ChatGPT as a darkroom tool

[benderscameracorner]

Personally I enjoy my darkroom time too much to spoil it by bringing in unnecessary technology Having the Massive Dev timer on my phone is about as far as I need to go.

You understand that there are free AI models in which the adage "the advice given is worth exactly what you paid for it" is very true; and commercial subscription based models which are increasingly being used at a professional level in many fields?

I know personally that some being used in engineering design are very good indeed, when driven by a qualified person in the field.

This is very true. In our engineering firm we have a corporate account (required to keep our IP private and track everything) and there's a significant difference in the, well shall we say competency of the paid business version, versus a free account. I've used the latter to verify some fairly middling electrical engineering calcs and its inability to get basic technical math correct reliably is a worry. I always love the response "oh, you're absolutely correct..." when you respond to its previous answer with a correction. Thanks buddy...

Lack of access to copyrighted Standards really does limit its usefulness, admittedly.

Credit where credit due, it is somewhat decent at coming up with complex or lengthy Excel tables. This summer I'm hoping to code a basic iPhone app with its help... that'll be an interesting test.

 

[MattKing]

I muse about this while thinking about much earlier experiences trying to bring newer technology to bear on older challenges - in particular my early efforts while reasonably successfully partially automating some custom legal document generation for my business use.
Colouring that musing is some of my marveling at being at the discussion ringside during some of the backroom discussions about how Photrio keeps running.
And the available, extremely computer nerdiness of some of the AI tools that are clearly simplifying and refining that process make it clear that those tools, when employed by knowledgeable hands - not mine - present great possibilities for huge improvements.
So I'm keen to see what those who have those knowledgeable hands might do with the melding of the creative and the logical/objective that my favourite avocation - photography - presents.
It reminds me at least slightly about what Ansel Adams et al tried to do with the Zone System.
That and some of the early technology like preview monitors and auto colour balance in the photofinishing world.
I think it is a mistake though to look at the sort of "replacement for an encyclopedia" applications of AI as being what we should expect of it.

 

[loccdor]

You understand that there are free AI models in which the adage "the advice given is worth exactly what you paid for it" is very true; and commercial subscription based models which are increasingly being used at a professional level in many fields?

I know personally that some being used in engineering design are very good indeed, when driven by a qualified person in the field.

Yes, and the one which gave me a false traffic ticket was doubtless in that suite of professional products. So is the one which replaced the human in the customer service situation and made it impossible to talk to someone to get that erroneous charge resolved. So is the product that flags my wife's college coursework as AI-generated: all professional, paid-for products.

In the original post I stated about 5% of uses are very positive, I think that's what we're talking about with your second sentence. I never denied that there is some successful application of AI.

 

[koraks]

People, the thread is about AI and its application in darkroom work. There are plenty of places to have a critical discussion on other applications and other technical & ethical issues; feel free to make a thread in e.g. the Lounge.

 

[xkaes]

It reminds me at least slightly about what Ansel Adams et al tried to do with the Zone System.

OMG, AI and the Zone system. The latter is confusing enough for most people without throwing AI into the mix. Talk about an opportunity for hallucinations!

Then throw in a good dose of darkroom practices? That's enough to put anyone in the Funny Farm.

 

[warden]

OMG, AI and the Zone system. The latter is confusing enough for most people without throwing AI into the mix.

Well you could always have AI deliver its advice in a rare endangered language to complete the trifecta. Tanema, for instance.

 

[Jan de Jong]

Another example how I use AI:
I wanted to see if walnut leaves could be used for anthotypes, so I started with the simplest approach: boiling the leaves in water to make an extract. The result was a strong brown solution that looked promising as a coating.

Normally, anthotypes bleach in sunlight, so I expected fading. But the coated paper behaved differently — instead of losing density, the exposed areas darkened. That was unusual enough that I asked AI about it.

Right away, it pointed me to juglone, the main compound in walnut leaves, and explained its quinone‑like structure. Quinones are known for their photo-chemical reactivity, and that gave me the clue: juglone doesn’t bleach like most plant dyes, it actually deepens under UV.

So the combination of hands‑on trial and AI’s chemical knowledge made the discovery clear. It wasn’t about AI giving me a magic answer — it was about connecting the dots faster and understanding why the process behaved differently.

This is a simple example but shows how AI can assist, the knowledge, facts come from the aggregate knowledge, the creativity and innovative combinations have to come from the human.

¨

 

[xkaes]

It's worked for me -- AKA, helped me out -- sometimes, but so far, more often than not, I'm asking it a pretty obscure question, and because it lacks enough information, it comes up with the best it can offer -- and it's wrong. Sometimes just a little, sometimes a lot.

So at this point, everything it comes up with, I end up having to verify by other means -- AKA, doing double duty!

 

[Jan de Jong]

It's worked for me -- AKA, helped me out -- sometimes, but so far, more often than not, I'm asking it a pretty obscure question, and because it lacks enough information, it comes up with the best it can offer -- and it's wrong. Sometimes just a little, sometimes a lot.

So at this point, everything it comes up with, I end up having to verify by other means -- AKA, doing double duty!

I’m lost here — what you wrote doesn’t connect to AI in the darkroom. Maybe you’re not using it at all? In that case, your comment is more about avoiding AI than applying it.

 

[gbroadbridge]

I’m lost here — what you wrote doesn’t connect to AI in the darkroom. Maybe you’re not using it at all? In that case, your comment is more about avoiding AI than applying it.

The human was hallucinating, not the AI.

 

[xkaes]

I’m lost here — what you wrote doesn’t connect to AI in the darkroom. Maybe you’re not using it at all? In that case, your comment is more about avoiding AI than applying it.

Maybe you missed my earlier post about AI thinking that my Yashica 15mm enlarging lens was a fisheye lens. I also suggested, like others, that AI in the darkroom can be helpful for neophytes, but is less helpful -- and even a waste of time -- for more seasoned darkroom users.

If you expect me and others to only write about how great AI is in the darkroom, you're barking up the wrong Forum.

 

[Jan de Jong]

Maybe you missed my earlier post about AI thinking that my Yashica 15mm enlarging lens was a fisheye lens. I also suggested, like others, that AI in the darkroom can be helpful for neophytes, but is less helpful -- and even a waste of time -- for more seasoned darkroom users.

If you expect me and others to only write about how great AI is in the darkroom, you're barking up the wrong Forum.

Ah, I see now — you're contributing to the thread by explaining how you didn’t find a use for AI in the darkroom, and offering guidelines on who should or shouldn’t use it. Apparently, my 47 years of darkroom experience disqualify me. I also understand that on Photrio, replying to the original question is less important than reinforcing a preferred narrative.

I have in previous posts tried to add some examples where it helped me and how I use it, besides the OP I have not seen many contributions to the question, I am really interested in that though.

 

[reddesert]

You can think of machine learning algorithms as, essentially, having learned a grammar from their training material (it can be a language grammar, or the equivalent kind of structure/internal logic for some other domain - it doesn't necessarily have to be words). The degree to which the ML algorithms can then produce recognizable results depends on how well the learning lends itself to constructing new output. This helps understand where ML/AI succeeds and where it can fall down.

For example, LLMs are now very good at constructing grammatically correct language output, so it sounds like you're talking to something intelligent. They are also very good at outputting computer code, which is in hindsight less surprising because software has an extremely structured grammar - the MLs have been trained on a lot of presumably working code, which has to obey the grammar. (This formulation isn't original to me, but distills remarks by Paul Kedrosky in an interview with Paul Krugman. The full interview is off-topic, but you can find it on Krugman's substack if you want.)

Anyway, that gives some insight into both the strengths and drawbacks. I am impressed that you can upload a test strip and it will make suggestions that aren't off-base. I don't think I would use it for something that has to be quantitatively correct, like the amounts in a developer formula - I would rather get that out of a book (the OP did say they were also using books). Another question is at what point it becomes a crutch. For example, although the fact that it can read a test strip print is impressive, that's also something that a human can learn pretty quickly from their own strips. It may be tending to push the exposures toward the mean, which is good for making a print that has an even range of light to dark values, (or to give ChatGPT more credit, it probably "understands" typical image properties, like that sky is usually lighter than foreground, and so on). However, not every print needs to be centered on Zone V or whatever. This gets back to the point about grammars: the grammar of visual and aesthetic representation is much fuzzier to define than, say, programming code.

 

[koraks]

AI in the darkroom can be helpful for neophytes, but is less helpful -- and even a waste of time -- for more seasoned darkroom users.

I wonder (doubt) if that's true. It might be if you use it as a kind of search engine where you ask a question and expect a definitive answer that doesn't require any further work. I think many people don't use AI that way and in a more 'fuzzy' setting a like a darkroom, it may not yield very usable responses if taken at face value. A different approach is nicely illustrated by @Jan de Jong and uses AI more like a virtual sparring partner to bounce ideas off of. I myself also do this (sometimes also photo-related, but not necessarily darkroom), where I ask questions, follow-up questions, and explore a subject. What AI typically does is offer several routes or options that you can then explore and/or further analyze. I find it's far quicker to make Ai burp up a list of options/alternatives than to manually assemble it. The latter may still be necessary to an extent; e.g. extending an AI-generated list of options with obvious options that it has missed, especially if you can reasonably expect there to be other options on conceptual grounds.

Maybe one of the crucial notions is a distinction between how people expect AI to work. Some use it interactively as I've described above; those seem to be the people who are reasonably (or even very) happy with it. Those who expect it will give hard & fast answers all the time, every time will likely always remain critical. Underlying this is the apt comment above from @reddesert: "LLMs are now very good at constructing grammatically correct language output, so it sounds like you're talking to something intelligent." LLM's aren't really intelligent in the sense of being capable of causation or conceptual soundness. They seem intelligent, but are inherently limited in some ways. I think if you're aware of those limitations, they're a powerful tool.

AI really is like a very complex hammer. If you need something to drive nails into a board, you'll find it works quite nicely. It may even work quite nicely for some other purposes, like gently nudging something back into shape - but only if used judiciously. But if you want to drive a screw into a metal housing and you expect that tapping it hard with the hammer will work - yeah, that'll be a disappointment for sure.

Much of the concern and criticism I read about using AI seems to boil down to this pattern of expectations. Using AI may currently be deceptively simple. Perhaps that's a bit of a pitfall.