The Dirty Little Secret of ChatGPT & Job Postings
By peterweddle
August 07, 2023
By Peter Weddle, CEO TAtech
Let’s begin with what’s known. ChatGPT is an extraordinary technological development. It is able to perform tasks humans find onerous or less important and help humans perform better in tasks they consider their forte or responsibility. The model is impressive, there’s no doubt about it. That said, ChatGPT is not without its shortcomings. And one of those imperfections, in particular, has the potential to significantly degrade the outcomes it can achieve in online recruitment advertising.
Like all AI models, ChatGPT is trained with historical data. It’s genius, if you want to call it that, is based on the 570 gigabytes of data or 176 billion parameters on which it was originally trained. The current model – ChatGPT4 – is breathtakingly smarter. It encompasses 100 trillion parameters or 5 times more than that “archaic” first generation model. In effect, ChatGPT has sucked up a whole lot of human knowledge and that’s made it pretty darn intelligent.
As you might imagine, this impressive developmental scale has made thinking up applications for the model today’s favorite pastime among the tech cognoscenti. Social media and the blogosphere have been inundated with use cases addressing everything from person-job matching and upgrading the candidate experience to interview scheduling and assessment. The one area that gets mentioned the most frequently, however, is its ability to help recruiters improve their job postings.
It would, of course, be a great benefit to recruiters if ChatGPT could lend them a hand in upgrading their job postings. Those ads definitely need it. Unfortunately, that’s not likely to happen. Despite the model’s extraordinary education, its capabilities are undermined by a dirty little secret: the data used for ChatGPT’s development is old.
As pointed out by Alexander Chukovski in his forthcoming TAtech book, ChatGPT Undressed & Unadorned (get a free copy by attending TAtech Europe & The EMEA Job Board Forum in London on December 4-6), the most current data on which ChatGPT was trained dates from 2021. That’s the equivalent of the Dark Ages in technology development.
Yes, ChatGPT has been updated since then. In fact, Open AI released the latest version of the model on March 14, 2023. And yes, as noted above, that model is significantly smarter than the previous one. But, here’s the rub: We don’t know if this updated version is smarter in ways that will be helpful to users in general and recruiters in particular because OpenAI hasn’t yet released the sources it tapped to train it.
One thing is certain, however, all those megabytes of data and bezillions of parameters describe the past. In fact, regardless of the sources, the data used to train the model detail the job postings we’ve always had. As a result, ChatGPT is simply unable to devise or even describe an improvement to such an ad. It’s never learned one. All it’s capable of doing is regurgitating what recruiters have always done.
And, what recruiters have always done in their job postings isn’t working particularly well. Thanks to programmatic ad placement, today’s job postings are being put in front of the right job seekers but their response to the ads is all too often a big yawn. Completed applications are way down despite over a quarter of the workforce now describing themselves as active job seekers according to a recent survey report from CompTIA.
Why such a dismal track record? There are several reasons, but the most important is that the way job seekers “shop” for a new opportunity has changed. They’re now much more likely to be impatient with or quickly skip over ads that are about as engaging as a brick and use language only a bureaucrat could love. Think value propositions buried in hundreds, even thousands of words and calcified terms such as Requirements and Responsibilities. Their unwillingness to tolerate such ads represents a dramatic demographic shift and as smart as it is, ChatGPT is clueless about it.
All of which means this: if we want to see job postings improve, we humans must first figure out how to do that so we can then train ChatGPT on our intelligence.
Food for Thought,
Peter
Peter Weddle has authored or edited over two dozen books and been a columnist for The Wall Street Journal. He is the founder and CEO of TAtech: The Association for Talent Acquisition Solutions.