AI Due Diligence for Mining Investors: Four Questions That Cut Through the Noise

Due diligence for AI in mining: ask the right questions to understand. (Source: Unsplash)

Companies applying AI in the mining industry make all sorts of claims. They will find you a "world-class deposit" by reading the old reports on an area. Their model identifies anomalies "with 94% accuracy." Add to that the fact that AI now commonly writes the marketing materials for AI products. The language is polished, plausible, and increasingly hard to distinguish from genuine technical substance.

The problem isn't bad AI. It's a shared language gap. At the technical level, it sits between a data scientist and a geologist. At the corporate level, it sits between people with different expertise and different stakes, where investors are being asked to evaluate claims against no agreed standard.

Here are four questions for AI due diligence in mining that cut through it. You don't need to understand machine learning to ask them. You need to know what a good answer sounds like.

What decision does this help me with, and how?

This question forces specificity. A good answer names the decision, describes the data inputs, and explains where the human sits in the loop. These answers often sound modest. "It flags intervals for review by the geologist" tells you more than "it identifies world-class targets." One describes a tool. The other describes a promise.

Watch for two failure modes at opposite ends of the spectrum. If the output creates work rather than resolving it, that's a problem created, not solved. An AI product that generates thousands of anomalies requiring follow-up has added a task, not removed one. But if the claim is that the AI makes the decision itself, that's equally problematic. A model that produces definitive high-value targets without meaningful human review has skipped the most important part.

The answer here sets the frame for everything that follows.

Where did the training data come from, and is it relevant?

This is the geological question underneath the technical one. Mining is site-specific in ways that make data transfer harder than it sounds.

A model trained on porphyry copper deposits doesn't transfer cleanly to a sediment-hosted copper system. Tenement-level data for regional exploration may give you the wrong scale, and patchy coverage. Using remote sensing data from active mines as training for greenfields exploration introduces biases that won't be obvious until you're drilling.

A good answer demonstrates understanding of the mineral system, the appropriate scale, and the relevance of each data input. The team should be able to show that the data represents the problem they're trying to solve, not just that there's enough of it. Volume isn't the same as relevance.

The same logic applies beyond exploration. Whether the application is predicting equipment failure, monitoring hazards, or optimising an ore sorter, a good answer follows the same sequence: understanding, scale, relevance. Our blog on how data is used and understanding it as a corporate risk may also be of interest.

Can you show me predictions against ground truth?

Note the deliberate phrasing. This isn't a question about model validation in the technical sense. It's a question about physical proof.

In mineral exploration, this is where AI products most often fall short. Targets can be generated cheaply. Proving them is expensive. Drilling costs significantly more than running code, requires different expertise, and takes time. Many products are working hard on the generating side and have limited results on the proving side.

When the answer is yes, probe further. Check whether the successful prediction was made on a genuinely new target, or whether it was already a known deposit. In some respects, this is reminiscent of the chicken-and-egg choice companies sometimes face of flying expensive geophysics surveys or drilling more holes (hint: the national government should be flying the survey), but it is often crucial to the successful application of AI in mineral exploration.

A good answer lays out the proof and links it back to the decision framework in the first question. The best answers put a polished core section in your hand, traceable back to the rest of the core, which is hopefully a shed full.

What happens if this is wrong?

Every model fails. Models are simplifications of complex problems, and in mining that is compounded by incomplete data layered on incomplete understanding. Failure is not a bug, it is both a feature of innovation and is a strength depending on how the team has thought about it.

Competent teams have reviewed their failure modes, built in safeguards, and can describe what happened when the model got it wrong and what changed as a result. "Our latest version does X" or "we built this feature based on customer feedback" are often euphemisms for failures the team has worked through. Dig a little deeper and find out why.

If failure modes haven't been thought about, the product hasn't been deployed in a real environment yet. That doesn't mean the work is a write-off, but the development is immature and, borrowing from software development, implementing basic alpha and beta testing will increase reliability when it gets there.

Building a picture

These questions don't require you to understand the model. They require the team to have done the work. A good answer is specific, honest about failure, and demonstrates rigour. That's the signal.

It's more than identifying bad AI. It's about finding a shared language between parties with different expertise and different stakes. That translation is no easy task, and it requires someone with domain experience in both data science and mineral exploration.

It is this experience that built the AI:CORE framework, a structured approach to evaluating AI in the mining industry for investors who want to cut through the noise and companies who want to implement best practice.