There’s currently a lot of discussion on how Artificial Intelligence (AI) might help address the talent shortage. However, technology alone is not enough – AI needs to be implemented with wisdom and human experience if it’s to be effective.
The Red Socks Conundrum – AI Pattern Recognition with a Human Resolution
We know that AI can be used to scan and interrogate data sets to find patterns, correlations and coincidences that a human being would just not have the capacity to.
A hypothetical example we often to use is the finding that the best salespeople wear red socks. We imagine that this finding has come about by an AI being fed all sorts of different information about salespeople, from their sales performance to their clothing choices, eye colour and a host of other ‘facts’.
The AI churns through all this information and out pops the finding that all the best salespeople wear red socks. Now, in the absence of a human in the loop, the AI could recommend that we should just go and hire anyone wearing red socks and they would be a great salesperson.
But with a human in the loop, we would question why it is that a person who wears red socks is a better salesperson? We might speculate that wearing red socks could be a sign of confidence and being prepared to stand out from the crowd, and of course, it is these characteristics, not the red socks per se, that are the better predictors of being a great salesperson.
As occupational psychologists we could then go and test this theory to be sure that ‘red socks’ were just a proxy for the more important latent variable of confidence and standing out.
In this way, the combination of human experience and AI in HR means that we can make much better decisions and get better outcomes through using AI as an assistant that makes recommendations but ensuring a human is the final arbitrator. Indeed, this philosophy is baked in to GDPR, where a data subject has the right to request a human review of any automated decision-making process.
The Limitations of Data
One of the areas where AI can come unstuck is if there is not enough data to be able to find reliable patterns. To be clear, AI can always find data patterns, but we need those patterns to be reliable if we are going to be able to do anything with them.
In things like political opinion polls, it’s generally accepted that a sample size of 1,004 people is representative of the population. We might consider that it’s the same for AI (although recognise that for the titans of tech like Facebook and Google, AI is more likely to run on billions of data points). However, in most cases where AI is being applied in HR, we won’t have sample sizes that are anything like as big as this, and so we could be faced with the challenges of sampling error – bias and error in our results. In a nutshell, the less confident we are that our sample represents the population, the greater the risk our results will be flawed.
This can lead to another problem that statisticians call ‘restriction of range’. Think of it this way:
If we lined up 100 people in order of their height, we would have the shortest people on the left and the tallest on the right. Then imagine that we took just the five tallest people from this group and focused on height differences between them. It’s possible that there may be some differences; equally, it’s more than likely that these five tallest people are going to be of a fairly similar height to each other – even though they are of a significantly taller than the rest of the other people in the line-up.
This is obvious when we look at it in this way. Yet this is what happens in recruitment validations all the time.
When people attempt to use AI to find the differences between the top and bottom of the cohorts that they have already selected, it’s like trying to find the difference in height between the five tallest people, rather than the difference in height between the five tallest people and the other 95 people in our line-up. There might be differences, but the biggest differences are likely to be between the people we already selected and the rest of the group – but we don’t know that because we don’t have the data on them.
Another area where AI can fall down is when we put the ‘wrong’ data into the model in the first place. By ‘wrong’, this could be data that has systemic errors in it, data that isn’t representative of the population, or data that is already biased in some way.
One example of this that reached the popular consciousness was the Amazon hiring algorithm that ostensibly worked out that male candidates were more likely to be better software engineers and so penalised and rejected female candidates. Of course, this is utterly untrue, but how can we expect the unchecked AI to know that, when it is only as strong as the data that feeds it – a dataset which comprised almost all male software engineers, leaving the AI to conclude that male = software engineer, female = not software engineer’.
The Invaluable Human Experience
AI has no conscience, no morals, and applies no ethical judgement to its decisions. It is simply a set of algorithms running through statistical calculations and correlations to find the pattern that makes the best sense of the data.
And this is why the human touch is invaluable. A human overseer can spot that something has gone awry, ask why and fix it, or can decide to take other appropriate action (which is actually what happened in the Amazon case, and the algorithm was never deployed to live).
When we blend the precision of AI and data analytics with the nuance for ethics, morality, perspective and intuition that human experience brings, that is where magic can really happen. That’s the future of AI in HR that I am most excited about.