How I Think About AI

Maintaining our living standards in Scotland will depend on integrating AI into core systems like education.

When I approach the Cashline, insert my card, enter my password, and request a statement, I expect to see an accurate list of debits and credits on my account. Repeating the same exercise at another ATM on a later date, I also expect a balance that reflects the previous position plus any account activity since my last inspection. This straightforward, accurate, and consistent process is enabled by traditional but rather mundane banking technology.

If my bank details were stored in an Artificial Intelligence (AI) system, I might see a recognisable balance and the latest expenditure, but there’s a chance I might not. I should also anticipate some inconsistencies in the statement. For tolerating these account irregularities (technically termed “hallucinations”), I might receive some interesting additional insights, such as recommendations to move money into a higher interest account, suggestions that I’m overpaying for energy, or advice to switch travel agents if the credit rating of my usual one has dropped by 50%.

Clearly, there are areas where AI is well-suited (like recommendations) and areas where it is not (such as core banking). However, there are also grey areas. For example, credit ratings are checked before approving a loan application. These ratings are typically calculated from various data points provided by large data-crunching companies like Experian. These credit checking companies generate data points from hundreds, if not thousands, of different data feeds. Do they fully understand how all those calculations were derived from their various data providers? Probably not.

To reduce costs and create more elaborate credit points, the temptation to incorporate AI will be overwhelming for companies like Experian or their downstream data suppliers. This should concern us because AI, while sophisticated, often lacks nuanced understanding. It navigates probable relationships and links data together, but this process is fundamentally different from human reasoning.

We may marvel at AI’s ability to win games of chess, but that’s largely a memory trick. Current AI should really be called Artificial Recall, and like human memory, it has gaps. Real human intelligence combines recall with intuition, and AI engines have yet to master the latter. Drop a human into a new situation, and they can adapt. AI engines, however, struggle with such adaptability.

Consider Deep Blue, the computer chess champion. It can remember and evaluate more chess positions than any human and win, but that’s recall and calculation, not true intelligence. Deep Blue didn’t get bored halfway through the famous game with Garry Kasparov and start playing poker with itself or invent a new game. It simply regurgitated what it had seen before. Many magic tricks rely on memory skills to impress the audience, and AI functions similarly.

Fortunately, as AI advancements continue, a Frenchman named François Chollet has developed a technique to measure the human-like capabilities of AI engines that exclude memorisation. It’s called the ARC test. Most humans can effortlessly score 80%, whereas the best AI engines struggle to achieve a respectable score. If you’re under pressure in business or government to introduce AI, just ask for the engine’s ARC score and observe the consultant’s face drop.

So, if AI is limited, repetitive, and sometimes misleading, why all the excitement? There are scenarios where being probably correct quickly is highly advantageous, such as in war, education, and art. AI should not be making final decisions, but it is excellent at checking decisions that have been made or flagging those that need to be.

Consider a teacher in a class of 40 trying to identify who is falling behind while still motivating the rest. Instead of testing once a month, AI can run continuous checks and make recommendations on how to improve scores based on each student’s level and learning ability. The teacher makes the final decision, but AI can guide them and can also continuously monitor the teacher’s performance. While the Scottish education system may not be ready for this, other countries with fewer resources will achieve better outcomes for students, advancing quickly and gaining an edge in both education and trade. In future, maintaining our living standards in Scotland will depend on integrating AI into our core learning systems.

ChatGPT has simplified the skill level required to use AI. Remember the surge in machine learning jobs five years ago? Today, that’s the fastest part of information technology being automated. ChatGPT allows us to direct a computer using natural language and get results that would normally take days or weeks. For example, it generated the following image for the previous YesThink article with just this prompt “create a cartoon image to represent tartan chickens flying home to roost in Edinburgh.” No creative talent was required on my part other than to think up the prompt!

ChatGPT has helped YesThink look professional with great images, but it can’t (unfortunately) recreate the brain or writing flair of Jim Sillars. If everyone used ChatGPT to create content and images, all publications would start to look very similar and boring.

Both the UK and the EU are rushing to legislate AI usage, which is be premature. Given the mess the data production legislation (GDPR) left in its wake, we should be cautious of government efforts. This EU/UK initiative is driven by an academic AI ethics community which isn’t doing enough practical AI work but still thinks it deserves a voice. Most groundbreaking AI developments these days are driven by industry not our Universities. While there are many pros and cons to using AI, the technology remains in its infancy, and industry must be free (within reason) to experiment, or our economy will lags those who are truly innovating.