AI error, hallucination and failure-state checklist

Know your errors

• The team has listed what counts as an error for this feature from the user's point of view, not only the model's.

  PAIR notes that people judge errors by their expectations: something working as designed can still feel like a failure. It calls these context errors. PAIR: Errors + Graceful Failure

• Each error type is traced to a likely source: training data, unexpected input, low relevance or confidence, or a conflict between connected systems.

  PAIR groups AI errors by source because the fix differs: better data, better input handling, or better fallbacks. PAIR: Errors + Graceful Failure

• Each error type is rated by stakes: what happens to the person if they act on it.

  PAIR asks teams to account for the situation's stakes. A wrong song is harmless; a wrong dosage, deadline or amount is not. PAIR: Errors + Graceful Failure

• The feature doesn't ask the model for facts it has no verified, current source for, and isn't used where a made-up answer could harm someone.

  Apple recommends avoiding requests for factual information unless the model has access to verified, up-to-date information, and avoiding AI-generated content where a hallucination could misinform and harm. Apple HIG: Generative AI

• Test sets include vague, badly phrased, out-of-scope, sensitive and deliberately harmful requests.

  Apple recommends testing exactly these cases and challenging your own policies and expected use cases. Apple HIG: Generative AI

Communicating uncertainty

• The interface says plainly that AI output can contain errors, close to where people read or use the output.

  Apple asks apps to clearly communicate that AI-generated content may contain errors. NN/g recommends placing disclaimers near the input and pairing them with an action. Apple HIG: Generative AI

• Uncertainty is signalled for the specific answer where possible, not only by a generic footer people stop reading.

  NN/g lists options: uncertain wording, the factors behind a prediction, confidence ratings, multiple responses, and visible sources, and warns that generic disclaimers are easily ignored. NN/g: AI Hallucinations: What Designers Need to Know

• Factual answers show their sources, and sources are easy to open and check.

  Sources let people verify a claim instead of trusting tone. Make checking cheap: link to the passage, not the site. NN/g: AI Chatbots Discourage Error Checking

• Formatting doesn't make weak answers look authoritative (confident headings, tables of invented figures, citations that don't support the claim).

  NN/g found that polished formatting and well-cited appearance lead people to over-trust output and skip checking it. NN/g: Explainable AI in Chat Interfaces

• When the request is ambiguous, the system asks a clarifying question or offers options instead of guessing.

  Microsoft's G10: scope services when in doubt; disambiguate or degrade gracefully when uncertain about the user's goal. Microsoft HAX: Guidelines for Human-AI Interaction

• Step-by-step “reasoning” shown to users is not presented as a reliable explanation of how the answer was produced.

  NN/g cautions that such walkthroughs are often rationalisations generated after the fact, not faithful accounts of the model's process. NN/g: Explainable AI in Chat Interfaces

Failure states to design

Each of these needs its own message and next step. A single “Something went wrong” covers none of them well.

• No answer: when the system can't find or produce a useful result, it says so, rather than producing a weak answer.

  PAIR calls these failstates and recommends giving people a path forward, including returning control to them. PAIR: Errors + Graceful Failure

• Blocked or refused requests explain, in plain words, what can't be done and suggest a request that would work.

  Apple recommends coaching people to be more successful next time, and offering example requests that lead to better results. Apple HIG: Generative AI

• Slow responses: there is a loading state that says what is happening, and people can keep working or cancel.

  Apple notes generative models take longer and suggests designing a loading experience or generating in the background. Apple HIG: Generative AI

• Timeouts and dropped connections keep the person's input and any partial output, and offer Retry.

  Losing a long prompt or a half-written draft turns a temporary fault into lost work.

• Partial results are labelled as partial (for example a stopped answer or a summary that skipped a file it couldn't read).

  An incomplete answer that looks complete is a silent error.

• Unavailable: when the model or provider is down, unsupported on the device, or the person is offline, the feature says so and the non-AI path still works.

  Apple notes models may be unavailable because of device compatibility, network access or battery, and recommends a non-AI fallback where possible. Apple HIG: Generative AI

• Usage limits, quotas and input-length limits are shown before people hit them, and the message at the limit says when it resets or what to change.

  A limit people couldn't see coming reads as a broken product.

• There is a plan, and a reporting route, for harmful or offensive output that gets through.

  Apple notes it may not be possible to prevent every harmful outcome and recommends policies and evaluation to minimise them. Apple HIG: Generative AI

Recovery and correction

• Edit, Undo and Retry are next to the output, and an AI change to the person's content can be reverted.

  Apple and Microsoft (G9, efficient correction) both ask for easy ways to refine or revert results. Apple HIG: Generative AI

• When the AI fails, people can finish the task manually without starting again.

  PAIR recommends returning control to people when the AI fails and always providing a manual alternative. PAIR: Errors + Graceful Failure

• When people correct the AI, the interface shows the correction took effect.

  Apple recommends acknowledging when corrections take effect, which helps people build an accurate mental model. Apple HIG: Generative AI

• Where the stakes are high, there is a route to a person.

  An AI that can't help and can't hand over leaves people stuck at the moment they most need help.

Learning from errors

• People can report a wrong answer in one or two steps, from the answer itself.

  PAIR treats feedback on errors as a way for people to help improve the system. PAIR: Errors + Graceful Failure

• The team reviews error reports from several channels: in-product feedback, support tickets, social media, metrics and research.

  PAIR recommends monitoring multiple channels to find new errors as people use the product in ways you didn't test. PAIR: Errors + Graceful Failure

• Reported errors are added to the evaluation set, so a model or prompt change that brings them back is caught before release.

  Generative output can change with small input or model changes; a fixed test set is how you notice.

• When a model update changes behaviour noticeably, people are told.

  Microsoft's G18: notify users about changes; G14: update and adapt cautiously. Microsoft HAX: Guidelines for Human-AI Interaction