Discrimination based on race and ethnicity in personnel selection is a well known and pervasive issue highlighted in numerous studies (Bertrand & Mullainathan, 2004; Kline et al., 2021; Pager et al., 2009).
Most of these studies report name-based inference of race and ethnicity by human reviewers leading to differential outcomes in the recruitment process. Linguistic racism is a form of discrimination that occurs based on one’s use of language, especially English (De Costa, 2020), and is highly associated with race and ethnicity.
As machine learning models are adopted to automate tasks like interview scoring, race or ethnicity encoded signals in language can lead to biased outcomes, if not mitigated. Hence understanding the level of ethnicity encoded signals in language is important when building natural language-based machine learning models in order to avoid biased outcomes, for example by using feature scores rather than raw text to score responses (Jayaratne, Jayatilleke, Dai, 2022).
In this work, we sought to quantify and compare the amount of ethnicity encoded information in over 300,000 candidates’ raw text interview responses to language-derived feature scores, including personality, behavioral competencies, and communication skills.
First, we trained machine learning models to predict candidate ethnicity from raw-text chat interview responses. Specifically, we trained an Attention-Based Bidirectional Long Short-Term Memory (Attn-BiLSTM) (Zhou et al., 2016) model for predicting ethnicity from textual responses.
Secondly, we tested the same for the language-derived features used in the automated scoring of the interview responses. We trained multiple models using a variety of machine learning algorithms (a linear model, tree models with bagging and boosting, and a neural network model with a single hidden layer) suitable for tabular data for predicting ethnicity from the 21 derived features.
Each model was then used to predict ethnicity for the 10% of the sample left out of the training dataset. The results from the classification tasks show a clear distinction between the ability to infer ethnicity based on natural language and inferred features. As hypothesized, we found that features derived according to a clearly defined rubric contain significantly less ethnicity information compared to raw candidate responses. That is, the models based on derived features recorded consistently weaker accuracy, precision, recall, and F1 values across all models compared to the model for the raw text candidate responses.
This research demonstrates the benefit of using algorithmically derived feature values in mitigating ethnicity related biases when scoring structured interview responses. Specifically, our results show that natural language responses to interview questions carry higher amounts of ethnicity information compared to features derived according to a clearly defined rubric for assessing interview responses. This further strengthens the case for using structured interviews that have been shown to reduce bias over unstructured interviews (Levashina et al., 2014) with much stronger criterion validity (Sackett et al., 2021).
References:
Bertrand, M., & Mullainathan, S. (2004). Are Emily and Greg More Employable Than Lakisha and Jamal? A Field Experiment on Labor Market Discrimination. American Economic Review, 94(4), 991–1013.
De Costa, P. I. (2020). Linguistic racism: Its negative effects and why we need to contest it. International Journal of Bilingual Education and Bilingualism, 23(7), 833–837.
Jayaratne, M., Jayatilleke, B., & Yimeng Dai (2022). Identifying and mitigating gender bias in structured interview responses [Paper presentation]. 2022 Society for Industrial Organizational Psychology Conference. Seattle, Washington, United States.
Kline, P. M., Rose, E. K., & Walters, C. R. (2021). Systemic Discrimination Among Large U.S. Employers (NBER Working Papers No. 29053). National Bureau of Economic Research, Inc.
Levashina, J., Hartwell, C. J., Morgeson, F. P., & Campion, M. A. (2014). The Structured Employment Interview: Narrative and Quantitative Review of the Research Literature. Personnel Psychology, 67(1), 241–293.
Pager, D., Western, B., & Bonikowski, B. (2009). Discrimination in a Low-Wage Labor Market: A Field Experiment. American Sociological Review, 74(5), 777–799.
Sackett, P. R., Zhang, C., Berry, C. M., & Lievens, F. (2021). Revisiting meta-analytic estimates of validity in personnel selection: Addressing systematic overcorrection for restriction of range. Journal of Applied Psychology.
Zhou, P., Shi, W., Tian, J., Qi, Z., Li, B., Hao, H., & Xu, B. (2016). Attention-Based Bidirectional Long Short-Term Memory Networks for Relation Classification. Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers), 207–212.
Related
Blog
New Research Proves the Value of AI Hiring
A new study has just confirmed what many in HR have long suspected: traditional psychometric tests are no longer the gold standard for hiring.
Published in Frontiers in Psychology, the research compared AI-powered, chat-based interviews to traditional assessments, finding that structured, conversational AI interviews significantly reduce social desirability bias, deliver a better candidate experience, and offer a fairer path to talent discovery.
We’ve always believed hiring should be about understanding people and their potential, rather than reducing them to static scores. This latest research validates that approach, signalling to employers what modern, fair and inclusive hiring should look like.
The problem with traditional psychometric tests
While used for many decades in the absence of a more candidate-first approach, psychometric testing has some fatal flaws.
For starters, these tests rely heavily on self-reporting. Candidates are expected to assess their own traits. Could you truly and honestly rate how conscientious you are, how well you manage stress, or how likely you are to follow rules? Human beings are nuanced, and in high-stakes situations like job applications, most people are answering to impress, which can lead to less-than-honest self-evaluations.
This is known as social desirability bias: a tendency to respond in ways that are perceived as more favourable or acceptable, even if they don’t reflect reality. In other words, traditional assessments often capture a version of the candidate that’s curated for the test, not the person who will show up to work.
Worse still, these assessments can feel cold, transactional, even intimidating. They do little to surface communication skills, adaptability, or real-world problem solving, the things that make someone great at a job. And for many candidates, especially those from underrepresented backgrounds, the format itself can feel exclusionary.
The Rise of Chat-Based Interviews
Enter conversational AI.
Organisations have been using chat-based interviews to assess talent since before 2018, and they offer a distinctly different approach.
Rather than asking candidates to rate themselves on abstract traits, they invite them into a structured, open-ended conversation. This creates space for candidates to share stories, explain their thinking, and demonstrate how they communicate and solve problems.
The format reduces stress and pressure because it feels more like messaging than testing. Candidates can be more authentic, and their responses have been proven to reveal personality traits, values, and competencies in a context that mirrors honest workplace communication.
Importantly, every candidate receives the same questions, evaluated against the same objective, explainable framework. These interviews are structured by design, evaluated by AI models like Sapia.ai’s InterviewBERT, and built on deep language analysis. That means better data, richer insights, and a process that works at scale without compromising fairness.
Key Findings from the Latest Research
The new study, published in Frontiers in Psychology, put AI-powered, chat-based interviews head-to-head with traditional psychometric assessments, and the results were striking.
One of the most significant takeaways was that candidates are less likely to “fake good” in chat interviews. The study found that AI-led conversations reduce social desirability bias, giving a more honest, unfiltered view of how people think and express themselves. That’s because, unlike multiple-choice questionnaires, chat-based assessments don’t offer obvious “right” answers – it’s on the candidate to express themselves authentically and not guess teh answer they think they would be rewarded for.
The research also confirmed what our candidate feedback has shown for years: people actually enjoy this kind of assessment. Participants rated the chat interviews as more engaging, less stressful, and more respectful of their individuality. In a hiring landscape where candidate experience is make-or-break, this matters.
And while traditional psychometric tests still show higher predictive validity in isolated lab conditions, the researchers were clear: real-world hiring decisions can’t be reduced to prediction alone. Fairness, transparency, and experience matter just as much, often more, when building trust and attracting top talent.
Sapia.ai was spotlighted in the study as a leader in this space, with our InterviewBERT model recognised for its ability to interpret candidate responses in a way that’s explainable, responsible, and grounded in science.
Why Trust and Candidate Agency Win
Today, hiring has to be about earning trust and empowering candidates to show up as their full selves, and having a voice in the process.
Traditional assessments often strip candidates of agency. They’re asked to conform, perform, and second-guess what the “right” answer might be. Chat-based interviews flip that dynamic. By inviting candidates into an open conversation, they offer something rare in hiring: autonomy. Candidates can tell their story, explain their thinking, and share how they approach real-world challenges, all in their own words.
This signals respect from the employer. It says: We trust you to show us who you are.
Hiring should be a two-way street – a long-held belief we’ve had, now backed by peer-reviewed science. The new research confirms that AI-led interviews can reduce bias, enhance fairness, and give candidates control over how they’re seen and evaluated.
Blog
AI Maturity in the Enterprise
Barb Hyman, CEO & Founder, Sapia.ai
It’s time for a new way to map progress in AI adoption, and pilots are not it.
Over the past year, I’ve been lucky enough to see inside dozens of enterprise AI programs. As a CEO, founder, and recently, judge in the inaugural Australian Financial Review AI Awards.
And here’s what struck me:
Despite the hype, we still don’t have a shared language for AI maturity in business.
Some companies are racing ahead. Others are still building slide decks. But the real issue is that even the orgs that are “doing AI” often don’t know what good looks like.
You don’t need more pilots. You need a maturity model.
The most successful AI adoption strategy does not have you buying the hottest Gen AI tool or spinning up a chatbot to solve one use case. What it should do is build organisational capability in AI ethics, AI governance, data, design, and most of all, leadership.
It’s time we introduced a real AI Maturity Model. Not a checklist. A considered progression model. Something that recognises where your organisation is today and what needs to evolve next, safely, responsibly, and strategically.
Here’s an early sketch based on what I’ve seen:
The 5 Stages of AI Maturity (for real enterprises)
- Curious
- Awareness is growing across leadership
- Experimentation led by innovation teams
- Risk is unclear, appetite is cautious
- AI is seen as “tech”
- Reactive
- Gen AI introduced via vendors or tools (e.g., copilots, agents)
- Some pilots show promise, but with limited scale or guardrails
- Data privacy and sovereignty questions begin to surface
- Risk is siloed in legal/IT
- Capable
- Clear policies on privacy, bias, and governance
- Dedicated AI leads or councils exist
- Internal use cases scale (e.g., summarisation, scoring, chat)
- LLMs integrated with guardrails, safety reviewed
- Strategic
- AI embedded in workflows, not layered on
- LLM/data infrastructure is regionally compliant
- AI outcomes measured (accuracy, equity, productivity)
- Teams restructured around AI capability — not just tech enablement
- AI-Native
- AI informs and transforms core decisions (hiring, pricing, customer service)
- Enterprise builds proprietary intelligence
- FAIR™/RAI principles deeply operationalised
- Talent, systems, and leadership are aligned around an intelligent operating model
Why this matters for enterprise leaders
AI is a capability.And like any capability, it needs time, structure, investment, and a map.
If you’re an HR leader, CIO, or enterprise buyer, and you’re trying to separate the real from the theatre, maturity thinking is your edge.
Let’s stop asking, “Who’s using AI?”
And start asking: “How mature is our AI practice and what’s the next step?”
I’m working on a more complete model now, based on what I’ve seen in Australia, the UK, and across our customer base. If you’re thinking about this too, I’d love to hear from you.
Blog
Beyond the Black Box: Why Transparency in AI Hiring Matters More Than Ever
For too long, AI in hiring has been a black box. It promises speed, fairness, and efficiency, but rarely shows its work.
That era is ending.
“AI hiring should never feel like a mystery. Transparency builds trust, and trust drives adoption.”
At Sapia.ai, we’ve always worked to provide transparency to our customers. Whether with explainable scores, understandable AI models, or by sharing ROI data regularly, it’s a founding principle on which we build all of our products.
Now, with Discover Insights, transparency is embedded into our user experience. And it’s giving TA leaders the clarity to lead with confidence.
Transparency Is the New Talent Advantage
Candidates expect fairness. Executives demand ROI. Boards want compliance. Transparency delivers all three.
Even visionary Talent Leaders can find it difficult to move beyond managing processes to driving strategy without the right data. Discover Insights changes that.
“When talent leaders can see what’s working (and why) they can stop defending their strategy and start owning it.”
Metrics That Make Transparency Real (and Actionable)
🕒 Time to Hire
What it is: The median time between application and hire.
Why it matters: This is your speedometer. A sharp view of how long hiring takes and how that varies by cohort, role, or team helps you identify delays and prove efficiency gains to leadership.
Faster time to hire = faster access to revenue-driving talent.
💬 Candidate Sentiment, Advocacy & Verbatim Feedback
What it is: Satisfaction scores, brand advocacy measures, and unfiltered candidate comments.
Why it matters: Many platforms track satisfaction. Sapia.ai’s Discover Insights takes it further, measuring whether that satisfaction translates into employer and consumer brand advocacy.
And with verbatim feedback collected at scale, talent leaders don’t have to guess how candidates feel. They can read it, learn from it, and take action.
You don’t just measure experience. You understand it in the candidates’ own words.
🔍 Drop-Off Rates, Funnel Visibility & Automation That Works
What it is: The percentage of candidates who exit the hiring process at different stages, and how to spot why.
Why it matters: Understanding drop-off points lets teams fix friction quickly. Embedding automation early in the funnel reduces recruiter workload and elevates top candidates, getting them talking to your hiring teams faster.
Assessment completion benchmarks in volume hiring range between 60–80%, but with a mobile-first, chat-based format like Sapia.ai’s, clients often exceed that.
Optimising your funnel isn’t about doing more. It’s about doing smarter, with less effort and better outcomes.
📈 Hiring Yield (Hired / Applied)
What it is: The percentage of completed applications that result in a hire.
Why it matters: This is your funnel efficiency score. A high yield means your sourcing, screening, and selection are aligned. A low one? There’s leakage, misfit, or missed opportunity.
Hiring yield signals funnel health, recruiter performance, and candidate-process fit.
🧠 AI Effectiveness: Score Distribution & Answer Originality
What it is: Insights into how candidate scores are distributed, and whether responses appear copied or AI-generated.
Why it matters: In high-volume hiring, a normal distribution of scores suggests your assessment is calibrated fairly. If it’s skewed too far left or right, it could be too hard or too easy, and that affects trust.
Add in answer originality, and you can track engagement integrity, protecting both your process and your brand.
From Metrics to Momentum
To effectively lead, you need more than simply tracking; you need insights enabling action.
When you can see how AI impacts every part of your hiring, from recruiter productivity to candidate sentiment to untapped talent, you lead with insight, not assumption. And that’s how TA earns a seat at the strategy table.