lel. OK hear me out.
Let’s assume that a company is able to accurately grade pokemon cards using “AI” with a very low defect rate (say a rate that matches PSA’s defect rate, yes, they aren’t infallible).
And let’s assume they get the economics right and are able to charge 20% less than PSA’s version of tiered pricing.
And let’s assume they announce this product to the market effectively and get the right people to say the right things.
ASSUMING all of this: do you think this company would be able to effectively compete with PSA? Because the only barrier at that point would be reputation I think.
You sound like an economist . It would prob look like CGC entering the market. They would take a decent chunk but not dethrone PSA. Also if a company successfully implemented this method, it would probably be PSA themselves.
In theory this sounds great. However PSA has the strongest inertia. They also could just acquire the technology.
I still think full AI grading has a long way to go. Also @sparky55603 point about cgc entering the market. They already changed their grading scale/standards. That is the main issue with perfect AI grading imo. Who decides the scale, and will it remain constant overtime, even if sales slump.
It’s probably inevitable. Question is, will it be one of the top company’s utilizing it?
can’t wait to hunt for the ultimate pristine AI transcendent gold label with 18/18 perfect subgrades
They probably won’t be able to ship the card back to you. What if one of the molecules were to shift during transit, invalidating one of the subgrades.
If the AI has the same defect rate as PSA whats the point of using it over PSA who has all of the years of inertia and higher value? Even if it was ‘perfect’ grading every single time, it would be a mammoth job to take over from a company so popular that their name is synonymous with a graded card. I don’t see it happening really
Here are the two hardest facts you’ll hear about AI grading
- it will happen
- you won’t notice
AI grading is virtually inevitable. The task of taking an image and classifying it on a scale from 1 to 10 is the most quintessential neural network task. I know there are some types of people who think AI will take over everything and that’s not my position. It’s just that grading is such a low-hanging fruit for a machine learning algorithm that it’s bound to happen. Some companies already claim to use it - though I remain skeptical about many of the claims.
The first idea people seem to trip over is that an AI would be “too good” and would be extremely harsh because a computer can see things a human can’t. This is an understandable intuition but it’s just not how it works. The simple explanation is that you train the algorithm to grade like a human so it would only learn to judge the features of a card that a human would.
The hard parts about creating a grading AI is collecting the data and training the algorithm. Why do you think PSA has started scanning every card (hmmm )? All this is to say that AI grading is both a reasonable thing that will eventually happen to this industry and I would argue is already in the R&D stages.
That being said, what are the consequences of having an AI grader?
There seems to be a lot of excitement about the topic but if you really think about it, there is not much exciting about it. Grading itself in it’s current state is basically sticking your cards in a black box and having a slab come out of the other end. You don’t know if Bob or Alice or a monkey or a pair of dice is grading your cards. And as long as the output is mostly consistent, you probably won’t care either! If you replace a mostly competent human with a 99.99% competent AI, you almost certainly wouldn’t notice unless you sent in cards on the scale of hundreds of submissions per year.
The benefits of AI grading really boil down to being able to fire the majority of your human grader staff. Maybe not a benefit to them, but to the company it means lower costs and they can sell their service for less (or increase profits). It may also mean higher throughput, so cards can be graded faster. The least important benefit is that card grades become (hopefully) more consistent and reliable (and well, the pop could become algorithmically fined-tuned if they want it to be). This is fine and dandy, but what does this all mean to the consumer? It means you might get cheaper grading with faster turnaround and probably an improvement in consistency that the average collector wouldn’t even notice.
It’s a long post to make two minor point, I know. To answer the question in the OP, if a company came out with AI grading that worked and tried to compete with PSA, it would probably look like this:
Which is something that both happened and you probably didn’t even notice
I’m not sure it will happen, at least in the way people think it will.
Backs, corners, edges, centering should be achievable. But the constant supply of new card arts makes it difficult to use AI grading for the entire card. New pictures, lines, drawings, patterns, textures, would need to be constantly updated into the model. Unless they figure out a way to make it agnostic to the picture/texture content and only identify the errors. That seems like very difficult model to build and maintain. Recognizing bubbles, bends, specs, dirt, and scratches on untrained art would be incredibly difficult.
I think one of the biggest areas we could see AI is in the intake process. AI scans, labels, and sorts the new cards as they come instead of using a human. Also, imagine not having to input your PSA submission by hand, but just snapping a picture of each card in an app and double checking the list!
Great point. The variability in art, holo patterns, textures and other forms of layout could be difficult. If there is a dirt looking spec as part of the artwork, how resilient is an AI to that?
I would assume they wouldn’t be very resilient. The models would have a harder time without previously learning what a clean copy looks like. Updating the model with new cards every release may help, but vision/image models are really difficult to build and update. In PSA’s case it may not be worth the risk for large amounts of blatant errors to slip through.
The grading part of the service is very cheap. You get less than a minute of a grader’s time per card. So maybe $.50 tops? I would imagine there are more costlier processes than the actual grading that could benefit from AI without the risk.
I don’t want AI grading. I don’t want grades that are dependent on minuscule microscopic imperfections completely invisible to the naked eye. I want a minimum wage worker’s biased opinion that my card is a PSA 10.
People are overlooking the obvious here — all grading is AI grading because the universe is a simulation.
So the total population of 10’s for any given card that is a 10 is predetermined?
Sounds like a scam to me…