Use of artificial intelligence in evaluating gemstones

[Sydneyphoenix]

I thought I had not too badly discerning eyes looking for clarity-impacting inclusions or surface-reaching inclusions on faceted gemstones but A.I. like ChatGPT seem to be much better at picking out those features. Even when I passed it off as eye-clean without surface-reaching fissures the A.I. seem to be able to find those features on photos and videos. Has anyone else used A.I. To help assess the gemstones or jewellery pieces before buying and what are the experiences like? Are they really good at picking minute inclusions and fissures, or are they over-calling the inclusions which might be found with loupe but not naked eyes? Would love everyone’s experiences and perspectives.

 

[LilAlex]

There's no doubt that, in a controlled setting, AI could excel at evaluating photos of gems. (There are cool examples of high-profile science in which an AI algorithm learned to accurately recognize mouse facial expressions (!) -- whereas the world never knew this was even a thing.) The utility all depends on the training set and how accurately those examples are annotated in that data base -- and none of us can speak to that for a ChatGPT in this context. So I would not trust any current general AI model more than a PS audience for characterizing particular light and dark spots in a 2D gem photo.

I could easily see a GIA or similar, with their near-limitless data set of gems and 2D and 3D images, backstopping their color and clarity grading with an AI "checker" -- and then perhaps moving entirely in that direction. Yes, something would be lost in that process. Same goes for, say, coin-grading and even real-estate pricing.

For a while, there were some cool AI-generated inspo photos appearing here and I wouldn't mind seeing more of those -- "15mm blue/green black opal with a halo of 3mm MRB diamonds, all in rose gold" etc. Like all of us, I'm really good at recognizing what I like when it's right in front of me but, unlike many of you, I have a hard time starting with a blank canvas!

 

[Autumn in New England]

I'm all for anything that helps us confirm composition and detect treatments (including degrees of clarity enhancement). And for diamonds, especially, this could be helpful in identifying and understanding internal characteristics a bit more thoroughly. But for me personally, I'm louping my CS at 10x and moving on.

Slightly O/T, but this reminds me of something that happened with my dentist. He got a new "cavity detector device." He placed it on each tooth to detect decay. Previously, he had always done this by examining my teeth manually... never had a cavity in my life. Suddenly he "found" 5. The device was no doubt sold to him by a very crafty dental supply rep as a money-making tool. I refused service and asked him to never use the device on me again. It's been 15 years of going to the dentist twice per year, and no one ever found a cavity again.

I guess my point is, sometimes focus can become excessive. Or maybe I'm just rambling and still holding a grudge. Love my dentist otherwise, BTW... started going to him when my first baby teeth came in!

 

[Pomelo]

I thought I had not too badly discerning eyes looking for clarity-impacting inclusions or surface-reaching inclusions on faceted gemstones but A.I. like ChatGPT seem to be much better at picking out those features. Even when I passed it off as eye-clean without surface-reaching fissures the A.I. seem to be able to find those features on photos and videos. Has anyone else used A.I. To help assess the gemstones or jewellery pieces before buying and what are the experiences like? Are they really good at picking minute inclusions and fissures, or are they over-calling the inclusions which might be found with loupe but not naked eyes? Would love everyone’s experiences and perspectives.

This is such an interesting question! I’ve not used AI to assess a stone before buying it, but I’ll definitely give it a go.

I think where it’s helpful is when the stone doesn’t come with a cert and/or a plot chart. I guess this is especially pertinent to coloured stones whereas diamonds often have GIA certs with inclusion plots.

However, as @Autumn in New England illustrates nicely, I think AI could be “perfect being the enemy of good”: by picking up on defects that we otherwise wouldn’t notice, does it detract from our enjoyment of the stone? Does it make a lot more stones fall into the non-mind clean bucket?

 

[Karl_K]

Ai chat bots aka L.L.M. is just a small subset of AI or more properly Machine Learning programing.
A significant amount of diamonds graded at GIA are done by computer alone using (ML)Machine Learning programs(AI in marketing speak but not LLM).
Expecting a general purpose LLM to do this is not practical or possible.
A LLM may have some image analysis and generation code but they will not be a replacement for specialized programs with training sets tuned for those task.
This can also be transparent to the user with online hosted AI.
If the LLM determines the load is better suited for a different program it can hand it off and return the result transparent to the user.

 

[Garry H (Cut Nut)]

Ai chat bots aka L.L.M. is just a small subset of AI or more properly Machine Learning programing.
A significant amount of diamonds graded at GIA are done by computer alone using (ML)Machine Learning programs(AI in marketing speak but not LLM).
Expecting a general purpose LLM to do this is not practical or possible.
A LLM may have some image analysis and generation code but they will not be a replacement for specialized programs with training sets tuned for those task.
This can also be transparent to the user with online hosted AI.
If the LLM determines the load is better suited for a different program it can hand it off and return the result transparent to the user.

GIA and IBM Research Join Forces to Transform Diamond Grading

CARLSBAD, Calif. – July 27, 2020 – GIA and IBM Research have joined forces to develop an advanced artificial intelligence (AI) system designed to revolutionize diamond clarity grading.

www.gia.edu

 

[PrecisionGem]

I sent a few images and weights to ChatGPT. It did a pretty good job of identifying the gems. I really liked the values it gave for the gems.

 

[Sydneyphoenix]

Thank you all.

Yes it is great for confirming inclusions or surface-reaching fissures we suspect but seem to be over-zealous in picking up minute inclusions or fissures that even the best trained eyes among us would not pick up without loupe. Personally I don’t care for loupe-clean as long as reasonably eye-clean so the AI indeed have danger of putting perfectly good stones into not-mind-clean basket unnecessarily. @Autumn in New England makes a great point about over-identification of problems when none exist.

I guess it is certainly a great tool in assessing gemstones that do not have certificates from reputable labs, in verifying vendor’s claims about the stone’s properties. So use it as part of preliminary research before you decide to buy.

 

[glitterata]

I agree that someone could train a machine-learning model to identify inclusions, but ChatGPT is the wrong tool. When I tried it, it coughed up a lot of nonsense about reflections and carbon, but it failed to properly identify reflections and carbon. When I gave it the dimensions of a star sapphire and asked it to calculate the carat weight , it neglected to include the specific gravity of corundum and came up with a nonsense number. When I told it to include the sg, it came up with a whole page of nonsense forulae and spit out another nonsense number. It's looking at all the combinations of words in its database and offering you some word combinations it calculates are similar to the target combinations; it has no idea what you're asking and whether its answers are right (or even what "right" means).

 

[Garry H (Cut Nut)]

I agree that someone could train a machine-learning model to identify inclusions, but ChatGPT is the wrong tool. When I tried it, it coughed up a lot of nonsense about reflections and carbon, but it failed to properly identify reflections and carbon. When I gave it the dimensions of a star sapphire and asked it to calculate the carat weight , it neglected to include the specific gravity of corundum and came up with a nonsense number. When I told it to include the sg, it came up with a whole page of nonsense forulae and spit out another nonsense number. It's looking at all the combinations of words in its database and offering you some word combinations it calculates are similar to the target combinations; it has no idea what you're asking and whether its answers are right (or even what "right" means).

When you finally correct the AI do you think it reprograms itself for the next similar question?

 

[Sydneyphoenix]

I agree that someone could train a machine-learning model to identify inclusions, but ChatGPT is the wrong tool. When I tried it, it coughed up a lot of nonsense about reflections and carbon, but it failed to properly identify reflections and carbon. When I gave it the dimensions of a star sapphire and asked it to calculate the carat weight , it neglected to include the specific gravity of corundum and came up with a nonsense number. When I told it to include the sg, it came up with a whole page of nonsense forulae and spit out another nonsense number. It's looking at all the combinations of words in its database and offering you some word combinations it calculates are similar to the target combinations; it has no idea what you're asking and whether its answers are right (or even what "right" means).

Yeah interesting that when I put the same video of the same ruby with exactly same prompts copied for grading and valuation twice, it provides slightly different dimensions estimates and wording, though same valuation point. Something’s not quite right in terms of internal consistency…

 

[glitterata]

When you finally correct the AI do you think it reprograms itself for the next similar question?

Possibly. But who's to say my corrections (or anybody else's) are accurate?

 

[LilAlex]

Yeah interesting that when I put the same video of the same ruby with exactly same prompts copied for grading and valuation twice, it provides slightly different dimensions estimates and wording, though same valuation point. Something’s not quite right in terms of internal consistency…

There are some hilarious memes where folks have asked ChatGPT (or similar) to generate an exact copy of a user-supplied image, and then the user iterates this process 100 times. The hallucination that ultimately emerges is completely wild!

 

[Karl_K]

When you finally correct the AI do you think it reprograms itself for the next similar question?

There may be exceptions but this is true of all the ones I know about.
If you use the payed version it will use your correction in response to you alone in future sessions.
It would not use it to answer anyone else.
The free version does not use the content of past chats in the current one so the correction means nothing.

 

[Garry H (Cut Nut)]

Possibly. But who's to say my corrections (or anybody else's) are accurate?

Mine are always 101% correct