Part 1 | Meeting generative AI
Chapter 4. Knowing what generative AI is good and bad at
Key points
- Generative AI is very strong at organising, summarising, explaining and translating
- It is weak at genuinely original ideas, and at guaranteeing accuracy
- Work on the assumption of hallucination: confident, well-formed wrong answers
You are allowed to ask the AI about the AI
The first step in getting good at this is
knowing what it is good at and what it is bad at
And the entertaining part is that you can ask it directly:
"What is generative AI good at, and what is it bad at?"
Using AI is less about learning the correct method first, and more about getting a feel for it through conversation.
What it is good at
Generative AI is quite strong at work like this:
- organising broad knowledge
- summarising long text
- explaining things clearly
- comparing information
- putting a vague unease into words
- translating and bridging between languages
Mapped onto therapy work, that means a good fit with:
- laying out an overview of a condition
- summarising papers
- drafting explanations for patients
- structuring slides for a study group or teaching material
- writing multiple-choice questions for exam preparation
It is also comparatively easy to organise an English paper in Japanese, or turn a Japanese clinical note into English. The barrier to information from abroad drops considerably.
It suits teaching especially well
Supporting teaching is a particular strength.
Ask
"Explain this so a middle-school student would understand it."
and you will get something genuinely broken down. Ask instead
"Organise this for specialists, using anatomical terminology."
and you get a specialist framing. In other words, it can shift the explanation instantly to fit
- the listener's level
- their age
- how much they already know
- their cultural background
which lines up closely with a therapist's work in patient education, student teaching and family explanations.
It is also easy to adjust as you go:
- "Simpler."
- "Add an analogy."
- "Use bullet points."
- "Warmer tone."
And what it is bad at
It is not all-purpose, of course.
- genuinely original ideas
- going from nothing to something
- leaps that cross contexts
- reading the fine texture of a particular relationship
- taking responsibility
are still not its strengths. It is better understood as something that
organises and recombines existing knowledge
Watch for hallucination
Because AI is probabilistically predicting the words likely to come next, it can produce, with total assurance:
- bluffing
- wrong information
- paper titles that do not exist
- guideline names that do not exist
This is called hallucination.
In other words, AI can be wrong with great confidence. So rather than
"the AI said it, so it must be right"
work on the basis of
"use it as reference; a human confirms it"
For the following, always confirm against primary sources:
- drug names and doses
- citations of specific papers or guidelines
- the fine detail of law or reimbursement
- medical judgement about an individual patient
The right distance
Generative AI works best when treated less as
"something to hand everything over to"
and more as
a thinking aid
Let it help you clear your head, get into a topic, gather material, draft and summarise — and let the final judgement stay with the human. That distance is the point.
A prompt to try
You are a specialist in generative AI. For therapists (physical, occupational, speech-language and so on) using generative AI in daily work, list the top five situations where it is especially useful and the top five where particular care is needed, as bullet points. For each item in the second list, give the reason why care is needed.
Chapter 4 summary
- Good at: organising, summarising, translating, explaining, simplifying
- Bad at: originality, guaranteeing accuracy, carrying final responsibility
- Assume it will be confidently wrong, and always have a human check