Faking or cheating by candidates in the selection process is not a new phenomenon. Depending on the selection method, candidates may fake responses to improve their chances of progressing to the next stage. For instance, some candidates incorporate keywords into their resume to manipulate resume parsers, provide socially desirable responses in a personality test, and utilize impression management in face-to-face interviews. One study found that 70% of a surveyed sample of 1914 employees admitted to lying in their resume, with 76% on cover letters and 80% on interviews.
Generative AI raises the stakes for online assessments
The rise of generative AI, particularly Large Language Models (LLM) like ChatGPT, has introduced a whole new way for candidates to cheat, posing a serious threat to the fairness and validity of online assessments. Candidates can now generate a whole new resume or use these tools to cheat on online assessments. Especially in asynchronous online chat interviews, the use of AI-generated content (AGC) can seriously impact both the effectiveness and the fairness of the outcomes.
What Sapia.ai already does
Sapia.ai has been innovating in the space of text chat based structured interviews since 2019. Our flagship AI screening product, Chat Interview™, is used by leading brands to increase the efficiency, fairness, diversity, and candidate experience at the top of the funnel filtering in high-volume recruitment. Candidates’ responses are already scanned for plagiarized content and flagged for the attention of recruiters or hiring managers, letting a human take action. Our work has also shown that only an average of 3% of candidates plagiarize across different role families, showing that the majority of the candidates want to stay authentic.
To combat the latest cheating challenge, researchers at Sapia.ai have developed a novel way to distinguish between interview responses generated by AI, such as ChatGPT, and genuine human-written responses. They leveraged Sapia.ai’s vast proprietary dataset of over 12 million human-written interview answers in training this detector (over 1.0 billion words). The detector capitalizes on a subtle distinction to identify responses that may have been generated by AI rather than the candidate. By observing the probability discrepancy between AI-generated and human-written text after random perturbations (a form of randomly changing words), the detector can identify, with high accuracy, whether a response is AI-generated or not. The AGC detection model proved highly effective, achieving a very high accuracy (ROC-AUC of 94.49%).
What we found
Based on the AGC detector outcomes, Sapia.ai researchers analyzed responses from over half a million candidates who participated in online chat-based structured interviews between January and September 2023 to understand the scope of the problem. January 2023 was chosen as the start of the period as ChatPT had already started getting popular by that time. During this period, 1.16% of candidates were flagged for using AGC in all five questions asked in the Chat Interview™, while 20.20% had at least one response likely generated by AI.
Now, detection is only one aspect of the solution. Deterring candidates from using AGC, on the other hand, is the better way to address this problem. To deter candidates from resorting to AI-generated content, we implemented several methods: instructions were added to encourage candidates to write their own answers, and warnings about tracking plagiarism attempts were included. Additionally, the ability to paste responses was disabled, and pop-up warnings were introduced when answers were flagged as using AGC.
With the implemented deterrence methods, we observed a drastic decline in AGC rates, showing that these measures have a tangible impact on curbing cheating. The weekly average of the candidates flagged for all five questions fell as low as 0.20% from a high of 2.4% (a tenfold reduction) prior to implementing the deterrence mechanisms. Further, we observed a clear alignment between the timing of the implementation of deterrents and the reduction in AGC rates, emphasizing the effectiveness of these measures in encouraging candidates to be themselves when completing their Chat Interviewf.
As technology continues to evolve, so do the challenges in maintaining the integrity of job interviews and assessments. The addition of AI-generated content poses a serious threat, but innovative solutions, such as the AGC detector and deterrence methods, show how Sapia.ai effectively de-risks cheating via cutting-edge innovation, in order to maintain the fairness, validity, and reliability of the hiring process.
About Author
