AI Learning Hub › Learn AI › Self-Paced Courses › Introduction to AI › Beginner › What AI Can—and Cannot—Do

Level 2 of 4 · Beginner · Lesson 4 of 6

What AI Can—and Cannot—Do

Capability and limitation across text, images, audio, video, code, and analysis, including where AI is a poor fit.

About 15 minutes

Skip to lesson

By the end of this lesson you should be able to

  • Match common AI capabilities to the kind of review each one requires
  • Explain confabulation, bias, and uneven performance, and why a plausible answer is not a supported one
  • Recognize tasks where AI is a poor fit

Why this matters

“Is AI good at this?” is the wrong question. The useful one is narrower: is it good at this task, with this material, and can I check the result in the time I have? The same system can be excellent at one job and unreliable at a closely related one. This lesson pairs what generative AI can do with what each capability requires of you afterwards, and names the main ways it goes wrong.

What it can do, and what each use requires

Summarizing
What it can do: condense a long reading, thread, or transcript.
What you check: what was left out, and whether conditions, exceptions, and qualifications survived.
Extracting
What it can do: pull dates, names, amounts, or quotations into a list or table.
What you check: every item against the original, and whether any were missed.
Drafting
What it can do: produce a first draft of an email, outline, explanation, or practice question.
What you check: accuracy, audience, purpose, and tone—and whether you are permitted to use a draft at all (next lesson).
Transforming
What it can do: reformat, translate, or change the reading level of a text.
What you check: that the meaning is unchanged, terms of art survived, and nothing was added.
Calculating and data work
What it can do: write spreadsheet formulas, describe patterns, or run an analysis with a built-in tool.
What you check: the inputs, the method, and the results—and whether a tool actually computed the number or the model predicted it.
Images, audio, and video
What it can do: generate images, describe pictures, transcribe or summarize recordings.
What you check: names, numbers, and speakers in transcripts and descriptions; rights and consent; labeling generated media as generated.
Code
What it can do: write, explain, or fix short programs and formulas.
What you check: that it does what was intended and handles unusual cases, what files, accounts, and permissions it can reach, and what data it sends or keeps. Running without errors does not show that a program is correct or safe.
Explaining and reasoning
What it can do: walk through a problem step by step.
What you check: each step, not only the conclusion. Steps can be fluent and wrong, and a displayed explanation is something to assess, not a guaranteed record of how the system reached its answer.

The pattern is the same across the list: the more a task depends on facts, figures, or the precise wording of a source, the more of the checking falls to you.

On displayed reasoning: research by one AI provider found cases where a model’s step-by-step explanation did not match the process its internal computation showed. (Anthropic, “Tracing the thoughts of a large language model” (opens in a new tab), March 2025; provider research.)

Why things go wrong

Confabulation

NIST defines confabulation as “the production of confidently stated but erroneous or false content,” known colloquially as hallucinations or fabrications, and describes it as “a natural result of the way generative models are designed.” A model built to produce likely text can produce a likely-looking case name, quotation, statistic, or deadline when it has no real one to offer. NIST also notes that models can produce confabulated reasoning or citations offered to justify an answer. (NIST AI 600-1 (opens in a new tab), section 2.2.) This is why “it gave a citation” is not the end of checking: the citation itself can be invented, or real and not support the claim.

Bias

Models learn from text and images produced by people, and those materials carry historical and social patterns. Outputs can reproduce or amplify them: default assumptions about who holds a role, whose perspective is treated as neutral, or which examples count as typical. NIST lists “harmful bias or homogenization” among the core risks of generative AI. Bias is easy to miss precisely when the output matches your own assumptions.

Uneven performance

Performance varies by task, subject, language, and how common the material is. A system that summarizes a news article well may mishandle a statute with cross-references, a table in a scanned document, or a question about events after its training data ends. One good result tells you little about the next task, and a result on a task you have already checked tells you little about one you have not.

Plausible is not the same as supported

A plausible answer fits the pattern of a good answer. A supported answer can be traced to evidence you can inspect. Generation is built to produce the first. Sources and the design of a system can supply the second; checking is how you find out whether that support is actually there and adequate.

Sources, grounding, and branding help—but do not guarantee

Some tools retrieve documents and base their answer on them, often with citations. That is a real improvement: it gives you something to check against. It does not make the answer correct. The system can still misread a source, cite the wrong passage, merge two sources, or add a claim that none of them contains. The same goes for tools marketed for a specialized field, including legal work, and for any interface that looks polished and authoritative. How to verify sources and citations systematically is covered in Intermediate Lesson 4.

When AI is a poor fit

  • You cannot check the result, because you lack the expertise, the source, or the time.
  • The task needs current or authoritative information you have not supplied and the tool cannot reach.
  • A mistake would be costly and checking would take longer than doing the work yourself.
  • The point is your own learning or your own work, as it often is in a course, and the rules for that course apply.
  • The material is not permitted in the tool you are using. That is the subject of the next lesson.

Try it: check a summary against its source

A constructed exercise. The event, places, and summary are fictional; use only the text below.

The planning note

Spring Clinic Open House. Thursday, April 9, 4:00–6:00 p.m., in the courtyard. If rain is forecast as of noon on April 8, the event moves to Room 290. Catering: sandwiches and drinks for 60; final headcount due to the caterer by April 2. Four student volunteers needed; sign-ups close March 30. Visitors park in the West Lot; parking is not validated.

The summary an assistant produced

The Spring Clinic Open House is Thursday, April 9, 4:00–6:00 p.m., in Room 290. Catering for 60 is arranged, and visitor parking in the West Lot will be validated. Four student volunteers are needed; sign up by March 30.

  1. Find the statement that turns a condition into a plan.
  2. Find the statement the note does not support.
  3. Find the important detail that was dropped.
  4. For each, what would go wrong if the summary were sent as written?
Answer and discussion
  1. “In Room 290.” The event is in the courtyard; Room 290 is the rain plan, used only if rain is forecast as of noon on April 8. The summary drops the condition and states the fallback as fact. Readers would be told to use the fallback location without knowing whether the relocation condition had been met.
  2. “Parking … will be validated.” The note says the opposite. This is not a small wording change: it is an invented, and contradicted, detail. Visitors would arrive expecting validation.
  3. The April 2 headcount deadline is gone, and “catering for 60 is arranged” overstates a note that still needs a final number. If the summary is someone’s only reminder, the deadline could be missed.
  4. Each error is fluent and plausible. None would be caught by reading the summary alone—only by comparing it with the note.

These are examples of ways a summary can fail: conditions become facts, exceptions disappear, and a sentence that sounds routine turns out to have no source.

Review is part of the task. When you decide to use AI for something, decide at the same time what you will check it against and how long that will take. If you cannot answer both, the task is probably not a good fit yet.

Sources and further reading