04  The three failure modes

Across all five tool categories, three failure modes show up again and again. They are not rare. They are predictable. Knowing them is the difference between using AI on purpose and using it hopefully. Each has practical defences you can build into your work today.

From the research

These three are not a tidy framework invented for a course. They emerged as the most consistently reported patterns across interviews with communicators working with AI in high-stakes contexts, supported by analysis of where AI-enabled communication has failed publicly over the last two years. Hallucinations and confidentiality slips have been documented for years. The composite voice is the failure mode practitioners report most often but governance frameworks discuss least, and the one most likely to undermine the work without leaving a visible trace.

Failure mode 1: Hallucinations

What happens

AI tools, particularly conversational AI, produce confident, fluent, plausible content that is wrong. The numbers can be invented. The citations can be invented. The history can be invented. The quote attributed to a public figure can be invented. The output looks indistinguishable from accurate output, and there is no internal cue that the system is hallucinating.

Why it happens

Conversational AIs are trained to produce plausible text, not to flag uncertainty. The same statistical mechanism that produces fluent prose produces fluent inaccuracy.

Practical defences

  • Treat every factual claim as unverified. AI output is a starting point, not a fact source. The verification step is part of your workflow, not optional.
  • Use citation-grounded tools where accuracy matters: Perplexity, Claude with retrieval over your own sources, or ChatGPT with web search return sources you can click. Even then, click them. Sources can be invented or misread.
  • Ask the AI to flag uncertainty. A simple addition ("if you are unsure of any fact, flag it") reduces fabricated confidence. Not perfect, but a useful first filter.
  • Build a pre-publication checklist. Three things to verify for any AI-assisted output: any number, any name, any date. Most hallucinations show up there.
  • Match the tool to the task. Hallucination risk is highest in open generation and lower in structured tasks (summarise this document, extract these fields, translate this paragraph).

Failure mode 2: Confidentiality slips

What happens

What you paste into an AI tool may leave your organisation. The destination depends on the tool, the account type, and the configuration. Public chatbot accounts often use input data for training unless explicitly turned off. Enterprise accounts are usually configured to keep data inside your tenant, but "usually" is doing significant work in that sentence. Meeting AI in particular captures comments people did not realise were being captured.

Why it happens

AI tools need data to function. Free and consumer products often default to retaining inputs for ongoing improvement. Enterprise tools default to not retaining, but the difference is in account type and configuration, not in the tool's name.

Practical defences

  • Know what your account does with your inputs. Check the data-handling terms. For enterprise accounts, your IT or legal team has likely reviewed them already; get the summary.
  • Apply a paste test. Before pasting anything, ask: would I be comfortable if this content appeared in another organisation's training data? If not, do not paste it. Faster than reading 40 pages of terms.
  • Classify before you paste. Know your data classification level (public, internal, confidential, restricted). Restricted data should not enter any AI tool.
  • Use enterprise tools for sensitive work. The marginal cost is small; the marginal protection is significant.
  • For meeting AI, consent and configuration are both required: consent from participants before the meeting, and configuration to control what the AI does with recordings afterwards. Both.

Failure mode 3: The composite voice

What happens

Conversational AI, by default, writes for a composite reader who does not exist. The tone is fluent, polished, neutral, and indistinguishable from every other AI-assisted output on the internet. This is the failure mode most practitioners notice last, because it does not look like an error. It looks like adequate writing.

Why it matters

Content that sounds like everyone else's content does not work in comms, where voice and audience specificity are what make the work land. The composite voice undermines exactly the thing comms practitioners are paid to produce: a recognisable, contextually appropriate, human voice.

Practical defences

  • Brief the AI with examples of your voice. Three to five samples of on-voice writing beat fifty adjectives. Most conversational AIs follow examples better than descriptions.
  • Brief the AI with a specific audience, not a generic one. "Write for senior public servants in regulatory affairs" beats "write for professionals."
  • Use the AI for ideation and structure. Do the voice work yourself. Many strong writers use AI for the outline and the rough draft, then rewrite for voice.
  • Read your output aloud. If it sounds like AI when you read it aloud, it sounds like AI when your audience reads it.
  • Closing this gap is the work of Week 2. Provenance and audience specificity are the two reflexes that put your voice back in.
Before you read the worked examples, hold one question in mind. Of the three failure modes, which is your work most exposed to right now? Think about your own work in the last month, not your organisation. The answer is the failure mode you will most want to design defences against when you build your shortlist.