AI is working its way into more and more aspects of daily life, and most discussions of card grading these days quickly turn to AI. This is especially true when the discussion revolves around alleged grader bias or human error.
So, what does the future of card grading look like in regards to AI? And what impact might those changes have on the millions of cards already graded by humans?
I talked with my pal (and Cardhound member) Brad Denenberg, who is currently waist-deep in AI grading software development. We chatted about questions like “What do we mean by AI, anyway?” and “What might card grading look like in 10 years?”
The future is impossible to predict but we can be certain that the presence of AI in card grading will continue to grow and evolve.
What is AI?
A first question to address is this: What do we mean by AI, anyway? Many of the recent upstart grading companies advertised AI overtly or subtly at launch, but the details never quite emerged. People seem to use the term AI as a trendy synonym for “technology.” But technology, broadly defined, has always been used in card grading.
IBM defines AI as “technology that enables computers and machines to simulate human intelligence and problem-solving capabilities.” So by this definition, scanning software that calculates exact ratios might be considered AI. But the simple application of “measurement” isn’t really AI and would likely assist a human rather than replace one. The question of what AI is (and what it is not) is not always easy to answer.
How Would AI Be Used in Card Grading?
I think the image most people have when they hear the term “AI grading” is feeding a card into a scanner. A few seconds later, presto, the card is graded. This AI grade could in theory include subgrades and a detailed report of every single card flaw. There’s also the idea that AI would be perfectly reliable, meaning that it would grade the same card the same way every time.
First, AI is already being used to assist humans in grading. PSA is allegedly using its Genamint software to fingerprint cards and detect frauds, for example. But the idea of a completely software-driven grading process seems a long way off. It is tough to say whether full AI grading is even desirable, given the current state of the technology.
But in practice, AI could be used in a host of ways. It could “assist” humans, or completely replace at least some human-driven decisions. For example, currently at most graders, a human assesses the card surface, assisted to some degree by technology, if not AI.
But AI could be trained to make this assessment independently instead. Here’s a sneak peek at Brad’s software in action:
Would AI grade the same card the same way every time?
This is the dream, right? Perfect consistency and reliability. No bias, no errors. Not so fast!
Upstart TAG Grading prefers the term “‘machine-learning’ over ‘AI’ when it comes to our describing our grading process and methodology.” But machine learning implies just that: humans continue to train and refine the software. This kind of process would result in the software evolving and improving over time. Improvement sounds good, yes? But this seems similar to how human grading has evolved and changed over time to become more exacting and strict.
Brad says “AI card grading could be designed in a way where the score would be evolving/learning over time, or in a way in which the score would remain constant. How it is actually implemented comes down to what collectors demand in different scenarios.” So, would customers accept “machine learning”? The concept seems to defy the calls for AI grading to eliminate human “inconsistency.” But what if the AI just isn’t very good at first?
Brad continues: “If collectors demand incremental improvements, then a grader could theoretically continue to train their models, and grades would change over time as more data is included in the training model. In most cases though, once the baseline is achieved, additional improvements wouldn’t be necessary.” My thought is: if the idea is that the AI is perfect and infinitely reliable on day 1, it better be great out of the gate. To me, that seems unlikely–at least any time soon. So if the market demands perfect reliability, there might be a tradeoff.
Implications for Vintage
Would AI affect every sector in the same way? Maybe, but it seems like a primary impact would be on those coveted GEM grade assessments–and there aren’t many vintage gems. It’s very difficult for most casual collectors and even many experienced ones to differentiate a 9 from a 10 with the naked eye. But the difference between a 3 and a 4 is typically more obvious. Presumably, AI could speed up the process, but the end results would not be too shocking. After all, the idea is to train AI to replicate the current human-driven process of applying the grader’s scale.
Some other impacts could include:
- Theft deterrent (fingerprinting)
- Crack / resubmit deterrent (fingerprinting)
- Fraud / counterfeit deterrent
- Improved speed / turn times
- Grading fee cuts
AI for Grading (and in general): A Long Way to Go
Clearly, AI is in its infancy. And I think the reputation of AI for card grading has been hurt by some bad early rollouts. I tried an app I won’t name here, and after I provided good front and back images of a card with massive paper loss, the AI grade was a “7.” Useless!
My college students use Chat GPT like crazy to cheat on their writing assignments, and the results are often comical–but always recognizable (so far).
And when I asked AI to generate an image for this article, it provided me with this inexplicable gem:
But then there’s Genamint, and its AI innovations that were touted 4+ years ago. Surely, PSA has continued to refine the massive potential for good AI.
What might grading look like in 10 years?
Obviously, the tech will grow up. As it does so, what are the implications for card grading?
Of course, one possibility is that it looks pretty much the same as it looks now. After all, PSA holds most of the cards–pun intended–and has a legacy of volume to uphold in terms of value. Would AI diminish the value of the tens of millions of cards already graded? And would collectors revolt, or would the resubmissions turn into another cash cow? My first thought is that AI will evolve to assist human graders in ways that might be invisible to collectors.
Brad has the following thoughts on what this future might look like: “I anticipate that grading will be vastly different in 10 years. Collectors are increasingly demanding more transparency, so I anticipate grading companies will not only be forced to include subgrades, but also to allow users to see detailed measurements and heatmaps of surface damage. The technology for this already exists and is in place behind the scenes at the top 3 or 4 grading companies, so I believe it’s just a matter of time. I also expect grading companies to begin tracking a specific card’s history and making that data public, including how many times a card has been graded (including cracking/resubmitting), and known fraud related to that card, and even the transaction history and provenance.”
The Human Element
Cardhound has already written about “eye appeal,” and PSA’s statement regrading the human element in card grading.This seems like at least a nod to the status quo: a human-heavy grading process. “Eye appeal” is perhaps the trickiest business of card grading–though Genamint claimed to have reduced it to an algorithm. Can AI ever truly replace humans in the actual grading process?
Brad’s thoughts: “I don’t believe we will see the market adopt fully-automated AI grading any time soon. Instead, AI-assisted grading will become the norm for a few different reasons. First, there are 100 million graded cards in the secondary market graded by PSA alone, representing many billions of dollars in gross market value. Grading standards have changed over the years, and the current process is subjective and incredibly inconsistent.
Too many collectors hold ‘over-graded’ cards, and a sudden move toward higher accuracy would devalue many cards graded a decade or two ago. Furthermore, for many years, PSA has had a financial guarantee, whereby if a card was erroneously awarded the grade, PSA will pay the difference in market value. Moving towards a fully automated, AI-based system (and removing the human subjectivity) would expose PSA to enormous liability.”
This potential liability seems tempered by the fact that PSA is both judge and jury regarding whether to enforce its own guarantee.
In Conclusion
Perhaps we raised more questions that we answered here. But if we reached anything resembling solid conclusions, here they are:
- The role of AI in card grading will continue to evolve and grow
- Upstart AI graders will have a hard time breaking into a legacy-driven industry
- For the foreseeable future, the role of AI will be to assist humans rather than replace them.
Other thoughts? Drop them in the comments! And if you like reading about Vintage, please Join Cardhound!
