Three entertaining examples of AI tools at their worst

This page has 0 threads | Add post

Blogs

Finding our blogs useful? Please consider helping us to share more!

3 good, 3 bad - the highs and lows of learning AI tools Part two of a three-part series of blogs
Learning to use AI tools has been a journey for us, although not one which we thought would take in broccoli cats and Neanderthal film reviews. In this blog we look at the 3 best and 3 worst things we've found when using AI tools like ChatGPT. 3 good, 3 bad - the highs and lows of learning AI tools Three entertaining examples of AI tools at their worst (this blog) Three examples of AI tools at their best, giving almost magical results

3 good, 3 bad - the highs and lows of learning AI tools

Part two of a three-part series of blogs

Learning to use AI tools has been a journey for us, although not one which we thought would take in broccoli cats and Neanderthal film reviews. In this blog we look at the 3 best and 3 worst things we've found when using AI tools like ChatGPT.

3 good, 3 bad - the highs and lows of learning AI tools
Three entertaining examples of AI tools at their worst (this blog)
Three examples of AI tools at their best, giving almost magical results

Posted by Andy Brown on 04 September 2024

In this blog

The AI version of a travel bingo card
Cryptic crosswords in the style of ChatGPT
The AI times table

This blog gives 3 examples of AI tools at their worst.

Let me stress: it's much harder making AI tools misbehave like this than it is getting them to give good results. In each of the examples below we've subtly loaded the dice against ChatGPT.

The AI version of a travel bingo card

AI tools seem to struggle to combine images and text. My favourite example is trying to create a travel bingo card using a prompt similar to this:

Create an image showing a travel bingo card for children aged between 6 and 12. The bingo card should contain 12 things that each person should try to see during a typical long car journey in the UK. The card should have 3 colums and 4 rows. Each cell in the card should have a large picture showing the thing that needs to be seen, and underneath this a smaller text description of what the picture represents.

Out of the 12 items, please create:

- 1 which is ridiculously easy to spot

- 4 which are would be very easy to spot

- 4 which you would typically see quite a few times on a car journey, but are a bit harder

- 2 which would be quite difficult to spot

- 1 made up thing which you could never see

Here's what ChatGPT just came up with:

The more you look at this, the more bizarre the answers appear!

It's worth trying this yourself as the answers can be very entertaining (although occasionally an AI tool will flummox you by getting everything just right).

Cryptic crosswords in the style of ChatGPT

Another thing that AI tools struggle with is cryptic crosswords (OK, I admit this is a niche subject). For example (and more or less at random) I tried this question:

Create cryptic crossword clues for the words "Cat", "Training" and "Numismatologist". Present your answers as a table with three columns: "Word", "Clue" and "Explanation"

Here are the answers given when I just ran this question through ChatGPT:

The 3 clues suggested.

These clues fail on every level: the surface reading of the clues is nonsensical, the explanations in most cases aren't true and not all of the definitions work either! Cryptic crossword setting is one job which should be safe from the predations of AI tools for some time to come.

The AI times table

A third example of something AI tools do badly is mental arithmetic. Consider this example:

Please calculate the product of the following three numbers:

- the number of lives a cat is said to have

- the number of faces on a dice and

- the number of players in a singles tennis match

What is this number? Please present your answer in the form:

The number you want is NNN

Do not include any other text in your answer.

On my ChatGPT just now this gave 288 as the (wrong) answer:

The answer should be 9 (lives) * 6 (faces) * 2 (singles players) = 108, not 288.

When challenged, ChatGPT breaks the question down and comes up with the correct answer:

I think it's time for a bit of humble pie - I've asked ChatGPT why it got the number wrong first time round.

AI tools can get apologetic when cornered!

Computer says ... sorry.

AI tools struggle when you don't give them a chance to do things step by step. By preventing ChatGPT from showing any other text we've trapped it into giving a false answer, and then into having to justiify this.

Parts of this blog
3 good, 3 bad - the highs and lows of learning AI tools Three entertaining examples of AI tools at their worst (this blog) Three examples of AI tools at their best, giving almost magical results

Some other pages relevant to the above blogs include:

AI (Artificial Intelligence) learning

This blog has 0 threads

Add a new post

BLOGS BY TOPIC

BLOGS BY AUTHOR

The AI version of a travel bingo card

Cryptic crosswords in the style of ChatGPT

The AI times table

Head office

London

Manchester