By Austin Purvis for CoinWeek …..
Collectors and dealers have traditionally needed experience, knowledge, and a keen eye to accurately identify and grade coins. While this is still true, new tools and technology are constantly being developed and improved that may impact the hobby significantly. One of those technologies is “Artificial Intelligence”, or AI.
The company Next Vision Limited is an AI technology group that developed the mobile collecting app CoinSnap. CoinSnap is described on the app store as a tool “that utilizes AI-driven image recognition technology to accurately identify any coin” and “has a coin grading feature and gives a reference price for each coin.” At the time of publication, the app has over one million downloads on the Google Play store and is the #3 reference app on Apple.
How accurate is this AI tool? We used the coin collecting app for an afternoon, and here is what we found.
How CoinSnap Works
After opening the app, users are asked to sign up for a subscription but they can still use the app on a free seven-day trial period. Note: you will need to cancel your trial before it’s over to avoid being charged for the application. For this test, the free version was used.
At the main menu, users can click a camera button at the bottom of their screen to immediately open the identification feature. Additionally, there are “Grade my coin” and “Identify my coin” options at the top of the screen. Users are asked to take obverse and reverse pictures of their coin or upload an image from their phone gallery. The app will then match these images to the information on their database to provide a result. The app offers an overview of the coin, including its name, country of origin, and year of issue, among other important details.
The grading option works exactly the same as the identification option, but the results also include a grade that the AI has determined the coin falls under based on the image provided. Note: this result does not include a number along with the grade. For example, a user would be shown an estimated grade of “MS” (Mint State), not “MS-65”. Users are given a price range that the coin may fall under and are prompted to provide mint mark and date information to receive a more accurate price estimate.
A Note on Privacy While Using CoinSnap
Since this application requires user information and content uploads, anyone who is interested in trying this application should be aware of and read the Terms of Service and Privacy documents. Below is a brief analysis but not a comprehensive overview of these policies.
While it reads like many user agreements one would typically see on applications, users should be aware of the following:
The Terms of Services agreement states that when using the application, you grant Next Vision “a worldwide, non-exclusive, irrevocable, royalty-free, fully-paid, perpetual, sublicensable (through multiple tiers), fully transferable license to use, distribute, reproduce, create derivative works from, publish, translate, publicly perform and publicly display any Content that you upload or publish.” Effectively, you do not have exclusive ownership rights to the photographs of coins you submit to this application. That is an extremely important detail for dealers or collectors who may have paid for high-quality photographs of coins to consider before they submit those same images to this application.
It is also part of a larger trend of developers teaching their AI to work better (and therefore make them more money) using datasets of other people’s information without fair compensation in the currently under-regulated AI field. Just look at news reports of how image-rendering AI is exploiting copyrighted material from artists for an indication of where this could be going. Use your own better judgment, but you should be doing that whenever you use apps on your phone anyway.
As of the publication of this article, there are no news reports or press releases that would indicate any issues with Next Vision Limited’s data security. On the CoinSnap website FAQ, Next Vision answers the question “Is my information safe?” with the following response:
“We are cautious about sharing personal information. Although we require access to your photo gallery, we never process any photos without your permission. Rest assured that we will NEVER share your information with anyone.”
Putting an AI Coin Collecting App to the Test
To test the app ourselves, we scanned 15 coins: 10 uncertified and five certified. Each coin was photographed in the same set-up, and each coin was scanned three times. This test could benefit from more coins to create a larger sample size, and we already have plans to test this app further.
The uncertified group included the following coins: U.S. coins in circulation (a penny, a nickel, a dime, a quarter); a U.S. Presidential $1 coin; a South Korean won; an Australian silver dollar; an Eastern Caribbean 25-cent piece; and a commemorative United States half dollar.
Four of the certified coins were all graded MS and included a Buffalo nickel; a Morgan dollar; a Peace dollar; and a Mercury dime. The final coin was an ancient Roman T. Quinctius Flamininus silver denarius graded XF. In total, we tested three coins certified by PCGS and two graded by NGC.
CoinSnap was able to accurately identify 14 out of the 15 coins provided, coming to 93% accuracy. The coin it was unable to identify was the single ancient coin, graded by NGC. To explore this further, we took ancient coin images from our latest article from Mike Markowitz and submitted them. The app was not able to accurately identify any of these additional coins. The app provided a list of potential matches for these coins, spanning from other ancient coins that were incorrect to a Canadian one-cent piece.
While anecdotal, these results seem to suggest the ancient coin category within the AI model needs additional training.
As for the grading tool, CoinSnap was able to accurately determine the MS grade of both the Morgan and the Peace dollar but it provided contradictory grades for the Mercury dime and the Buffalo nickel (AU and XF). No grading results for the ancient coin were obtained because it could not identify the coin. For this modest test group, only 40% were graded correctly.
Can the CoinSnap AI Accurately Identify and Grade Coins?
Based on these results, it seems that this AI collecting app can be extremely accurate when identifying many coins but significantly less so when faced with ancient coins and niche examples. Especially with popular series/types of U.S. and world coins, CoinSnap should be able to let you know what type of coin you have. Additionally, users should be aware that CoinSnap does not support the identification of errors or counterfeits at this time.
As for grading, there are a couple of factors to consider. First and foremost, many aspects of coin grading can be subjective, so translating what an experienced dealer, grader, or collector may value into a specific AI model could be difficult. One grader may interpret a detail on a coin differently than another. There has also been debate among experts on certain coins as to whether they are over-graded or under-graded. While certain objective measures of quality exist that a computer may be able to identify, the nuances of grading may be lost when asking an artificially intelligent program of the current generation to do it.
Photo quality is also a huge factor when using this app.
While quality was controlled for this test, the ability of the AI to identify details on the coin naturally depends on the quality of the photo. This means the grade offered by the app may be more of a reflection of your photography skills and/or the quality of your camera than the actual grade of the coin.
Another idea to consider is the strength of the AI model that CoinSnap is utilizing to identify different coins or grade coins. AI programs are “trained” by being given data (photographs, text, etc.). They use these data points to better understand something, like how the obverse and reverse of a coin may look (with different dates, conditions, et al.). Using this data, the application can compare user-provided input against the database of information it was “trained” on. The more examples an AI has been given, the better the program will be at performing its desired function. So if AI is poorly trained (low data quality, or too little data), then it won’t produce accurate results. This application has an option for users to provide input if the results they received are wrong, which can also help further train the AI model.
A final idea to consider is the following hypothetical: if the photo provided is excellent quality, and if the AI has been accurately trained, then, ceteris paribus, is the grade CoinSnap gives more accurate than a grade given by a third-party grading service? This debate spans many areas, as AI and computer learning challenge the conventional anthropocentric approaches of many industries. AI is now helping doctors diagnose diseases, businesses handle complicated logistics, and factories automate many tasks. So one may ask: can a computer grade a coin better than a human? Let us know in the comments.
Overall, CoinSnap seems to be a useful option for people in the hobby, especially if you’re new to coin collecting. Using this app for the simple identification of coins found in an attic or inherited by a relative would be extremely helpful for people with limited numismatic knowledge. The AI grading tool could potentially help a new collector get an idea of quality, but it probably should be confirmed by a trained eye if you are considering selling the coin with confidence. As for veteran collectors or dealers, we say the jury is still out. Along with the concerns raised earlier, these two groups would usually be able to identify a coin and would have an idea of what to look for grade-wise. Look out for future articles as we continue to explore this app and any additional improvements that roll out.
* * *