Research shows that AI can rival kriging in estimating gold reserves.
getty
Search for “AI gold mining” and you’ll be flooded with promises. Increases discovery rate by 30%. Drilling costs are reduced by 40%. One vendor site claims a hit rate of nearly 75 percent, compared to the industry average of 0.5 percent. The pitches are everywhere, they’re loud, and they’re almost all about one thing: finding gold. Where to point the drill. Where in the ground is the ore hidden?
Now let’s read the actual mining documentation. Get the latest technical reports on large gold projects from the public file system. This is the type of report that companies must publish before they raise a dollar for a deposit. Now onto the part that shows how much gold is actually in the ground. That number, or resource estimate, is what sets a company’s value, backs its financing and drives its stock price. And what method was used to calculate it? Kriging. A statistical method for estimating borehole-to-hole grade that geologists have relied on since the 1960s. I can’t see the neural network. This quote is signed by the person who is legally responsible for the license and its correctness, or in the language of the industry, a qualified person.
That’s the gap in this story. Mining has trust issues. We trust artificial intelligence to go find gold, but we don’t trust it to count it.
numbers that actually matter
Most of the press about AI in mining is about exploration, and it’s easy to see why. Pointing a model at satellite images or old drilling logs to suggest where to dig is exciting and cheap to talk about, but nearly impossible for an outsider to verify. Companies can claim that their algorithms have sniffed out a promising target, but who will argue? Nothing has been built. No one will sign it.
Resource estimation is a completely different animal. This is a number that tells investors how many ounces a project holds and what grade it is. Banks take advantage of this by lending money. Stock prices will follow suit. When miners say it’s 5 million ounces, that number comes from a formal estimate, and somewhere in the document is the name of the person qualified to attest to it. If the numbers prove to be illusory, the person could be brought before securities regulators. There are real skins in the game, the professional and legal kind.
So this is the final stage of a high-stakes business. And that’s exactly the outcome of not allowing AI to pass through that door.
Keep the two jobs clear, because the whole story depends on the difference between them. Exploration is hunting. Scan geology, geophysics, old records, and satellite imagery to decide where to drill next. One wrong move can mean wasting the drilling budget, which is painful for investors, but I won’t lie. An estimate is a count. The analysis results are obtained from the drilled holes and the amount of metal filled in the space between the holes is calculated based on statistics. If you get that wrong, you’ve published fake numbers for people to buy and sell. One job is gambling. The other thing is a promise. AI ran riot through betting, but the promise was not fully kept. Once you see that line, the overall pattern comes into focus.
What is actually stated in the submitted documents
I’ve looked into places that can’t confuse you. It is a technical report itself filed under Canada’s NI 43-101 regulation and its U.S. equivalent regulation, SK 1300. These are audited disclosures and have a legal basis. Across recent gold reports for major projects, resource estimates are based on conventional geostatistics, prepared based on the mining industry’s own best practice guidelines, and signed by geologists trained in the methodology. A qualified individual’s listed qualifications include a specialized geostatistics program from the French School of Mines, classical genealogy, and the Kriging tradition. Machine learning makes little assessment of where the actual estimation takes place.
This is not a unique phenomenon unique to one company. If you read the names of famous people, the pattern will fit. Exploration press releases brag about AI. Audited Estimates are just a few clicks away on the same filing system, quietly using mathematics that is older than the internet.
Meanwhile, in the laboratory
This is what makes the gap not just alarming, but strange. This technology is clearly not ready. Researchers have been pitting machine learning models against kriging on real deposits for years, and the results are quite interesting.
A 2025 study of Indian iron ore blocks found that machine learning models barely worked, even with regular kriging. The coefficient of determination for Random Forest was 0.74 compared to 0.74 for Kriging, and the error was small. It’s basically a tie. A recent study published through the Institute of Mining Engineers found that graph neural networks actually outperformed kriging in predictive accuracy in messy, heterogeneous ore deposits, while preserving short-range details that kriging tends to obscure.
But this is the part that the hype guys ignore, and the scoreboard is mixed. In the Itakpe iron ore deposit, regular kriging completely defeated the neural network and produced tighter resource yields. Same contest, winner on the other side. The lesson buried in this literature is not that “AI wins.” It depends on the rock. Sometimes the model wins, sometimes the old method wins, but we often don’t know which one until after the fact. If your number comes with legal liability, it’s difficult to sign off as “it depends.”
why does the door remain closed
So why is the industry willing to let AI go on a treasure hunt, but keep it out of the tally room? There are several reasons, but they pull in different directions.
The most obvious one is responsibility. The qualified person signs the quote and owns the results. Kriging is a known quantity. It has been litigated, audited, taught and trusted for decades. Its assumptions are transparent, allowing peers and regulators to follow the logic from drill holes to block models. When you hand a job off to a neural network, some of that transparency is obscured. When a model that can’t fully explain itself produces a number, who’s going to answer if that number is wrong? The Qualified Person framework was built for a world where humans make estimates. I don’t yet have the vocabulary to model sitting on a chair.
This auditability point is so core that it’s not worth a shrug. If a regulator or independent reviewer wants to check the kriging estimate, they can rerun it and look at the variogram to see exactly which nearby samples determined the grade of which block. This method shows how it works. Some machine learning methods can do something similar. Others are more like sealed boxes that capture the drill data at one end and extrude the material at the other end, with the inference trapped inside the weight. A qualified person asked to prove a number that cannot be fully recovered is asked to vouch for a stranger’s homework. There are plenty of people who don’t, and it’s hard to blame them. The problem isn’t that the model is stupid. That said, “Trust me” is not the way to defend yourself in front of a securities commission.
Then there’s the mixed scoreboard. If the academic results were a total failure, the pressure to recruit would be enormous. it’s not. When a new method can only occasionally beat a reliable old method, and the old method does not pose career risk, sensible professionals stick to what they can protect.
And there is a palpable conservatism in this corner, closer to prudence than stubbornness. The estimate is the basis of the entire project. You cannot rebuild the foundation with half-tested materials.
Questions floating around in the counting room
None of this is settled, but that’s why it’s worth paying attention to. The gap between what labs prove and what applications accept is what industry chooses to do, not a wall set by physics. As the model improves and the pressure to use it increases, that choice becomes harder to maintain.
The interesting battle is not whether AI can estimate gold reserves. Evidence suggests that in some cases this is possible, and in others it is not, and the field is closing down. The battle is about accountability. When a model begins to shape the numbers that determine a mining company’s value, what happens to those numbers and the people behind them? The rulebook governing resource estimation is written for human judgment. It’s about to be tested by something else, but no one has decided yet who will be responsible when the algorithm receives votes.
For now, that line holds. AI is all for the money. Measuring is not yet welcome. Whether that line is bent or not, and on whose terms it is followed, is something that is actually worth following long before regulators decide what to call it.
A final thought for those who want to read this as Luddites vs. Progress. it’s not. The miners who keep AI out of the tally room are making defensible claims about numbers that deliver real results, and the researchers who prove their models work are doing honest science. Both could be correct. The tension between technology that is almost ready and liability regimes that are not yet complete is the most interesting aspect of mining today, and it will be resolved through applications and rule changes that almost no one will read. That’s usually where the real change lies. Look at the methodology section, not the press release. The day a major gold estimate is approved by the model and a qualified person is attached to it, the rulebook will quietly turn a page.
