Second test - following instructions

This page has 0 threads | Add post

Blogs

Many people around the world enjoy and benefit from our blogs. If you're one of them - and can afford to - please make a small donation to help keep this and future blogs, exercises, skills assessment tests, tips, tutorials and videos freely available to everyone!

Some other pages relevant to these blogs include:

Training courses in AI (Artificial Intelligence)

You can also book hourly online consultancy for your time zone with one of our 7 expert trainers!

Recomparing the main AI image creation tools 2 years on Part three of a five-part series of blogs
In this blog I re-run AI image creation comparison tests I performed way back in July 2024 to see how the results - and the tools which produce them - have changed. Recomparing the main AI image creation tools 2 years on First test - photo-realism Second test - following instructions (this blog) Third test - being creative My overall verdict on the results of the tests

Recomparing the main AI image creation tools 2 years on

Part three of a five-part series of blogs

In this blog I re-run AI image creation comparison tests I performed way back in July 2024 to see how the results - and the tools which produce them - have changed.

Recomparing the main AI image creation tools 2 years on
First test - photo-realism
Second test - following instructions (this blog)
Third test - being creative
My overall verdict on the results of the tests

Posted by Andy Brown on 18 May 2026

In this blog

What the test involves
July 2024 results
Today's results (May 2026)
Verdict for this test

This page gives the results of the second of the three tests run.

What the test involves

The challenge posed by the second test was to see if I could get an AI tool to recreate the sort of treasure island maps I used to doodle during lessons at school. Like this, in fact:

Art wasn't my best subject at school ...

Here's the prompt to create this:

Create a line drawing of a treasure island with a curvy coastline. The background of the island should be white apart from child-like drawings of the following features:

Hell’s Kitchen (a volcano)
Hangman’s lookout (a gibbet with a skeleton hanging from it)
Dead man’s desert
Sulphur Springs
Magwitch marshes
Tombs Town
Lone Palm
Treachery Hills

These features should be roughly equally spaced around the island. Do not show any text on the picture – I will add that later.

The coastline of the island should be clearly delineated, and child-like drawings appear in the sea of the following:

2 or 3 child-like drawings of whales
A line-drawing of a large bird like an albatross
A pirate ship in a bay

Keep everything simple. Avoid having much fill or background colouring. When in doubt, use simple outline drawings of things. Use pastel colours throughout.

The challenge is to see how well the AI tools can follow these instructions.

July 2024 results

Here were the best results from my original set of tests:

The one on the left was the best map from the original test; the one on the right was from re-running the test about a year ago, out of interest.

Today's results (May 2026)

Here's what I got when I ran the same test now:

AI tool	Time taken	Image generated
ChatGPT Plus	Under a minute
Claude Pro	About 90 seconds
Copilot Chat	Forever! See notes below.
Gemini Pro	About 40 seconds

Verdict for this test

OK, I'm regretting including Claude in this test. Here's what Claude claimed about its picture:

Really?

I also think I can draw whales better than this! Copilot was possibly even more annoying. It took at least 5 minutes (I confess I stopped timing) to come up with this message:

Seriously? And couldn't Copilot have told me this at the beginning?

I said I was happy for Copilot to make this change, and after a few more minutes got this:

OK, this is getting annoying.

I replied with this message:

Just give me a picture!

After nearly 3 minutes more, I finally got a picture:

Copilot has not only imposed ridiiculous safety constraints, but then ignored them in the final picture by drawing a skeleton hanging from a gibbet anyway!

Here's how the tools did (apart from Claude, which I refuse to even judge). I'm judging the Copilot picture through gritted teeth:

Request	ChatGPT	Copilot	Gemini
Hell’s Kitchen (a volcano)	Good	Good	Good
Hangman’s lookout	Good	OK	Good
Dead man’s desert	Good	Poor	Good
Sulphur Springs	Good	Poor	Good
Magwitch marshes	Good	Poor	Good
Tombs Town	OK	Poor	OK
Lone Palm	Good	Good	Good
Treachery Hills	Good	Good	Good
2 or 3 child-like drawings of whales	Good	OK	Good
A line-drawing of a large bird like an albatross	Good	Good	Good
A pirate ship in a bay	Good	Good	Good

Gemini gets points deducted, however, for labelling its drawings - despite my request not to do this.

This test had 2 clear winners - there wasn't much to choose between ChatGPT and Gemini - and 2 clear losers (Claude and Copilot).

Parts of this blog
Recomparing the main AI image creation tools 2 years on First test - photo-realism Second test - following instructions (this blog) Third test - being creative My overall verdict on the results of the tests

Some other pages relevant to these blogs include:

Training courses in AI (Artificial Intelligence)

You can also book hourly online consultancy for your time zone with one of our 7 expert trainers!

This blog has 0 threads

Add a new post

ALL BLOGS

ARTIFICIAL INTELLIGENCE BLOGS

ARTIFICIAL INTELLIGENCE (AI) BLOGS

What the test involves

July 2024 results

Today's results (May 2026)

Verdict for this test

Head office

London

Manchester