This article is a follow up to my previous article, where I laid out the case as to why today's ubiquitous AI narratives have distorted many people’s expectations with regard to what AI is, and what AI is actually capable of doing. In a personal case study, I expounded on how I was susceptible for the narrative as well. In more detail, I looked at a tech company making use of AI to increase the odds of successful resource discovery. Looking from a naive perspective in hindsight, I thought AI would function as a magic ‘dowsing rod’ - in essence pinpointing to where the next big gold deposit could be found using a single powerful algorithm. It was at a later date that I figured this premise was very flawed in nature. AI wasn’t the magic dowsing rod I expected it to be. Instead, it was something much more down to earth, namely a new workflow that resulted in simple, but effective, additive incremental improvements. I understand this doesn’t sound half as exciting at first sight. However, it appears that outside of the fantasy realm, it is much more effective in achieving real world results.
Mineral exploration is all about where to drill. As a mineral exploration company, you want the absolute best bang for your buck, and that means that a drillhole should result in as much information about a geological system as possible. Misses are costly and the margin for error is slim. An untimely or repeated miss can be the difference between a “multibagger” and versus a bankruptcy filing. It is of no surprise therefore, that executives want to be as confident as possible when delineating a drill target.
One of the most prominent ways AI can increase the confidence level on where to drill, and thereby enhance human decision making, is by what can be dubbed as the ‘the dual perspective method’. In essence this method allows for a simultaneous, but independent, target generation by both the AI and the geologist. It is important to note that when I mention ‘AI’ - I refer to a geologist making use of AI-related technologies and workflows. Anyway, both the geologist making use of the AI, and the geologist using conventional techniques, will come up with exploration targets. Subsequently, these targets are compared, and where they overlap and/or correlate the confidence level is significantly increased - as both the AI workflow and the conventional workflow have produced the same output using the same input values.
Although this makes perfect sense to most people, simply under the motto of ‘two independent perspectives tend to be more accurate than one’ - it’s constructive to emphasize this with the following story, which I thought is a magnificent manifestation of the concept. Although not related to geology, but to crowd behavior, it lucidly illustrates how the building of one narrative can prove to significantly increase levels of erroneousness. The following story is a citation from the book ‘The Delusion of Crowds: Why People Go Mad in Groups’ ;
“Our modern understanding of how crowds can at times behave wisely began in the fall of 1906, when the pioneering polymath Francis Galton attended the annual West of England Fat Stock and Poultry Exhibition in Plymouth. There, he performed an experiment in which a large group of people acted with surprising rationality. Approximately eight hundred participants purchased tickets for an ox-weighing contest at six pence each, with prices awarded for the most accurate guesses of the weight of the dressed animal, that is, minus its head and internal organs. Amazingly, the median guess, 1,207 pounds, was less than one percent off the actual weight, 1,198 pounds. The average estimate was 1,197 pounds. Galton’s conclusion about the accuracy of the collective wisdom has since been repeatedly confirmed. More recently, New Yorker writer James Surowiecki summarized this concept in his bestseller ‘The Wisdom of Crowds’, in which he laid out three requirements for effective crowd wisdom: independent individual analysis, diversity of individual experience and expertise, and an effective method for individuals to aggregate their opinions. So what qualifies, for our purposes, as a “crowd” - the wise one of Francis Galton and James Surowiecki, or the unwise ones of Luc Jouret and David Koresh?
What separates delusional crowds from wise ones is the extent of their members’ interactions with each other. It’s doubtful that all, or even most, of Galton’s eight hundred contestants ever physically gathered into a single group. A key, and usually overlooked, feature of his experiment is that it involved the ‘dressed’ weight of the ox. Contestants had to fill out an entry card with their address so that the winners could be notified, and since the result would not become known until the ox was later butchered, this would have discouraged the contestants from congregating before completing their card.
A few years ago finance professional Joel Greenblatt performed a clever variation on the Galton experiment with a class of Harlem schoolchildren, to whom he showed a jar that contained 1,776 jelly beans. Once again, the average of their guesses, when submitted in silence on index cards, was remarkably accurate: 1,771 jelly beans. Greenblatt then had each student verbalize their guesses, which destroyed the accuracy of their aggregate judgment - the new “open” estimates averaged out to just 850 jelly beans.
Thus, the more a group interacts, the more it behaves like a real crowd, and the less accurate its assessments become. Occasionally, crowd interaction becomes so intense that madness results.
Accordingly, the accuracy of a group’s aggregate judgment rests on the participants not behaving like a crowd. It also, as Surowiecki points out, depends on the diversity of the group; the more points of view a group brings to bear in an estimate, the more accurate that estimate is liable to be.
Diversity of opinion also benefits the individual as well; as put by F. Scott Fitzgerald, “The test of a first-rate intelligence is the ability to hold two opposing ideas in mind at the same time and still retain the ability to function”. Over the past three decades, psychologist Philip Tetlock has examined the forecasting accuracy of hundreds of well-regarded experts; he found that those who took into account a wide variety of often contradictory viewpoints performed better than those who viewed the world through a single theoretical lens. In plain English: beware the ideologue and the true believer, whether in politics, in religion, or in finance.”
So what can we distill from this story? A couple of important things. First of all, it shows that more opinions do not necessarily result in more wisdom. In actual fact, when not under the right conditions - and as embodied by the jelly bean story - it even has an adverse effect on a potential correct outcome. So what makes for good wisdom? Surowiecki listed three conditions:
- Independent individual analysis
- Diversity of individual experience and expertise
- Effective method for individuals to aggregate their opinions
When these three conditions are met, one can safely assume that the aggregate of diverse opinions will prove to be more accurate than a single theoretical lens.
It is because of these very fundamental reasons that artificial intelligence can provide a potential breakthrough in the success rate of mineral discovery. Not because it magically points to where the gold is, but because it lets us humans see things from a different perspective, thereby greatly increasing accuracy.
When looking back to the three conditions, it is clear how AI can be of incremental value. Let’s dissect how exactly.
- Independent individual analysis - a target delineation using AI-techniques can be done wholly independent from an analysis using conventional techniques. An AI-geologist can independently generate drill targets, and so can the conventional geologist.
- Diversity of individual experience and expertise - target generation with use of AI tools is suitably different than target generation using conventional tools. Machine learning algorithms can detect correlations otherwise missed or deemed unimportant by humans. AI is like a ‘second pair of eyes’. Not influenced by human bias or heuristics, the workflow of artificial intelligence is idiosyncratic in nature, and with that, diverse.
- Effective method for individuals to aggregate their opinions - the delineation of drill targets results in an output in the form of a heatmap. As both the AI-method and the conventional method produce maps with the same dataset as input, they can be compared. Where meaningful overlaps and/or correlations occur, the confidence level is increased.
AI holds true to all three conditions. An AI analysis can be independently done, is diverse in nature, and allows for an aggregation of the outcome. Frankly, it looks like including AI in the decision process will likely result in a better accuracy of the outcome, namely successful resource discovery.
This is not a sneer to conventional geology. Frankly, I’d argue it is even quite the opposite. Many geologists are skeptical about AI, and not because they inherently think it is useless, but because they are afraid they might become obsolete once AI workflows make the entrance in their field. In the case of mineral exploration, I think there's no need for geologists to feel uneasy about that prospect. As shown in the described use case, AI is only of incremental value when used in combination with conventional geology. That means that when either of the pair of eyes is left out, that is; AI or conventional geology, you are back to square one. However, it also turns out that conventional geology is not as fundamentally sound as many ought it to be. Simply because it consists of only one theoretical lens. Artificial intelligence can function as a ‘second pair of eyes’. By providing a different perspective, it should greatly increase accuracy.
The concept of a ‘second pair of eyes’ is as old as the world itself, and has proven to be very effective over time. When viewing AI through this lens, it becomes a lot more intuitive in nature. Wizardry is being swapped for common sense.
To conclude, when a company is figuratively putting all of its chips on one number, additive incremental probability improvements may not have the sex appeal of the AI fantasy, but it will be an AI reality that any wise executive will jump on...
