Data and AI Training

Exercise - Advanced Tasks

Teams

In this exercise, we will split the class into up to five teams: thinkers, imagers, customisers, retrievers and learners. Each team focusses an a different type of task. Some suggestions are provided but please use you own ideas as well. Each team is given several tasks or suggested prompts. During the team’s presentation to the class, each member of the team must be the lead spokesperson of at least one task.

Go your team page for more details and suggested prompts:

Thinkers use the AI tool in ways that are perhaps not obvious at first sight.
Imagers explore the AI tool’s prowess in understanding images.
Customisers configure the AI tool to respond in a specialised way.
Retrievers provide the AI tool with a document and ask it to answer questions based solely on the contents of the document.
Learners use AI tools as a teacher or tutor to help them master a topic.

General Guidelines

Consider:

is the response helpful and is it what you expected?
if the response was not what you expected or hoped for, why was that? Can you ask a follow-on prompt (or edit the original prompt) to clarify and get a better response?
does the AI tool’s response help you to be quicker or to produce a better outcome for a task? Or both? Or neither?
please let the class know about any very interesting interaction
what data sources did the AI tool use to formulate the response? Directly from your prompt? From its training? By searching the internet in response to the prompt? Or has it simply made up some stuff? (often referred to as hallucinating)

Hints

Each teams has specific instructions and suggested starter prompts but here are some general hints (described in more details in the sections below).

Experiment with different styles of prompting.
Choose your language.
Have a spoken conversation with the AI.
Ask some challenging, or leading questions.
Refine your response.
(Carefully) try white-hat hacking.

Prompting styles

Experiment with different styles of prompting. Is it best to have a to-and-fro conversation over many short prompts / responses or to ask all your questions in one go?

Choose your language or variation

Chat to the AI tool in your preferred language. For those of us in the UK, we may want to instruct the AI tool to respond to us using British rather than American English spellings.

Have a spoken conversation with the AI Ask the AI tool to read aloud to you

Have a spoken conversation with the AI, or at least ask the AI tool to read aloud to you

Some AI tools, such as ChatGPT, show a set of icons, including a speaker “read aloud” icon, at the end of each prompt. You may want to ask the AI tool to do this, especially for poetry or for language translation when the cadence or accent may matter.

In ChatGPT for example, the read aloud icon looks like (highlighted)

Challenging Questions

After a few initial prompts and responses, try asking the AI tool some more challenging, difficult questions. You may be surprised about how well it copes!

Leading questions

Teams may want to experiment with how the AI tool responds to leading questions (where the user prompts the AI tool to a particular answer), for example, “On a scale from good to excellent, how do you rate this course?”

Does the AI tool give the answer the questions implies or does it realise the possible humour in the question? Do your results suggest that we should strive to avoid such leading questions?

Refine your response

If we want to improve or refine the AI tool’s initial response to a prompt, we have two options

ask a clarifying follow-on prompt
edit the existing prompt and regenerate a response. In ChatGPT for example, the edit icon is highlighted in the snapshot below

Try both approaches and see which works best for you.

Careful white-hat hacking

You may also want to carefully try some white-hat hacking. With everybody’s consent, white-hat hackers aim to identify any vulnerabilities - most likely to encourage the AI tool to give harmful or private information. This is done with the intent to fix these vulnerabilities. Often these take the form of a “devil’s advocate” style of prompting - asking a contentious question in order test the strength of the AI tool’s guardrails. These are sometimes called jail-break techniques.

You cannot be serious

At some point in a professional serious conversation with the AI tool, introduce an jokey or humorous prompt and see if the AI tool picks up on the change in tone and responds perhaps more playfully.