How will ChatGPT impact recruitment?

In the short run, the LLMs have the potential to help recruitment in several aspects. This technology can be used in two broad modes - extractive and generative. Extractive AI extracts value and information from unstructured data. Generative AI generates new content automatically, without a human having to write it. In both cases, it is used to drive or aid automation and automated decision making. This technology is incredibly advanced and has the potential to dramatically improve pieces of the recruitment process. However, there are a few aspects of generative AI that are likely to make hiring challenging in the future.

Hallucinations - If robots can dream, they can hallucinate too! In extractive mode, with the right configurations, and most importantly, the right code and human checks and balances, the risk of the AI hallucinations impacting any workflows is minimized. The risk still exists, but can be mitigated with technology and process design. However, in generative mode, LLMs are unpredictable. 

• If leveraged to write job descriptions, will they suggest skills that are not really pertinent to a job? Or miss skills that are really needed? Our research and indeed, our experience has proved that the strongest determinant for getting qualified applications to apply to open jobs is having the right skills in the job description.
• If leveraged to evaluate candidates, will they recommend certain people over others for unexplainable reasons? Explainability, the ability for a company to explain the reasoning behind AI decisions or predictions, is already an area of intense focus. However, even state-of-the-art in explainability leaves a lot to be desired. Decision making in recruitment is incredibly high-stakes for both companies and candidates. Leveraging LLMs in this space needs to be carefully thought through.

Dated information - The way the LLMs are currently designed (a centralized model that takes a large amount of resources to train and test), updating them with the latest information is quite the chore. 

• In their current avatar, their information content will always be a year or two behind the latest "picture of the world". For example, searching for Bard or LLAMA (other LLMs) on ChatGPT yields nothing. As does searching for other recent news.
• This has the potential to impede jobs and candidates in fields that are emergent, a growing sector of the labor market.

A Trojan horse? - The models might have a more subtle but insidious impact. Since the user interactions with these models are via "user prompts". The model output, and hence, its fitness for use, seems to be sensitive to the prompt in varying degrees.

• There is an increasing emphasis on making sure the prompt is "just right" - not surprisingly, entire subreddits and "prompt marketplaces" have come into existence to help users find and refine their prompts. 
• Are prompts the new "resume templates" of the new hiring world? Will this lead to people using the same (or very similar) prompts to refine their resumes or job descriptions? This could lead to a "great resume convergence" in which all resumes read similarly, highlight the same aspects of one's career, and use the same keywords.
• No doubt this happens today as well, to some extent, but LLMs could supercharge this phenomenon. The same could happen for job descriptions as well. If this is indeed the future, then how will hiring processes adapt? Just as we are drowned by a deluge of SEO and bot-generated content spam on websites and web search, hiring teams could drown in similar LLM-generated overstatement.
• With all new technology comes attempts to trick it. There are already websites and content creators dedicated to adding hidden prompts to resumes so that a LLM would recommend that resume. Companies will need to be vigilant if they intend to use a LLM to evaluate candidates.

Ultimately, we believe that there is more regulation inbound around the use of AI in various industries, especially where decision making will directly impact humans. With the size, complexity, and aspects like the fact that the models are only available via an API to end consumers of LLMs, this question becomes even more complex.

In conclusion, recruitment will not be immune to changes brought on by Generative AI. There is potential to streamline hiring processes and shorten the time investment required from hiring managers and recruitment teams on processes like writing job descriptions. However, LLMs have unignorable shortcomings. Are these shortcomings fixable, or so inherent a manifestation of the way LLMs are trained that it will never be easy to solve for them without changing these models fundamentally? There may indeed be a GPT-HAL of the future that starts off helping organizations find the best candidates possible, but eventually just refuses to consider any candidate that does not meet its definition of the "ideal candidate."

Want more learn more about how we use Generative AI?

Learn how Meytier uses LLMs to streamline the hiring process