Key takeaways
- Generative AI fails in predictable ways — knowing the patterns lets you spot them faster.
- The most common failures are plausible falsehoods, outdated information, narrow reasoning, and quiet omissions.
- Human review is not a formality; it is the control that keeps specific errors from reaching real readers, customers, or decisions.
- A layered routine — set expectations, review against sources, and keep a named owner — catches most problems.
- No review process is perfect, so pair it with a way to correct mistakes when they surface.
AI tools produce their most impressive results and their most dangerous ones in the same way: by generating text that looks finished. A draft arrives structured, readable, and confident, and it is easy to send it onward. The discipline that separates safe use from risky use is understanding how these systems fail, and reviewing output with those specific failures in mind.
This article catalogs the main failure modes and offers a practical review routine. It is the companion to Fact-Checking AI-Generated Content, which focuses specifically on verifying claims, numbers, and citations, and to the wider Limitations & Verification section.
Why these tools fail in the first place
Generative models are built to predict likely text. They are very good at producing language that fits a pattern, and they have no built-in mechanism that separates "sounds right" from "is right." A model does not know when it is guessing. It generates the most plausible continuation it can, whether or not a reliable answer exists.
Two consequences follow. First, errors and correct answers arrive in the same tone, so fluency tells you nothing. Second, the system's confidence, such as it is, is not calibrated to reality — it can be as certain about a fabrication as about a well-known fact. Human review is the step that supplies the judgment the system lacks.
A catalog of common failure modes
Learning to name these failures helps you scan for them. Most AI mistakes fall into one of the following categories.
Plausible falsehood (hallucination)
The model states something false with full confidence: an invented statistic, a citation that does not exist, a rule that is not real, a quote that was never said. This is the best-known failure and often the most damaging, because fabricated specifics carry the authority of detail.
Outdated information
Unless a tool is connected to a live source, it reflects its training data, which has a cutoff. It may describe things as current that have since changed — a policy, a price, a leadership role, a product version — and it rarely volunteers that it is out of date.
Mixing up specifics
Names, numbers, and dates are easily swapped, blended, or misattributed. A model may correctly describe an event but attribute it to the wrong year, or correctly name a concept but credit the wrong person. The overall shape is right; the cost is in the details.
Overgeneralization and overreach
Asked to explain a topic, a model may flatten genuine complexity into tidy rules that are not true in all cases. It can turn "in many situations" into "always," and can offer confident advice in areas where the honest answer is "it depends."
Silent omissions
Summaries and answers can leave out material that matters — a caveat, an exception, a whole counterargument. Because the output is coherent, the missing piece is easy to miss. What is absent is often harder to notice than what is wrong.
Inherited bias
Models absorb patterns from their training material, including stereotypes and uneven representation. This is especially risky when AI output touches decisions about people, such as hiring, evaluation, or access to services. See our Responsible AI guidance for how to approach fairness.
Inconsistency between runs
Ask the same question twice and you may get two different answers, both stated confidently. If two runs disagree, at least one is likely wrong — and neither output reveals which.
Overconfidence in tone
Where a careful human would say "I'm not sure," a model may state a claim flatly. The absence of hedging is not evidence of accuracy; it is simply a style the tool learned.
Mental model
Picture a fast, well-read assistant who never checks sources and has no way to feel unsure. You would not hand their draft to a client unread. The faster they work, the more valuable — and the more necessary — your review becomes.
Why human review is the control that matters
Review is what converts a quick draft into something you can stand behind. It adds the qualities the model cannot supply on its own: grounding in reality, awareness of what is missing, and accountability for the result.
Review also catches a second class of problem that is easy to overlook — the problems AI creates in your process rather than in the text. Automating drafts can invite overtrust, where people stop reading carefully because the output usually looks fine. It can spread one unverified claim across many documents quickly. And it can blur ownership, so that no one feels responsible when something is wrong. A deliberate review step counters all three.
The goal is not to distrust every sentence. It is to match the depth of review to the stakes: light for a low-risk internal note, thorough for anything a customer, regulator, or broad audience will see.
A practical review routine
The following sequence works as a checklist for almost any AI-assisted output. Run it before anything is sent, filed, or published.
- Set expectations before you start. Decide who is responsible for the final result and how careful the review needs to be, given who will see it and what they might do with it.
- Read for claims first. Do a pass that flags every statement a reader could check. You are not editing yet; you are building a list of things that need a source.
- Check the specifics. Verify numbers, dates, names, titles, and citations against sources you can open. Treat any citation you cannot confirm as one to remove.
- Look for what is missing. Ask what a fair treatment would also include — the caveat, the exception, the opposing view — and add it if the draft omits it.
- Test the reasoning. For advice or conclusions, ask whether the recommendation actually follows from the evidence, or merely sounds reasonable.
- Check tone and fit. Confirm the output matches your voice and the situation, and that it does not overstate certainty the material does not support.
- Decide, then edit. For each item, choose to keep, revise, soften, or cut. Prefer honest hedging over confident phrasing you cannot back up.
- Keep a record and a name. Note what you verified and who signed off, so the work can be defended and corrected if needed.
Building review into the workflow
A routine only helps if it fits the way work actually moves. A few habits make review sustainable rather than heroic.
- Match review depth to risk. Rank tasks by consequence — internal brainstorm versus public announcement — and reserve heavy review for heavy stakes.
- Never let the tool be the final checker. Asking the same system to "verify" its own output can reproduce the same error. Use an independent source or a second person.
- Keep a source trail. Recording where facts came from makes later review, updates, and corrections far easier. The techniques in Organizing Research with AI apply here.
- Make corrections easy. A visible path for fixing mistakes, like the one in our corrections policy, limits the harm when something slips through.
- Review the process, not just the text. When an error reaches an audience, ask how it got past review, not only who typed the words.
Limitations of review itself
Honesty about the tool should extend to honesty about the safeguards. Review reduces risk but does not remove it.
Reviewers get tired, rushed, and anchored by a draft that already looks polished. Errors that survive a first pass can survive several. Review is also limited by knowledge: a reader cannot catch a wrong fact they do not already know is wrong, which is why independent sources matter more than a careful reread.
Review cannot resolve material that is inherently uncertain. Predictions, judgments, and contested questions do not become safe just because someone checked them; they need to be framed honestly as what they are. And no process, however careful, is fully reliable — which is exactly why verification and correction belong together.
Treat review as part of the work
AI is genuinely useful, and its usefulness grows when you understand its limits. The failure modes here are not reasons to avoid the technology; they are the map you use to navigate it. Know that confidence is not accuracy, that details drift, and that important omissions are easy to miss — and review accordingly.
Make review a named step with a clear owner, apply it consistently, and match its depth to the stakes. Continue with Fact-Checking AI-Generated Content for claim-by-claim verification, and Protecting Confidential Information for handling sensitive data safely. The site's Limitations & Verification and Responsible AI sections collect the rest of our guidance on using these tools with care.
Editor's note
This article is educational and describes AI limitations in general terms. It does not claim that any tool is always accurate or always wrong, and it is not professional advice. Specific products change frequently; verify current behavior and consult the relevant experts for high-stakes work. See our editorial standards.