Artificial Intelligence-based interview scoring learns from past interview answers, which makes it hard for it to determine if a candidate is legitimately answering the question if their response includes context or an example rarely seen in training data.
Moreover, AI interview platforms may be susceptible to adversarial inputs where an irrelevant answer may receive a high score. Both scenarios raise fairness concerns and can erode trust in AI job interviews (Madaio et al, 2020).
This is why identifying outliers that differ significantly from the majority of answers and flagging them for manual review become crucial steps toward responsible and fair use of AI interview software. While simple rule-based methods (Reiz and Pongor, 2011) could help filter out some irrelevant answers based on answer length and regular expressions, these methods do not take into account the context and content of the answer and question. Someone may describe a very unique, yet relevant, situation in response to an AI for interviews question, which you wouldn’t want to disregard.
In this study, we introduce an unsupervised, question-aware, multi-context outlier detection model that can help detect anomalous answers contextually and semantically. The unsupervised approach is deemed to be more practical compared to a supervised model that requires a large labeled dataset of outlier answers. It helps bootstrap an outlier detector that can then be enhanced through human feedback.
We tested the outlier model to ascertain how well it is able to correctly identify 177,691 actual hired candidate interview answers from outliers, (e.g., movie reviews, news articles, nonsensical text, and sentences generated using BERT (Vaswani et al, 2017) with random starting words).
Our model outperformed the baseline One-class SVM outlier detector (Li et al, 2003), in detecting outliers from actual interview answers. The performance of our model over the baseline unsupervised model can be explained by both question-aware learning and multi-context learning, which help yield better contextual representations for detecting outlier answers from typical interview answers.
We also conducted a human evaluation on 10,689 interview answers of candidates who were not hired and might have provided outlier answers. Our model predicted 0.16% of the answers as outliers with only 5.9% of them being false positives. All of these false predictions describe contexts related to family and personal life in their answers but are relevant to the question. It is reasonable that these answers are labeled as an outlier by our model since they are contextually and semantically different from most interview answers.
While a data-driven AI interviewer can help counter flaws in human interviewers, answers that are significantly different to training data can lead to spurious predictive outcomes. In this study, we show how a
References:
Dai, Y., Qi, J., & Zhang, R. (2020). Joint recognition of names and publications in academic homepages. In Proceedings of the 13th International Conference on Web Search and Data Mining (pp. 133-141).
Li, K. L., Huang, H. K., Tian, S. F., & Xu, W. (2003). Improving one-class SVM for anomaly detection. In Proceedings of the 2003 international conference on machine learning and cybernetics (IEEE Cat. No. 03EX693) (Vol. 5, pp. 3077-3081). IEEE.
Madaio, M. A., Stark, L., Wortman Vaughan, J., & Wallach, H. (2020). Co-designing checklists to understand organizational challenges and opportunities around fairness in AI. In Proceedings of the 2020 CHI Conference on Human Factors in Computing Systems (pp. 1-14).
Reiz, B., & Pongor, S. (2011). Psychologically Inspired, Rule-Based Outlier Detection in Noisy Data. In SYNASC (pp. 131-136)
Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A. N., … & Polosukhin, I. (2017). Attention is all you need. Advances in neural information processing systems, 30.
We can’t hide from reality anymore. Talent needs are shifting overnight, and AI is redefining what it means to work. Traditional talent frameworks are no longer fit for purpose. At Sapia.ai, we believe the future of talent strategy lies in a smarter, fairer, and more adaptive way of defining what great looks like.
Our AI hiring platform is built on the largest proprietary dataset of interview answers globally – we’re a data company at heart, and we’ve seen the power of data-driven people methodology in transforming how organisations hire and retain good talent.
So, when it came to building a new Competency Framework that could be leveraged globally for hiring for any role at any scale, of course, we used a ground-up, data-led methodology that bridges the gap between organisational psychology and AI.
Conventional frameworks are typically crafted through expert interviews and focus groups. While valuable, they tend to be subjective, static, and too slow to keep pace with evolving job demands. As roles become more fluid and technology augments or replaces task-based skills, organisations need a new way to understand the human capabilities that genuinely matter for performance.
We wanted to identify enduring, job-agnostic competencies that reflect what drives success in a modern workplace – capabilities like adaptability, resilience, learning agility, and customer orientation.
(Why competencies and not just skills? Read why here.)
Sapia.ai’s methodology is rooted in the science of human behaviour but powered by cutting-edge AI. We asked two core questions:
The answer to both: yes.
We began with a rich dataset of over 37,000 job descriptions across industries and role types. Using large language models (LLMs) and advanced NLP techniques, we extracted over 200,000 behavioural descriptors. These were distilled down through a four-step process:
This resulted in a refined list of 25 human-centric competencies, each with clear behavioural indicators and practical relevance across a wide range of roles.
Our framework is intelligent, but importantly, it’s adaptive. Organisations can apply this methodology to their own job descriptions to discover custom competencies. This bottom-up, role-data-led approach ensures alignment to real work, not just theoretical models.
And because the framework integrates directly with our AI-powered hiring tools, you get a connected system that brings your talent strategy to life.
Our framework comes to life in the following tools:
Skills alone cannot predict success. Competencies do. As AI continues transforming how we work, Sapia.ai’s Competency Framework offers a scalable, scientific, and fair foundation for hiring and developing the talent of tomorrow.
If you’re a CHRO or Head of Recruitment at an enterprise today, chances are you’ve been inundated with messages about the importance of “skills-based hiring.” LinkedIn’s recent Work Change Report (2025) is full of compelling data: a 140% increase in the rate at which professionals are adding new skills to their profiles since 2022, and a projection that by 2030, 70% of the skills used in most jobs today will have changed.
This is essential reading. But there’s a missed opportunity: the singular focus on “skills” fails to acknowledge the real metric that talent leaders need to be using to future-proof their workforce — competencies.
But skills on their own — even soft ones — are generic, disjointed, and often disconnected from real-world performance. In contrast:
Put simply, competencies answer the all-important question: Can this person apply the right skills, in the right way, at the right time, to deliver results in our environment?
The Work Change Report outlines a future where job titles are fluid, roles evolve quickly, and AI is a constant disruptor. This creates three massive challenges for hiring at scale:
Skills alone don’t tell us whether someone can succeed in a role that will look different 12 months from now. But competencies can. Because they measure not just what a person knows, but how they apply it.
The LinkedIn report highlights a critical insight: organisations now prioritise agility in entry-level hiring. And there’s a good reason for that. With professionals expected to hold twice as many jobs over their careers compared to 15 years ago, adaptability is not just a nice-to-have. It’s core to success.
But you can’t measure agility with a keyword on a CV. You measure it by looking at competencies like:
When you shift the focus away from skills to behavioural competencies that can be defined, observed, and assessed in structured ways, you open yourself up to a much more dynamic and more useful way of managing talent.
To hire effectively at scale, particularly in a technology-driven world of work, talent leaders must shift their lens:
LinkedIn’s data shows that people are learning more skills more quickly than ever. But the real question for talent leaders like you is: Are those skills being applied in ways that drive value? Are we hiring for task proficiency or performance?
The truth is that the organisations that will thrive in an AI-driven, skills-fluid economy aren’t the ones chasing the next hot skill. They’re the ones designing systems to identify, develop and scale competence.
Sapia.ai has developed a comprehensive Competency Framework using a data-driven approach. Download the full paper here.
Every day, we read stories of increased fake or AI-assisted applications. Tools like LazyApply are just one of many flooding the market, driving up applicant volumes to never-before-seen levels.
As an overwhelmed hiring function, how do you find the needle in the haystack without using an army of recruiters to filter through the maze?
At Sapia.ai, we help global enterprises do just that. Many of the world’s most trusted brands, such as Qantas Group, have relied on our hiring platform as a co-pilot for better hiring since 2020.
Our Chat Interview has given millions of candidates a voice they wouldn’t have had – enabling them to share in their own words why they’re the best fit for the role. To find the people who belong with their brands, our customers must trust that their candidates represent themselves. Thus, they want to trust that our AI is analysing real human answers—not answers from a machine.
The Rise of GPT
When ChatGPT went viral in November 2022, we immediately adopted a defensive strategy. We had long been flagging plagiarised candidate responses, but then, we needed to act fast to flag responses using artificially generated content (‘AGC’).
Many companies were in the same position, but Sapia.ai was the only company with a large proprietary data set of interview answers that pre-dated GPT and similar tools: 2.5 billion words written by real humans.
That data enabled us to build a world-first:- an LLM-based AGC detector for text-based interviews, recently upgraded to v2.0 with 99% accuracy and a false positive rate of 1%. An NLP classification model built on Sapia.ai proprietary data that operates across all Sapia.ai chat interviews.
Full Transparency with Candidates
Because we value candidate trust as much as customer trust, we wanted to be transparent with candidates about our ability to detect artificially generated content (AGC). As an LLM, we could identify AGC in real time and warn candidates that we had detected it.
This has had a powerful impact on candidate behaviour. Since our AGC detector went live, we have seen that the real-time flagging acts as a real-time disincentive to use tools like ChatGPT to generate interview responses.
The detector generates a warning if 3 or more answers are flagged as having artificially generated content. The Sapia.ai Chat Interview uses 5 open-ended interview questions for volume hiring roles, such as retail, contact centre, and customer service, and 6 questions for professional roles, such as engineers, data scientists, graduates, etc.
Let’s Take a Closer Look at the Data…
We see that using our AGC detector LLM to communicate live with candidates in the interview flow when artificial content has been detected has a positive effect on deterring candidates from using AI tools to generate their answers.
The rate of AGC use declines from 1 question flagged to 5 questions – raising the flag on one question is generally enough to deter candidates from trying again.
The graph below shows the number of candidates, from a total of almost 2.7m, that used artificially generated content in their answers.
Differences in AGC Usage Rate by Groups
We see no meaningful differences in candidate behaviour based on the job they are applying for or based on geography.
However, we have found differences by gender and ethnicity – for example, men use artificially generated content more than women. The graph below shows the overall completion ratios by gender – for all interviews on the left and for interviews where the number of questions with AGC detected is 5 or more on the right.
Perception of Artificially Generated Content by Hirers.
We’re curious to understand how hirers perceive the use of these tools to assist candidates in a written interview. The creation of the detector was based on the majority of Sapia.ai customers wanting transparency & explainability around the use of these tools by candidates, often because they want to ensure that candidates are using their own words to complete their interviews and they want to avoid wasting time progressing candidates who are not as capable as their chat interview suggests.
However, some of our customers feel that it’s a positive reflection of the candidate, showing that they are using the tools available to them to put their best foot forward.
It’s a mix of perspectives.
Our detector labels it as the use of artificially generated content. It’s up to our customers how they use that information in their decision-making processes.
This concept of having a human in the loop is one of the key dimensions of ethical AI, and we ensure that it is used in every AI-related hiring product we build.
Interested in the science behind it all? Download our published research on developing the AGC detector 👇