Demystifying AI for HR Leaders, and the HR / AI Ecosystem
AI as a technology is often misunderstood – here are some practical tips to avoid pitfalls and capitalize on missed opportunities.
The Explosion of AI in HR
There was a lot of hype around AI in 2022 with stunning developments like DALL-E and ChatGPT, but the hype had already arrived in the HR space. The excitement is warranted - we are in the midst of a technological revolution that has democratized the technology. For example, a report from the Stanford Institute of Human-Centered Artificial Intelligence found that between 2017 and 2021, the cost of training an image recognition system decreased from $1,000 to $4.60. Creating a machine learning model that would have taken tens of thousands of dollars of equipment plus half a dozen developers in 2013 can now be produced by a single data scientist on their laptop in Python.
In addition to greater technological availability, organizations now have more data to feed AI models. Most medium to large sized companies have hundreds of thousands of data points in their HR information systems, and HR tech vendors with talent intelligence platforms can now access data from a multitude of large companies, as well as data scraped from the web, have upwards of billions of data points.
This revolution has created opportunities for HR functions to apply AI and find operational scale and strategic muscle that was previously out of reach. Unfortunately, a number of high-profile misuses of AI in the hiring process have also garnered a lot of (negative) publicity. Without diving into the specific issues of those cases, this article serves as a primer for HR leaders and subject matter experts to recognize and understand AI – what it means, how to think about key surrounding elements to ensure effective and ethical use, and how to think about some underrated use cases.
Creating a machine learning model that would have taken tens of thousands of dollars of equipment plus half a dozen developers in 2013 can now be produced by a single data scientist on their laptop in Python.
What’s Under the Hood?
Demystifying AI is key to having a better conversation in HR. Sadly, much of the conversation around AI in HR stems from an incomplete or inaccurate understanding of the technology. First, even the term “AI” can be extremely misleading. The vast majority of AI actually used today falls into a subcategory of AI called machine learning. Machine learning systems are not intelligent by human standards but are powered by a statistical process of pattern-finding (learning).
Machine learning systems, or models, have the potential to help us to make more accurate predictions or create smarter tools. For example, if you use google mail, emails that are likely to be spam are automatically flagged. Behind the scenes, google has created a machine learning model that “learned” through pattern recognition – finding the most common attributes of emails that are frequently categorized as spam.
You can imagine a scenario where you trained a machine learning model, but the emails (the data) you used to train the model set it up for failure. Maybe there weren’t a sufficient number of emails for the machine learning algorithm to identify statistical patterns. Even worse, maybe the emails marked as spam were not actually spam! This is what a bad AI model looks like, and as an HR leader/professional, you may already be imagining situations in HR that could have more serious consequences for real people.
The HR/AI Ecosystem
Most “AI” is not actually intelligent by human standards, so we don’t need to worry about Skynet enslaving the human race anytime soon. Instead, the risks and pitfalls of AI in HR are more nuanced and more systematic. Treating AI not as an isolated technology, but as a broader ecosystem allows the HR leader to ask better questions about the effectiveness and ethics of AI projects. Some key components of the HR/AI ecosystem include the data inputs (training data), the business problem / question, the models, the data scientist, the user, and the subjects (employees/candidates).
Any single part of the HR/AI ecosystem can operate as a single point of failure. The bad training data example already given can be especially devastating from a diversity standpoint. If candidate or employee datasets used to train machine learning models are not representative of the broader subject population, then models can amplify and scale bias. Just like within broader HR analytics, identifying the right business problems to solve is key to ascertaining value from HR in AI.
Potentially the most interesting and under-explored components of the HR AI ecosystem are the modeler, the user, and the subjects. These three can be thought of as the “decision engine” that translate AI into action. What’s unique about the use of AI in HR is that unlike a lot of consumer-facing technology, the buck doesn’t stop with the user. In order to realize value from a predictive turnover model, the user must leverage an additional intervention to influence an employee or candidate's behavior. This is where HR’s consultative acumen, and some behavioral science come into play.
Some Underrated HR/AI Use Cases
Most HR / AI hype is limited to a small number of use cases, like automated resume screening and predictive turnover models. Those use cases are powerful, but machine learning is an incredibly versatile technology that can be applied to a broader suite of problems. Here are a few lesser-known use cases for HR practitioners to take away:
1. Creating markets – the talent marketplace is one of the most exciting recent innovations in HR. Most talent marketplace platforms are powered by (read: the AI is trained by) skills data, so it’s crucial to have a strategy for the collection, validation, and governance of this data. Don’t think of talent marketplace as being just an internal mobility platform – AI has the ability to power markets for gigs, mentors, or even the services of teams within an organization.
2. Job families and workforce planning – many organizations have job titles and job descriptions that are a total mess. This makes doing analysis on what talent the organization has, and what talent it needs in the future a challenge. One type of machine learning, unsupervised learning, can be used to read data from job descriptions and cluster them together, creating data-driven job family taxonomies that can be used for workforce planning.
3. Adding color to performance management – a really cool paper by Andrew Speer from Wayne State University outlined a new method for measuring employee performance: Most organizations have a lot of data in the form of text in manager reviews as well as peer and self-reviews, but this is an underutilized data source. Speer demonstrated that numerical performance scores created from natural language processing (NLP) correlated highly with human ratings of performance.