SearchResearch (4/23/26): AI image gen tools are great, as long as you don't ask for accuracy

 Every few weeks... 

The wreck of the M/V Oduna on a remote beach on the
southeast side of Unimak Island, Alaska. P/C US Dep. Interior

... there's an advance in AI in one way or another. Maybe the text is smoother, or there are fewer hallucinations.  It's always something. 

Sometimes it's just a wreck.  

This past week it was OpenAI announcing a new and improved image generation model.  (Their press announcement of 2.0)  

And it's true--much has been improved. 

BUT... don't look for accuracy in the images, especially if you're creating a diagram to help you understand how something works.  

The little details count, and all of the AI image gen tools get details wrong.  

Looking back at our "tea kettle" Challenge from Nov, 2023 ("How does it work?")  I asked for a diagram about how a tea kettle operates.  The prompt was simply: [create a diagram of an electric steam kettle showing how it knows when to shut off when the water is boiling]  

Here's what I got from ChatGPT (2.0):  

Which isn't terrible... EXCEPT... there's no thermostat or temperature-sensing device shown.  It says there's a "bimetallic strip assembly," but most tea kettles use a circular bimetallic device to detect shutoff temps.  They look like this with the bimetallic device shown as a gold circle (illustration from above on the left; a side-view on the right).

More annoyingly, the "simplified circuit diagram" is nonsense. That's just two parallel switches side-by-side.  Both of them are open.  It's unclear how this helps the story of the diagram. 

Meanwhile, here's Google Gemini's version--very glossy, but missing in the important details: 

Here, the thermostat is shown in the handle (which is odd, they're usually in the bottom) and it's fairly clunky and cyberpunky in appearance.  

I'm not the only person pointing this out.  Gary Marcus managed to win with the most-ridiculous diagram by this week's OpenAI 2.0 model. Here he asked for a:  

P/C Gary Marcus and OpenAI

We can go on in this vein. 

But here's the big take-away for you:  Do NOT trust diagrams generated by AI systems (at least at the moment).  They often have obvious bugs (and even worse, sometimes subtle bugs).  

And don't get me started on Claude's image generation skills... 

It's good to know that Claude is really good at code generation, it's not going to make it as a technical illustrator.  

Back to our regular programming next week!