Comment by michaelt | Hacker News Reader

Comment by michaelt | Hacker News Reader

Parent Full thread

The most important thing to keep in mind, when considering AI images for helping to illustrate or explain the actual content, is that AI-generated images have a big trust problem.

Take a look at the Google Image Search results for 'Engine Diagram' https://www.google.com/search?q=Engine+Diagram&tbm=isch

In the first row I get two correct-looking diagrams, then a load of AI-generated ones. Three cylinder heads for four cylinders? Two crankshafts? Oil pan labelled Crankshaft? Oil pan labelled Connect in rod? A four-cylinder engine with seven valve springs? Parts labelled up to 39 but a 9-item key? Spark plug wires that loop back on themselves? A crank shaft where the pistons aren't offset in phase? Four-cylinder engines whose exhaust manifolds have three or five openings? Then in the second row, I get "Parts of a V6 engine" showing a four-cylinder engine - with an 'Exhaust malvol', an 'extaust manivol', an 'exsoust manivol' an 'oilolaivt manivol'

Savvy technical audiences have noticed this - and so when a diagram has a whiff of AI about it, they become extremely skeptical and distracted.

Personally I only use AI-generated diagrams very sparingly, and very carefully.

To be fair, the problems you describe existed before AI, though I concede it has gotten worse.

What baffles me is why anyone would think it a good idea to post and host these images? What’s that upside?

Same reason fishermen put baits on the hooks, or hunters make weird sounds: it helps attract unsuspecting victims closer, increasing the chance of extracting some value out of them.

> the problems you describe existed before AI

No, the problem of mass-produced, outright wrong, educational graphics flooding image-search results did not exist before 2023.

> To be fair, the problems you describe existed before AI

To an extent, I guess?

But in the past if one wanted a diagram and didn't care about copyright they'd grab an image like https://i0.wp.com/www.engineeringchoice.org/wp-content/uploa... and reuse it unmodified.

In the age of AI, one can hand it to an LLM and get https://mechaniclove.com/wp-content/uploads/2023/12/engine-p... in a different style. And with messed up labels and the wrong number of holes in the gaskets.

> To be fair, the problems you describe existed before AI, though I concede it has gotten worse.

I do think it's fair to say it's gotten a lot worse.

Before AI-generated imagery, an hilariously bad (yet human created) engine diagram as described above, might actually be shared among my friends for laughs. Because it didn't happen that often. Now it happens a lot more often, and they aren't human created (which was only funny for a very short while).

No human drawing an engine would have called an oil pan a crank shaft. A thing that in no way resembles a crank or a shaft.

Diagram-wise I had better luck in telling it to make a text version for diagram generator, not only AI edits tend to be more correct, it's easier to fix any errors too.

As a bonus you can then prompt AI to validate whether the diagram matches article and vice versa, I've caught some small bugs like that when we migrated our docs

> you can then prompt AI to validate whether the diagram matches article and vice versa

No repro with Nano Banana with an infographic prompt.

I used Claude code that I pointed at local Kroki instance for rendering, and made it note how to use it in CLAUDE.md for validating

Reply on news.ycombinator.com