:-) Interesting AI hallucination ;-)

[Hert]

Screenshot 2025-04-10 124101.jpg
Looks more like a dinosaur too me
I know I cover Water Sports as well but its taking it a step to far!

What model identified this bird as a crocodile?

[Hert]

I found this photo, but have no idea what I'm looking at...OpenAI helped:

flower.png (635.14 KiB) Viewed 32107 times

I love to take pictures of flowers, but know little about them. Until I ran it through OpenAI (I wouldn't have even guessed it's a tulip):

tulips.png (766.12 KiB) Viewed 32106 times

AI is such a helpful tool

[Hert]

Summarized:

In all these samples, I ran the images through the AI service *without* processing the results. Still, it gave me a world of search data.

I can now search on wetsuit and find the statue image.
I can search on tulip and find the great parrot tulips.
I can search on gothic and find the image of the Trinity Church in New York.
I can search on velvet deer and find the image of the Père David's deer

And of course I can create all kinds of combinations eg pink tulip or feathered tulip or cultivated ruffled petals tulip outdoor

All without processing anything.

I could also use the AI results and start assigning catalog labels based on the descriptive text (which I would only do if I needed the keywords).

[andrew.heard]

Often the AI labels are very minor variations, just synonyms, of each other. Is there a fast way to ctrl+click any number of labels & mark as synonyms? The current method (context menu > Details > Synonyms > manually enter details) is far too tedious for large #s of AI labels.

[Hert]

It’s pointless trying to maintain the labels inside the AI Labels category. AI maintains these.

Either:
1. Keep AI Labels as-is (why would you)
2. Move (or merge) the labels you “like” to your Catalog structure, and then delete the left-over AI Labels category (pick your wins)
3. Completely delete the AI Labels category

Don’t blindly process the Ai results so you don’t end up with things you don’t want.
Either don’t process the suggested labels at all (as I tried to explain with all the samples above) or pick your winners in the results before processing.

As I stated a few posts up:

Nothing processed = Nothing to cleanup

Most important: use a good AI model to raise the quality of the results.

Ai will not take over your cataloging. It help you do the cataloging by providing input and suggestions.

[Hert]

And adding so many pedantic labels does nothing to enable me to filter my images.

Don’t blindly take over the suggested labels. You don’t even have to process the results at all!

Filter with Advanced Search and you can filter on free text or AI results.

Even without processing the results you get a world of additional search terms. Catalog labels are not your only search instrument.

[snowman1]

I found this photo, but have no idea what I'm looking at...OpenAI helped:
I love to take pictures of flowers, but know little about them. Until I ran it through OpenAI (I wouldn't have even guessed it's a tulip):

Off topic for the thread, so I don't want to start a discussion on it here - if anyone does let's start a separate thread - but I do recommend the Plantnet mobile phone app. It's excellent. (I read an article a while back in New Scientist that basically said most plant id apps tested were poor - but that Plantnet was the exception).

NB1: I have no idea how good OpenAI is, perhaps it's just as good!
NB2: I too know little about flowers but I like identifying ones I come across with the app! And when you dig a little deeper they all have their own back-story of what they were traditionally used for which I also find interesting, like stepping back in time sometimes.

[andrew.heard]

I mainly upgraded to 2025 because of the chance/ hope/ interest that AI may reduce my labelling & speed up searching, especially older thumbs that aren't well cataloged, or not at all, but I'm going to give it a miss now. When I get separate (Open)AI results, for example, Boulder, Rocks & Stones, all equally valid, it is far too tedious to "pick my wins". Maybe in future if I search for "Boulder", the AI will also know to include "Rocks" & "Stones".

[Hert]

I’ve been trying to highlight that AI-generated descriptions are what truly matter. That’s your real search data moving forward, not the suggested labels or keywords.

We need to move away from thinking that catalog labels are the primary tool for finding images. The real power lies in the search box and Advanced Search, and neither of those depends on catalog labels.

Historically, catalog labels were essential because they were the only way to describe and retrieve your images. That made sense…then.

But now, with AI, everything changes. Imagine running your entire image archive through AI: you’d be able to find almost anything you need, without assigning a single label. Less manual work, better search results. Ask yourself: are you cataloging to build a structured system or to actually find images?

That’s not to say catalog labels are useless. They are a precision tool: to shape search results and give you control over how specific data is represented.

But looking to the more distant future, we need to ask: will we even need keywords at all? They’re a holdover from a different era.

The future of cataloging and search lies in rich, AI-generated descriptions and in AI-powered tools that transform how we search and interact with content. As AI evolves, both the data and the search experience will become smarter, more intuitive, and more powerful.

So yes, I fully envision a future where searching for “rocks” effortlessly returns boulders, stones, and everything in between: automatically, based on the meaning in the descriptions, not just the exact words used.

[andrew.heard]

I'm happy for that "future" to come ASAP. In the meantime I made a copy of a "test" JPG (thumb attached), deleted the labels & description, and did an "Analyse with AI". The AI description then was

The image depicts a construction scene featuring an orange excavator positioned on uneven ground, with two workers standing nearby. One worker is dressed in a gray jacket and blue jeans, while the other wears a white shirt over a dark shirt and blue jeans. The excavator's arm is extended, and the site appears to be in the process of excavation or groundwork preparation, with large rocks and soil visible. The background consists of a forested area with tall trees, indicating a rural or semi-rural location. The excavator is a hydraulic type used for earthmoving tasks, characterized by its tracked chassis and articulated arm.

gobsmackingly absolutely amazing, to be true - 100% accurate.

But if I ignore all AI labels & do a search for "boulder" or "stones" rather than "rocks", the thumb still won't be found as it only mentions "rocks", so how does that help/ what am I missing? If it did help, is there a XMP place where I could move the my existing description(s), and allow the AI to overwrite the existing description(s) so I can retain both?

[Hert]

Double click on the title or description in the AI Panel to edit the text.

[Hert]

Use build 7632 or higher.
Reanalyze your image and search for boulder, or rock, or stone

[andrew.heard]

Wow - how does that magic work?

But if one thumb AI description includes "boulder", then another "rock" and another "stone", will a search for "boulder" now find all three thumbs? PS. I could test this myself - being lazy.

[Hert]

But if one thumb AI description includes "boulder", then another "rock" and another "stone", will a search for "boulder" now find all three thumbs? PS. I could test this myself - being lazy.

If a thumb shows a stone, and that stone is also a boulder then yes.
All boulders are stones, but not all stones are boulders.

If a thumb shows a rock, and that rock is also a boulder then yes.
All boulders are rocks, but not all rocks are boulders.