The goal of this tutorial is to:
- Get you using a chatbot for common business use cases
- Identify common issues with accuracy you should keep in mind while
using GenAI
- Provide some tips for effective prompting
Data Security and Privacy
Before we begin, it is very important to
remember:
- Public chatbots may use data you upload in responses from future
chatbots
- You shouldn’t upload any data to an AI service without company
approval, especially:
- Internal company data
- Customer data
- Check with your IT department for approved AI services and what data
is appropriate to upload to them
Open a new chat
- Open chatgpt.com or a similar
tool
- We’ll start working in a new chat
Brainstorming and Drafting
Let’s start by using the chatbot for some brainstorming for kicking
off a new project at work.
Type in the following request - what’s referred to as a
prompt:
Write an agenda for a project planning meeting
- AUDIENCE QUESTION: Have a quick skim, any feedback
on the response?
Some possible observations to discuss:
It might provide suggestions for follow-up prompts at the end of
the response.
It’s quite verbose, but we can ask to make it shorter:
Make it shorter
- It “remembers” the history of the chat, so we don’t need to provide
the original prompt again.
It’s quite generic, it would help to be more specific about our
project:
Make it specific to a project to {AUDIENCE CHOICE; e.g. launch a new website}
Chatbots can be a great source of inspiration when brainstorming:
- They can draw from ideas in a massive amount of text scraped from
the internet and other sources.
- You still need to review the content and decide whether the ideas
are relevant.
- Sometimes coming up with your own ideas first helps you better think
through a topic.
- You can use a chatbot to provide a second opinion to see if there’s
anything you’ve missed.
Similar to brainstorming, chatbots can also be a good way to make a
first draft of a document:
Turn this into a meeting invite
Chatbots are also quite good at mixing ideas together, like applying
a style of writing to a piece of text:
Say it like a pirate
Generating images
Many chatbots now also have built-in capabilities to generate images,
try asking it to draw a city skyline you know well:
Generate an image of the {city} skyline
This might take a while, so let’s come back to this later
Automating web searches
We can also use chatbots to search the web and summarise results for
us.
Let’s start a new chat for this so that:
- You can easily revisit specific chats in the history along the
sidebar.
- You avoid the chatbot getting confused by previous messages in the
chat.
Enable the web search option in the chat (sometimes
chatbots will do this automatically if needed), then prompt it to find
out about upcoming events you’re interested in:
What {event type e.g. tech conferences} are coming soon in {your city}?
- You can see a list of sources it used - it’s a good idea to
double-check these.
- This is effectively the same thing as Google’s “AI Summary” at the
top of search results.
- AUDIENCE QUESTION: Would you trust this answer to
be complete?
- Probably not, it’s only used a source or two.
- You can use Deep research (as a homework exercise)
to get it to search more thoroughly.
- These days, there is a risk that the sources themselves are
AI-generated web pages!
- If it’s important, you should do your own thorough review.
Reviewing the generated
image
Go back to your first chat with the skyline image.
- AUDIENCE QUESTION: How well did it do?
- Overall, usually a pretty good approximation
- But you’ll see that it will often make mistakes in the details
- Point out a few mistakes you can see in your image
Summarising Documents
- Now let’s try using the chatbot to summarise a document.
- Start a new chat for this
- We can upload word docs, PDFs, and more.
- Again, it is important to not upload company data into
unapproved systems
- Let’s use a spreadsheet (as a CSV file) of hotel reviews:
From technoon.org/ainoon, download hotel_reviews.csv
- Each row contains a (fake) review with an associated rating.
- The file was actually generated by ChatGPT 5 with the following
prompt: “Write 123 unique 3-sentence reviews for a hotel with associated
ratings out of 10 in a table”, along with some manual tweaks
- It actually returned 182 reviews, with quite a few
duplicates!
Upload the file to the chat
Prompt it with:
What are the common issues in these reviews?
- If we were running this hotel, it’s given us some issues to
investigate, but how do we know these are the most important, or that it
hasn’t missed any common issues?
- If we ask twice, we might not get the same result.
- If it’s used code to analyse the data (which ChatGPT often
does):
- We might get more consistency in results
- We can check the code to see if we trust it’s approach - but only
helpful if you can read code
- In general: good for getting ideas to investigate, but not
for formal reporting or decision making
What about asking it more detailed questions?
How many rows are in the file?
Should be 123
What is the average rating?
Should be approximately 6.5
- If it doesn’t use code to generate the answer, it is likely going to
be wrong because chatbots answer based on patterns in language - they’re
not a calculator!
- If it generates code to answer the question, it is more likely to
give a better answer, but the code should still be reviewed.
You can try force ChatGPT to not use code by saying “don’t use
code to answer”. But do this in a new chat so it can’t read the previous
code-based answer.
Prompting Tips
Refer to slide in the slide deck.
Conclusion
- We’ve seen how we can use a chatbot for a bunch of useful tasks:
- Brainstorming
- Drafting
- Web searching
- Generating images
- Summarising documents
- It’s quite useful, but it isn’t perfect
- It’s just a tool
- It doesn’t actually “think”
- You need to do the thinking!
- That’s why the logo for this course is just a hat
- Refer to “Golden Rule” slide in the slide
deck.
- Next week, we’ll learn more about how a chatbot generates responses,
which will help us better understand these limitations