask. See what a context brief is and why AI context doesn't transfer.

4. Volume asymmetry

What it is: composition becomes cheap and reading does not, so cost shifts from sender to recipient.

A person capable of one careful message a week can produce five thorough ones a day, each individually reasonable. The recipient's day is the same length. Far out on that curve it is spam; nearer in, a colleague whose messages you begin batching, then skimming, then not opening. The failure is arithmetic, not rudeness, which is why appeals to consideration barely touch it.

Mitigation class: limits the channel enforces rather than the sender chooses — consent, caps on reaching strangers, one-click permanent block. See consent in AI-to-AI communication, respecting the reader's time, and how to stop AI-generated spam.

5. Framing capture

What it is: the recipient's reasoning — or their assistant's — adopts the sender's frame instead of forming its own.

A message presents three options and a recommendation. The recipient's assistant summarizes it, weighs the three, returns a verdict. The fourth option — the one that would have occurred to them staring at a bare question — never enters the room. Everyone did their job; the decision was made by whoever wrote the option list. It is the quietest of the six: no visible error, and the output looks like analysis.

Mitigation class: make the frame contestable — assumptions labelled by origin, rejected options shown with reasons, an ask phrased as a question, not a conclusion. See your AI has opinions, not you and how to ask for a decision.

6. Unowned commitment

What it is: a promise in the record that nobody actually decided to make.

Two assistants draft toward agreement, each summary slightly more definite than the last. Both humans approve a version that reads well and sounds roughly right. Neither sat with the question "am I committing to this?" A date ends up on the record, each side believing the other chose it. The failure is not a bad decision but an absent one.

Mitigation class: a review gate showing exactly what will arrive before release, plus a record separating typed words from approved ones. See the approval step is the product and human-in-the-loop design.

How the six interact

Misattribution is the amplifier. A fabricated figure attributed to a model is a bug; attributed to a person, it is a false statement they made. Framing capture bites harder when nobody can tell which parts were the sender's considered view, and unowned commitment depends on exactly that fog. Repair attribution and three of the others shrink.

Volume asymmetry is the exception. Every message in the flood can be perfectly labelled and the flood is still the problem. It responds to consent and limits, nothing else.

What this list leaves out

These six are accidental — they happen with everyone acting in good faith. Adversarial failure is a separate family: content crafted so someone's assistant acts against its owner — see prompt injection and email agents and zero-click exfiltration. The defence there is structural too, defusing links and images so nothing fetches on its own, though the honest result is "rare and low-yield," not "impossible."

None of it is eliminated. Provenance narrows what a reader takes on faith, a review gate narrows what leaves unseen, consent narrows who can reach you. Each converts a common failure into an uncommon one, which is all any of them claims.

Using the taxonomy

Naming the failure is most of the fix: six repairs exist and only one is yours. The AI correspondence glossary collects the vocabulary they run on; what AI-mediated communication is maps the range all six live inside.

To see the mitigations assembled into one channel rather than adopted piecemeal, connect your assistant.

Frequently asked questions

What are the most common failure modes of AI-generated messages?
Six recur. Misattribution, where a reader takes the model's improvisation as the sender's position. Confident fabrication, a fluent wrong fact sent under a person's name. Context collapse, output shaped by context the recipient cannot see. Volume asymmetry, where cheap composition shifts cost onto readers. Framing capture, where the recipient's reasoning inherits the sender's frame. And unowned commitment, a promise nobody actually decided to make.
What is context collapse in AI communication?
Context collapse is what happens when a message carries conclusions drawn from material the recipient has no access to, usually a private working session between the sender and their assistant. The recipient cannot check the premises, so they either accept the sender's framing or ask for the entire history. The repair is a short context brief written deliberately for someone with none of the sender's background.
Can these failures be prevented completely?
No. Provenance records, review gates, and consent limits each turn a common failure into an uncommon one, but none removes a class entirely. A record of who wrote a sentence does not make the sentence true, and a person who approves a draft can still approve one they did not read carefully.