What are your views on AI grading, it’s accuracy and consistency?
Will this affect the likes of PSA and Beckett?
My perspective is that it’s a natural progression with the technology improving.
Do I think it’s there now? Absolutely not.
However, businesses would love to reduce staffing costs and assuming it improves it’ll be far more objective in assessment.
It’s definitely not there now though, this is the kind of development which could leave companies not adapting in the past.
1 Like
I dont think AI will make much of a dent on Psa’s grading process (getting rid of the human aspect). Then again, maybe I am a robot in disguise 
1 Like
Do you not think that if AI picks up flaws that human eye cannot. If it were to grade top level, surely an AI card would command a fee potentially higher than Black Label?
Black labels are more flex culture than anything. At that level, does it really matter how much cleaner the card is? I’d also argue that ai picking up flaws no one can see doesnt really matter. If we cant see them in the first place, why place a premium on cards without these flaws?
Ai is useful for some things, but I dont ever see it replacing the human element of grading, nor do I see any sort of premium from cards graded from it.
1 Like
If it was that advanced it’d need to be based on a % score rather than 1-10.
No one wants to miss out on a 10 because a fancy camera picked something up that the human eye can’t even see in the final result.
It would be very difficult to select cards to send in if you can’t do an assessment yourself.
3 Likes
What intrigues me a lot more than full AI grading (which usually is just a load of hype from people who don’t actually understand how AI works) is AI or computer-assisted grading. I believe PSA has already implemented this in some ways.
I think actually assigning the grade to a card is always going to be something done by a human, outside of gimmicky companies like AGS. But computers can do things like match scans/color readouts of a submitted card to authentic examples in a database, reducing time spent authenticating. Computers/AI could also be used to examine centering and provide an exact measurement to the grader, look for any dents or creases, and catalogue print errors and variations.
I see the future of AI/computers in grading to mostly be in streamlining the information available to the grader when they make their grading decision.
8 Likes
I agree with others here that even if AI could theoretically pick up on flaws not visible or just barely visible to the human eye, does it really matter? In the end, the product is to be enjoyed by a person, and I dont think this “AI black label” is going to carry much of a premium over a “normal” black label or PSA 10 because of microscopic flaws that the buyer can’t even perceive.
But then again, there is the current black label market, so who knows 
Absolutely, the “eye appeal” aspect quoted in PSA’s grading protocol will only ever be a subjective human decision. Inb4 sentient subjective AI.
Generally though I enjoy the idea that other humans have examined my cards in person and passed a standardised(ish) judgment over them that, hopefully, agrees with mine. It just feels more authentic.
1 Like
It’s interesting, because in my opinion we’re definitely able to do it now. The limiting factor isn’t ‘AI’ but rather creating a setup which takes ‘perfect’ photos of the cards at different angles, different lighting in an efficient, quality and repeatable process. If enough effort was put into it I think it could be done quite easily. Comment above on producing a % rather than a grade - it would of course be possible, especially since it’s likely the method chosen would actually first generate a single continuous number which is then scaled & bucketed into the 1-10 grading system we’re familiar with. But honestly just because it can be done doesn’t mean it necessarily should.
For the record, I think this should happen, but I’m a little biased. (EDIT: this refers to AI grading, not a % scale)
2 Likes
I think it will be inevitable, but I can imagine a future with a cheaper “AI submission tier”, “human check” tier in the middle and the most expensive “100% human manual review” tier.
I’m sure at first people would be upset, with Ai causing some comical miss grading (for example mistaking a print defect for a scratch, and so on), and maybe some minor grading company will try to be the “still 100% human one”. But quickly all the majority of people will be ok with this and it will be a staple.
1 Like
I had a thought the other day while scrolling a thread on here, and I can’t remember what thread it was (think it was something about print quality and people arguing about scratches and lines). Anyways, I’ve never graded a card but really interested to hear how others would respond that do collect graded.
Okay here goes, be kind:
Imagine if AI was used to make grading relative and adaptive. You submit a card and it gets authenticated and assigned a cert number… but that is all the label contains.
An AI has graded the card and you find the grade by typing your cert number into the PSA website. Now, here is the fun/crazy/ridiculous/amazing part…depending on how you feel. The grade is relative and can change based on what is submitted. You have a PSA 10, but then someone submits an even cleaner version of the card. When you type your cert number in you see you’ve been bumped down to a 9.7 grade. 
AI is used to measure absolutely every minuscule part of the card and assigns it a condition based on its results in comparison to other cards of the same variety. So a card notoriously difficult to achieve a 10 because of a bad print run would theoretically have a 10, but not based on standard 10 across the board metrics. all other would be graded relative to the worst and best cards submitted.
How much do people hate this?
Artificial intelligence and machine/deep learning models are only as good as their training and validation datasets. There are millions of cards out there when you consider the intellectual property or sport, size of the card, color palette/variants, etc. and finally, card condition.
It is possible that companies have started to compile datasets of scanned images to classify cards (i.e., from what set or year is this?), but I highly doubt that they have used these algorithms accurately for grading. Surface is just one area that a machine would need multiple images with different lighting/direction to accurately discern, and even then, it may not fully understand texturized or embossed cards vs. dents/damage.
When you get into the weeds here, it looks much murkier than the average person is claiming. Also, most “AI” discussed here is not really AI, but convolutional neural networks (CNN) that use training and validation datasets to visually discriminate between images. Those types of neural nets are helpful (their utilization in medicine is groundbreaking), but they aren’t likely to be helpful enough for actual grading (yet).
2 Likes
If I understand your idea well, it would be some kind of “percentile” grading, where having a 10 means having the best specimen (or one of the best) of all the submitted one at this specific time.
Really interesting, pretty dystopic lol, but cool nonetheless.
2 Likes
First off, it will NOT be actual AI (or AGI). Any AI grading will basically be one of two things:
- A comparison to an average based on previous examples available to the AI model, such as
all these are examples of 10s, 9s, 8s, etc, so this card is most similar to ____.
- Examine the features of the card, and compare them to a pre-set matrix of features, and grading parameters. (basically a more exact way of doing what is done now.)
This card has these key features (dents, scratches, edge whichtening, etc) and with these data points, it most closely matches the pre set parameters for ___ grade.
Ultimately, AI grading would be no different than human grading, in the first point. If it’s based on examples of previous grades, then there’s a big range. MIGHT even be Worse. 
The second point is probably what would happen, but this wouldn’t be so good, I think, since it totally upends the previous grades, and has absolutely no history or precedent. It would not be a “grade” so much as a measurement. This card has these data points. Perhaps a machine graded perfect 10 would have some significance, but the other points… ?
Imagine, “I’m sorry. Your card is off by .005 mm and does not qualify as Gem,” or “Your card has a nearly invisible-to-the-eye unknown defect on the top right corner and cannot be a Mint specimen.”
True AI or General AI, what many call “AGI”, (I don’t know why they used a tax term when “TAI” or “GAI” would be good) would be more like the first point, however, it’s not needed to grade cards. Consider that it’s theoretically, a person, so in order for a True AI to grade cards, it would have to have a desire to spend its time grading cards for $25 an hour (IDK what they pay), when it could probably be solving world hunger, breaking the great firewall of China, or something.
edited for clarity
1 Like
This is possible without AI even now. You just scale grades based on a full database of grades a given card has received (which PSA has access to, for example). You’d of course need to remove the label with the number and replace it with just the QR code. Once AI comes in and can provide a continuous scale for condition, we’d then go from ‘9/10’ to ‘second best condition in the world for this card’.
This conversation pops up a lot, but I don’t think AI grading its practical yet. When human grading costs what $.25-.50 per unit? Even if you had an accurate grading model, a human would still need to take the card out, load it/out of the machine, wait time… I don’t think there would any $ savings. So many other places in the production line that have better ROI for automation/AI.
Please PSA give us an AI/automated submission form.
I head an AI group for what it’s worth.
1 Like
I think the most realistically outcome with it in the future would be PSA adopting it, thus eliminating any need ever for crossgrading, and they completely consume the entire hobby.
I think the only way it happens and happens successfully is if they do some sort of an era cut off point where all cards after a certain year are AI only and Vintage is still manually graded. For a while anyways. Otherwise it will basically kill the vintage part of the hobby.
I mean, this is a terribly performing model and not something that should ever surface.
If anyone’s interested in how a good performing system here would work I could note down some ramblings. But it’s definitely definitely definitely possible with today’s technology.
3 Likes