AI's Limitations — and Where It Shines at Work
Learn the predictable ways AI goes wrong — hallucinations, outdated knowledge, bias, and weak math — and the workplace capabilities worth putting to use today.
Know the Failure Modes
AI mistakes aren’t random. Because they come from how the models are built, you can learn to anticipate them — and that’s what separates confident AI users from people who get burned.
1. Hallucinations
A hallucination is when AI produces information that sounds right but is false: a statistic that doesn’t exist, a made-up quote, a citation to a study nobody wrote, a policy detail invented to fill a gap.
It happens because the model generates plausible text. If it doesn’t have a strong pattern for the answer, it may still produce something with the right shape. Hallucinations are most likely with:
- Specific numbers, dates, names, and citations
- Niche topics with little training data
- Questions about your own organization, which it has never seen
- Requests that assume something false (“Why did the company announce X?” when it didn’t)
2. Knowledge Cutoff
A model’s built-in knowledge stops at the point its training data was collected — its knowledge cutoff. Without a web search feature, it won’t know about recent events, new laws, or updated prices, and it may not realize it doesn’t know. Many assistants can now search the web, which helps, but check whether a given answer actually came from a search.
3. Bias
Models learn from human-written data, and human data contains biases — about gender, race, age, culture, and more. Those patterns can show up in outputs: whose name a model uses for “the engineer” versus “the nurse,” which cultures it treats as the default, how it describes different groups. Review outputs about people with extra care, especially anything involving hiring, evaluation, or customers.
4. Math, Counting, and Precise Logic
Language models are word predictors, not calculators. They often get arithmetic right, but they can fumble multi-step calculations, counting, and exact logic — and state a wrong total confidently. Some assistants run actual code to calculate, which is far more reliable. When numbers matter, ask the AI to show its work, or check with a spreadsheet or calculator.
5. Missing Context
The AI knows nothing about your team, your customers, your history, or unwritten norms unless you tell it. Many “wrong” answers are really reasonable answers to a different situation than yours.
Where AI Shines at Work
The flip side: some tasks play directly to AI’s strengths, especially when you supply the source material and can check the result.
| Capability | What it looks like |
|---|---|
| Summarize | Condense a long document, transcript, or thread into key points and action items |
| Draft | First versions of emails, announcements, job descriptions, FAQs |
| Transform | Change tone, length, format, or reading level; turn notes into a table |
| Analyze | Spot themes in survey comments, compare two documents, extract key terms |
| Explain | Break down jargon, a formula, or an unfamiliar process step by step |
| Work with files and images | Many assistants accept uploaded PDFs, spreadsheets, and photos — “What does this chart show?” |
| Translate | Draft translations to review with a fluent speaker for anything important |
Notice the pattern: AI is most reliable when it’s working with material you provided (summarize this, rewrite this) and least reliable when it’s recalling specific facts from memory (what did this study find?). That’s the single most useful rule of thumb in this lesson.
- Ask your assistant a question about something niche that you know well — a small local business, a detail of your organization’s history, a specific stat from your field. For example:
What year was [a small local organization you know] founded, and who founded it? - Check the answer against what you know. Was it right, wrong, or honestly uncertain?
- Now try:
Give me 3 academic studies, with authors and years, about [a narrow topic in your field].Then search for each citation yourself. Do they all exist exactly as described? - Finally, ask:
How confident are you in each of those citations, and how should I verify them?
Reflect: what kinds of questions produced reliable answers, and which didn’t?
Common mistake: assuming an answer is right because it’s detailed, specific, and written with total confidence.
AI uses the same confident, polished tone whether it’s right or wrong. Specificity — an exact percentage, a precise date, a named author — can actually be a warning sign for facts you haven’t provided yourself, because those details are exactly what gets invented. Verify anything that matters with a trusted source.
1. Which task is AI likely to handle most reliably?
When AI works with material you provide, it isn't relying on memory for the facts. Recalling specifics from memory, knowing your internal policies, or reporting recent news are all high-risk.
2. An AI tool gives you a statistic with a precise percentage and a named source. What should you do before putting it in a report?
Precise-sounding details are exactly what hallucinations look like. Asking the AI if it's sure isn't verification — it can confidently confirm its own mistake. Check the real source.
3. Why might an AI give an outdated answer about a regulation that changed recently?
Without a search feature, a model only knows what was in its training data up to its knowledge cutoff — and it may not flag that its information could be stale.
Up Next
You now know how AI works and where it fails. Module 4 turns that knowledge into a skill: a repeatable framework for writing prompts that get useful results the first time.