AI in Recruitment: Fair or Biased?

[00:00:00] Lauren:
Screening is fair and consistent, whether it’s done by a human or a machine. We would never allow AI scoring to run without human involvement. There is always a human wraparound in place. We have the upfront calibration and ongoing oversight.

[00:00:23] Darren:
Welcome everyone to the Talent Download podcast. This is our first episode, and I’m delighted to welcome:

• Martin Kavanaugh, Head of Assessment at Amberjack
• Lauren Edge, Principal Consultant within Assessment

Both are occupational psychologists who do an amazing job for us.

Today we’re discussing AI in the screening process. Recruitment is currently full of noise about AI tools, and I’m proud of how Amberjack uses technology to improve the candidate experience. We’ve developed a live tool that screens video interviews, but a big question remains: Is it fair? Is it fair for organisations to use bots to support screening?

Those are the questions we want to explore today. We’ll touch on the science, our thinking, and where the next phase might take us. But first, before talking about AI, we should explain our quality assurance (QA) process. From an Amberjack perspective, what does that look like?

[00:02:33] Lauren:
The most important thing in screening is ensuring fairness, consistency, and the absence of bias. At Amberjack we have a dedicated QA team and strong QA principles.

When screeners join us, they receive rigorous training based on British Psychological Society best practice. But humans are humans: tiredness, distractions, or unclear scoring guidelines can affect performance.

So QA is essential. One thing we’re particularly proud of is our calibration process. Before screeners begin volume scoring, we sit with them, score sample candidates together, review best practice, and ensure alignment. We involve the client in this as well.

After calibration, screeners move into volume, and QA continues. A QA consultant reviews scores, provides feedback, and if needed, temporarily removes screeners from volume until realigned.

[00:04:35] Darren:
You recently wrote a blog about this, and you mentioned that good people can make the wrong decisions. Can you share an example?

[00:05:04] Lauren:
Absolutely. When scoring guides are overly complex or cognitively demanding, screeners can struggle to apply them consistently, especially if they’re tired or distracted. This can impact scoring.

That’s why we simplify scoring guidelines before going live — so assessors can use them effectively.

[00:05:50] Darren:
And someone scoring at 9am on a Monday may behave differently from someone scoring at 5pm on a Friday. QA helps manage that. When it comes to tech, though, AI is constant — which we’ll come back to.

How do our human screeners react to feedback?

[00:06:31] Martin:
Our people genuinely care about getting decisions right. Even with huge volumes — sometimes 10,000 video interviews per campaign — they see the individual behind each response. They know how important these decisions are for candidates.

They’re very receptive to feedback. In fact, they’re often harder on themselves than we are. Mistakes happen in any volume process, which is why QA exists, but they’re always focused on improvement.

[00:07:56] Darren:
Application rates have increased massively. In 2002, the average was around 35 applications per job; now it’s about 145. That matters for how we operate.

What are the key QA principles we must follow?

[00:08:45] Martin:
Two main principles:

1. Use a small, highly trained team.
   As volumes grow, training and consistency become even more important.
2. Design assessment materials that set screeners up for success.
   Indicators must be clear, simple, and directly connected to what candidates are asked to demonstrate. High cognitive load reduces scoring accuracy.

[00:10:27] Darren:
So, let’s talk about AI. Why introduce AI if we already had strong human QA?

[00:10:53] Martin:
Because of scale. Application volumes are rising quickly. Our previous approach worked, but with growth — both in the market and within Amberjack — we needed additional tools to maintain quality and efficiency.

[00:11:27] Darren:
Is AI essential in modern recruitment?

[00:12:02] Martin:
Yes — but only if implemented safely. There has been too much imperfect action in the industry. Some organisations rushed into AI without proper oversight, which creates risk.

We’ve taken a careful, transparent, science-led approach, which builds trust with clients.

[00:13:16] Darren:
When speaking to clients, what problems are we solving with AI?

[00:13:39] Martin:
Consistency, scale, and alignment to campaign-specific behaviours. We calibrate the AI to each client’s indicators, context, and role requirements. It isn’t a generic model — it’s tailored.

[00:14:27] Darren:
Lauren, how do we blend AI with the human element?

[00:14:44] Lauren:
We follow the same principles as with human screeners. The AI goes through a calibration process before volume scoring. It must demonstrate that it can use the scoring guides effectively and provide rationales for its scores.

Humans remain involved throughout:

• We QA at least 30% of AI‑scored interviews.
• Candidates can opt in to AI.
• If they don’t consent, humans score their interviews.
• The AI scores only the transcript, not video or visual data.
• If it’s unsure about a transcript or a score, a human reviews it.
• At this stage, anything below 90% confidence goes to a human.

[00:16:56] Darren:
So just like a Teams meeting transcript — AI sees only the words?

[00:16:56] Lauren:
Exactly. It doesn’t see the candidate’s appearance or environment — only their words.

[00:17:38] Darren:
In your blog, you mentioned “AI with a human wraparound.” What does that mean?

[00:17:38] Lauren:
It means AI is never left unsupervised. We calibrate upfront, monitor scoring patterns, analyse demographic data, and intervene if needed. If anything unexpected appears, we pause and adjust. Humans always oversee the process.

[00:18:32] Martin:
We also focused heavily on candidate communication. We explain what AI is doing, why, and what safeguards exist. Candidates can opt out at any point. Because of that transparency, 90% choose AI scoring.

[00:19:57] Darren:
What’s the candidate experience like?

[00:20:13] Lauren:
Candidates:

1. Enter the blended assessment
2. Watch an intro video
3. Accept the privacy policy
4. Reach a dedicated AI consent page explaining fairness, QA, and safeguards
5. Tick a clear “I consent” box

It’s transparent and easy to understand.

[00:21:02] Martin:
Research suggests candidates prefer human screening, mainly because they don’t trust AI. But when communication is transparent and supportive, they do trust it. Early trials show positive feedback.

[00:22:51] Lauren:
We track NPS, and feedback is very positive. AI helps prevent bottlenecks and reduces waiting times, improving the overall experience.

[00:24:01] Martin:
We measure every campaign based on:

1. Candidate quality
2. Candidate sentiment
3. Efficiency
4. Diversity outcomes

AI is only released on campaigns once we’re confident in all four areas.

[00:25:23] Lauren:
Exactly. When AI has low confidence in transcription or scoring, a human automatically reviews it.

[00:25:58] Darren:
There’s a lot of talk about AI bias. Have we seen any bias in our tests?

[00:26:24] Martin:
No — we monitor demographic progression constantly. Clients can see the data in real time, and we adjust if anything unexpected appears. We’ve been recognised for our real‑time improvements, including awards and shortlisting from the Association for Business Psychology.

[00:27:34] Darren:
How have clients responded to the AI product?

[00:28:05] Lauren:
Overall, very positively. Some clients who previously used AI elsewhere were more cautious, but our data reassures them.

We currently use AI only on clients with a 1–5 scoring scale because we know it performs consistently there. Other scales will come in future.

Some clients who don’t use Amberjack for screening are interested in adopting our AI tool because of resource challenges or inconsistency in their own internal processes.

[00:29:12] Darren:
Is the QA process different for AI versus humans?

[00:29:28] Lauren:
It’s similar — but even more rigorous. We took our refined human QA principles and applied them to AI, with additional checks.

[00:29:51] Darren:
Does feedback for candidates change?

[00:30:28] Lauren:
If anything, it improves. Human feedback varies depending on the comments screeners leave. With AI, we automatically provide consistent rationales for each score, meaning candidates receive more detailed, structured feedback.

[00:31:20] Darren:
Are there misconceptions about AI screening?

[00:31:48] Martin:
Yes — but mainly misconceptions about how others apply AI. Many organisations rushed in without safeguards. When we explain our process — calibration, client involvement, transparency — clients become much more comfortable.

[00:33:16] Lauren:
Some clients who screen internally are interested in using our AI to support resource gaps or improve consistency.

[00:33:50] Darren:
Application numbers are high, and organisations need smarter processes. Even outside early careers, volume roles create pressure. AI can help ease that pressure.

What’s next for Amberjack?

[00:36:25] Martin:
We’re exploring exciting applications:

1. Candidate management powered by AI

AI can help candidates feel more comfortable raising concerns or requesting adjustments. Early trials show candidates are more open with AI about sharing their needs.

2. Positive action support

Lauren is leading a project using AI to provide real‑time guidance to candidates from underrepresented groups — for example, through WhatsApp — giving them information and reassurance at key moments without giving unfair advantage.

3. AI within assessments themselves

We’re exploring how AI can evaluate responses within real assessment tasks — but more on that in a future episode.

[00:40:23] Darren:
Do you think the QA process will change as AI evolves?

[00:40:23] Lauren:
Not in the short or medium term. Consistency is essential, and we’ll always maintain human wraparound. As more data emerges, we may adjust thresholds or processes, but human oversight will remain.

[00:41:04] Darren:
In one sentence each — what does fair AI mean?

[00:41:31] Martin:
Fair AI means building trust with clients and candidates.

[00:41:44] Lauren:
Trust and consistency.

[00:41:50] Darren:
To finish, some quick true-or-false questions. Answer at the same time.

AI will replace assessors.
Martin: True.
Lauren: True.

AI can’t be trusted to make hiring decisions.
Martin: False.
Lauren: False — with the caveat that humans must be involved.

AI screening is faster but less fair.
Both: False.

Speed matters for candidate experience, especially with rising application volumes. Candidates are excited and ambitious, and we must support them with efficient processes. I think we’re on the right track.

Thank you both for your time. If anyone has follow‑up questions, please contact us at Amberjack. And do read Lauren’s blog about this topic on our website.