AI detection tools are one of the ways to flag the use of Generative AI within the hiring process. The problem, however, is that the current detection tools available on the market are largely inaccurate, resulting in a high volume of false positives.
In recent benchmarks we have found the following on average:
- Answers are correctly identified as Human or AI 65.9% of the time
- 19% of the time they gave a false positive (i.e. flagged a human as AI)
- 15% of the cases were inconclusive
Given the potential for inaccuracies using AI detection models, Applied's approach is to provide a referent AI-generated answer to view alongside candidate responses.
In our research, we found that users naturally used their pattern recognition skills when suspecting that candidate answers had been plagiarised; observing similar structuring of paragraphs, common wording and a basic level of depth in the suspected answers. These suspicions were often corroborated when the Sift Questions were put into Large Language Models (LLMs).
Further, this approach is less burdensome for users than comparing candidate answers to each other for pattern recognition, saving time for hiring teams and providing better consistency and accuracy.
Where are the referent AI answers visible?
Referent AI answers are visible by members of a role's hiring team who are reading and scoring answers to Sift Questions and Work Samples. They are displayed in a tab next to the review guides set up by the job admin.
How can the Referent AI answer be used?
This will depend on your company policies regarding using Generative AI. The goal of providing the Referent AI answer is to give reviewers more control over the hiring process, making it easier to discern if candidate answers have used Generative AI, and to what extent it may have been used. Using this as a guide, you may wish to then flag certain candidate answers.
It is worth noting that, while the content of the AI answers can differ between users, the structure and format is the most predictive indicator that an answer has been generated.
How are the Referent AI answers generated?
To generate our answers we use OpenAI, the research laboratory and creators of ChatGPT. Of the current LLMs on the market, it is the most widely used and offers an Application Programming Interface (API) that best suits our software. We are currently using version gpt-3.5-turbo.
How does Applied generate the answers?
As soon as a job admin allocates sift answers for review, Applied triggers a request to OpenAI to generate a 250-word answer for each of the questions created by hiring teams for that specific role. The same generated response for each question is then stored on our system and visible to each reviewer for that role. This ensures consistency in the referent information being used throughout the review process and makes it easier for the hiring team to trace back any flagged answers.
Can reviewers generate new Referent AI answers?
To ensure consistency across the review process, it's important for all reviewers to be working from the same referent answer. For this reason, reviewers cannot generate a new answer to the question from within Applied itself. The answer can, however, change over time if a job admin updates the question itself and the reviews are reallocated.
Will the Referent AI answer change over time?
The referent AI answer will not change within a specific role unless a specific action is taken. A job admin would need to edit the question and reallocate it for review to generate a new answer.
If the same sift question is being utilised across different roles, Applied triggers an AI generation process specific to each role at the point of review allocation. This means that teams will likely see a different AI answer for Job A than for Job B. We have learned while researching this solution, however, that the patterns of the answers are similar (structures, length of paragraphs) despite potentially referring to different topics asked by a question.