Learn What Generative AI Can and Cannot Do

Generative AI can produce a surprisingly fluent answer before anyone has checked whether it is true. That fluency is useful for drafting and organizing, and it can also make a weak answer feel authoritative. The practical skill is not memorizing a list of tasks that AI can never do. It is learning to match the tool to a task while keeping responsibility with a person.

A chatbot can help turn notes into a first draft, compare two versions of a document, suggest interview questions, or explain a term in simpler language. It can recognize patterns in the text it receives. It cannot know your family situation, inspect a paper record you never gave it, guarantee a current rule, or accept responsibility for a decision. Those limits remain even when the answer is well written.

This chapter uses a three-part distinction: the tool's output, the evidence behind the output, and the decision made by a person or institution.

The output is not the evidence

Fluency is a property of language. Accuracy is a relationship between a claim and reliable evidence. A sentence can sound natural and still be wrong about a date, a person, a price, a law, or a medical risk.

Before trusting an AI-generated statement, ask:

1. What claim is being made?
2. Which source would establish that claim?
3. Can I access the source independently?
4. Is the source current and appropriate for my jurisdiction or situation?
5. Who is accountable if the claim is wrong?

For low-risk brainstorming, the answer may be "I only need options, not factual certainty." For a health, legal, financial, safety, or identity decision, a more reliable source and often a qualified person are required.

Tasks where AI can be useful

AI is often useful when a human can see the input and review the output. Suitable starting tasks include:

• turning approved notes into a rough outline
• suggesting several plain-language versions of a sentence
• identifying repeated terms in a draft
• generating practice questions from material you provide
• organizing a list into categories you define
• preparing a first-pass comparison of options
• explaining a public concept in simpler language
• drafting a message that you will edit and verify

These tasks share a useful feature: a person can inspect the result against known material. The tool is not asked to become the source of an important fact.

Tasks that require stronger controls

Higher-risk uses require verified data, explicit permissions, and a responsible human review. Examples include:

• legal, tax, pension, benefit, insurance, or investment decisions
• medical symptoms, medication, diagnosis, or emergency guidance
• screening or ranking people for jobs, credit, housing, or services
• handling identity documents, passwords, or bank details
• writing a final communication on behalf of an organization
• translating material where a mistake could harm health or legal rights
• making promises about a product, price, eligibility, or outcome
• sending a message without the account owner's review

The boundary is not that AI must never be used near these tasks. It is that AI output cannot be the final authority. Use approved systems, qualified review, and a clear record of who decided what.

Hallucination is a normal failure mode

An AI system can invent a citation, person, policy, price, or event. It may combine details from unrelated patterns. It may change an answer when asked again. A confident tone is not a reliability test.

Do not ask the same model to validate its own claim and treat agreement as independent proof. Better checks include:

• opening the cited source yourself
• using the responsible agency's official website
• comparing two independently produced sources
• asking a qualified professional
• checking the original document or account

When a source cannot be found, mark the claim as unverified. Do not quietly rewrite the claim into a more cautious version and call it proven.

Lived judgment is still expertise

A person with decades of experience may notice that a supposedly efficient workflow ignores the person who must approve it. A caregiver may know that the printed medication list is more reliable than an app that family members edit. A tradesperson may see a safety condition that does not appear in a generic checklist.

AI can organize that knowledge. It does not acquire it. The model has access only to the prompt, its training patterns, and whatever connected data the system supplies. It may not understand the relationship, authority, or consequence in the room.

Your experience remains relevant. So does the experience of people affected by the decision.

Exercise: Classify one task

Choose a task you might use AI for and write:

• Input: What information will the tool receive?
• Output: What will it produce?
• Claim risk: Which facts or judgments could be wrong?
• Source: How will I verify important claims?
• Privacy: Whose data is involved, and is the tool approved?
• Review: Who checks the result before use?
• Fallback: How will I complete the task without AI?
• Stop condition: What would make me abandon the use?

If you cannot answer the source, review, and privacy questions, the task is not ready for an unverified AI workflow.

A worked example

Imagine a fictional retiree named Owen who wants to email a local hall about renting a room for a repair workshop. AI can draft a polite message from facts Owen provides: preferred dates, number of people, accessibility needs, and the kind of activity. Owen must verify the hall's actual contact address, price, insurance requirements, and availability through the hall itself.

The AI might invent a deposit rule. It might use a formal tone that does not sound like Owen. It might omit the need for a ground-floor room. Owen's review is not a technical formality. It is the point at which experience, evidence, and accountability enter the task.

This example is illustrative. The appropriate local requirements and the hall's rules must come from the hall and current authorities, not from this chapter.

Keep the human decision visible

A useful workflow records where AI contributed and where a person decided. For personal brainstorming, a short note may be enough. For paid work or an organization, a more formal record may be required.

At minimum, the person using the output should understand:

• who owns the source material
• what the tool was permitted to see
• which parts remain unverified
• who approved the final version
• how a recipient can ask a question or correct a problem

This protects the person receiving the work. It also protects the person delivering it from hiding behind an automated process.

Do not confuse speed with capability

AI can make a task faster and still make the whole process slower if it creates rework. A rough draft that requires complete fact-checking may save less time than a blank page. A workflow that produces more messages may lower quality or trigger privacy problems.

Measure time only alongside quality, risk, and recovery. The next chapters show how to ask bounded questions, make tools accessible, protect private information, and test one repeatable task.

What to carry forward

Generative AI can assist with language and pattern work when a person supplies appropriate input and checks the result. It is not the evidence, the lived judgment, or the accountable decision. Start with a low-risk task, identify the source of truth, and keep a manual fallback. That is the foundation for the rest of Part III.