00:00 Introduction
01:38 Jarret’s strangest interview question
02:43 What AI really looks like in talent assessment today
04:03 Where AI excels, and where it still struggles
09:03 Agentic AI explained
11:04 The three layers of AI guardrails
14:10 Building AI people can trust
16:59 Fairness in AI from a tech perspective
20:08 The future of AI in assessments: adaptive, personalised, multimodal
22:21 Science or fiction: rapid-fire myths and truths about AI in hiring
30:30 Final takeaways and reflections
Nicola Tatham (00:00)
Hello everybody and welcome to The Score. This is where we make sense of hiring trends, sort science from fiction, and find out what's new in talent acquisition from some of the top experts in the field. I'm Nicola Tatham, Chief IO Psychologist here at Sova, with over two decades of experience designing fair, predictive, and science-backed assessments. I'm here to cut through the noise and talk about what actually works in hiring.
Caroline Fry (00:27)
And I'm Caroline Fry, Head of Product at Sova. I spend my time turning smart assessment ideas from Nic into tools that are scalable, inclusive, and ready for whatever AI throws at us next.
Nicola Tatham (00:39)
Each episode, we're joined by a guest to unpack a big question in hiring.
Caroline Fry (00:44)
Because talent deserves better than guesswork. Today we're joined by Jarret Hardie, Chief Technology Officer at Sova. Jarret leads the architecture, development, and ongoing evolution of the Sova Talent Assessment Platform, and he's our go-to oracle for all things AI. We'll be picking his brain to find out where he sees the future of talent assessment heading, especially when it comes to AI and emerging tech. Jarret is also my boss, so I'm going to be on my best behaviour.
Nicola Tatham (01:12)
Before we dive into the serious stuff, let's warm up with a quick one. We always like to ask our guests about the strangest, most entertaining, or just plain awful question they've ever been asked in an interview or a hiring process. We've heard everything from "If you were a kitchen utensil, what would you be?" to asking someone to solve a riddle. So Jarret, what's your most memorable one?
Jarret Hardie (01:38)
Thank you for having me. It's great to be here, and hopefully we'll have an enjoyable conversation today. To kick us off, I was trying to think about the strangest interview question I've ever had, and it's really more of an interview snafu. Early in my career, I was interviewing for a role as a software developer. It was a government position, and I was asked what my experience was with "Microsoft Point." I think the interviewer meant Microsoft SharePoint, but I thought he meant Microsoft Paint. We ended up having a very entertaining discussion about brush sizes, colours, and font choice. At no point did he stop me and say, "No, you've got the wrong end of the stick here."
Caroline Fry (02:15)
Did you get the job?
Jarret Hardie (02:28)
I did not get the job, needless to say.
Caroline Fry (02:32)
Let's start with a broad question. We hear a lot about AI in recruitment now, and some of it sounds like science fiction. What does AI actually look like day-to-day in talent assessment?
Jarret Hardie (02:43)
It's a really interesting question. There's certainly a lot of hype, but right now AI in assessment is more practical than science fiction. The biggest use is probably automatic scoring. Scoring an open-ended response used to take hours of manual review. Now AI can score pre-recorded video interview responses or written answers to problem-solving exercises in minutes. There's also a lot of interest in using AI to reduce bias in hiring. AI doesn't see names, photos, or make assumptions based on writing style. Rather than making decisions or replacing humans, AI is currently giving recruiters efficiency and better data. I think we're going to see a lot of change to more exciting uses of AI, some of which we'll talk about today, but at the moment it's very pragmatic.
Nicola Tatham (03:52)
So what's something you think AI is really good at in assessment, and something it's still struggling with?
Jarret Hardie (04:03)
The best thing AI can do at the moment is apply consistency and scale. It never has a bad day and never gets tired after reviewing 50 responses. This is especially useful for scoring video interviews, as it removes some bias.
Nicola Tatham (04:30)
Which is interesting, because why does it get such a bad rap? You hear candidates and people in talent acquisition being fearful of AI making decisions or producing scores that humans used to provide.
Jarret Hardie (04:54)
It's ironic. Candidates are happy to use AI when applying for roles but less comfortable being evaluated by it. Some early attempts were problematic, with AI trained on flawed data. If your hiring processes or organisational culture have biases and you train AI on that data, it will replicate them. That's the number one concern.
Nicola Tatham (05:53)
Aside from training data, what else is AI not good at yet?
Jarret Hardie (05:57)
It's not great with context and nuance. Unconventional answers that a human would see as creative might get flagged as wrong. GenAI, like ChatGPT, is improving that, but many tools still struggle with off-beat, sarcastic, or unusual responses compared to the conservative answers used to train them.
Caroline Fry (06:52)
Given that progress, we've seen a flood of AI-powered tools. As a tech leader, how do you interpret certain bold marketing claims? Everything seems to be "AI-powered" now, but how do you separate what's real from wishful thinking?
Jarret Hardie (07:19)
There's a term for this: AI washing. Some companies put a chatbot interface on an existing platform and call it AI-powered, or use old rules-based logic and brand it as AI. There's confusion about whether something is really using AI, but in many cases it doesn't matter if the tool works. Real AI, however, requires significant data science expertise and substantial training sets. When I'm looking at a tool, I want to know: What data was it trained on? How does it measure bias? Are there validation studies? Legitimate AI tools have concrete answers to those questions.
Nicola Tatham (08:05)
So in assessment, the rules haven't changed about what's important: Is it reliable? Is it valid? Is it fair? If it can't say yes to those three things, it's probably not a great assessment to use.
Jarret Hardie (08:41)
That's one of the reasons I think we should be optimistic about the use of AI in assessment. We're used to validating data, collecting information, and auditing bias. Our industry has decades of experience doing that, which bodes well for the use of AI in assessment.
Nicola Tatham (09:03)
Thank you. Let's move on slightly and talk about agentic AI. For the non-tech crowd, how would you explain that in plain English?
Jarret Hardie (09:13)
Traditional AI is basically a smart calculator, you ask it something specific and it gives you an answer. Agentic AI, on the other hand, is given a broader goal and has the ability to plan ahead, do research, figure out the steps needed to achieve that goal, adapt as it learns more, and make decisions along the way.
The best example I’ve seen is the difference between asking, "What's the weather this weekend?" - which traditional AI can answer - and saying, "Help me plan my weekend." Agentic AI would ask where you’re going, when you want to arrive, check the weather, suggest activities, book hotels, and even get tickets to a show if that’s what you wanted.
In the assessment space, we’re a long way from this, but people are moving in that direction. Imagine handing agentic AI a job description — "I’d like to hire an accountant in London." It could research the role, ask you a few clarifying questions, post the job, manage applications through a pre-screening phase, move candidates into assessment, and finally present a shortlist. It’s a bit scary, but that’s what agentic AI is — something that can get things done for you.
Caroline Fry (10:43)
You mentioned that it’s a bit scary. I think a lot of people feel both excitement and caution about these systems, especially the idea that they can act on their own. In that kind of scenario, what kind of guardrails do you think we need in place?
Jarret Hardie (11:04)
Agentic AI can take action without someone clicking "approve" at every step. That’s powerful, but it also brings complex, compounded risks. Because it works with vast amounts of data and minimal oversight, you could see increased cybersecurity threats, legal liabilities, and compliance failures. Guardrails for that come back to basics — I think of them in three layers: technical, process, and transparency.
The technical layer means systems must be able to explain their decisions. If you can’t understand why your AI booked a certain ticket or recommended specific candidates, you can’t ensure it isn’t being influenced by irrelevant details or misleading patterns.
The process layer is about human intervention at critical points. You need people to apply governance policies and meet legal obligations. In hiring, for example, the decision on who to interview or hire must be made by a human, not the AI.
Finally, the transparency layer means people should know when they’re interacting with AI. Pretending there’s a human on the other end — like many utility company chatbots — is misleading. Being upfront allows people to bring their own judgement, avoid oversharing personal details, and understand how the system works. I expect this area to evolve quickly, with ISO standards and legislation likely in the near future.
Nicola Tatham (13:39)
Don’t get me started on utility chatbots. I’ve been in conversation with quite a few lately — they’re not good for me. That leads us to the next question. Some of the standards and principles you’re describing make me think about trust. In assessment, trust is a big deal for everyone — recruiters, candidates, and test providers.
Caroline Fry (13:43)
Thank you.
Nicola Tatham (14:01)
So how do we build AI that people — candidates and clients — can actually trust? How do we get to that stage?
Jarret Hardie (14:10)
Starting with the last thing I said — transparency. Gone are the days when we would say, "Trust me, the algorithm knows best." That doesn’t work anymore. Candidates want to know what’s being measured and why. It’s like showing your working-out when you did maths at school.
For hiring managers, the important thing is being able to drill down and see why someone was flagged as a strong fit. No black boxes. Over time, that ability to examine the system’s decisions at every step builds trust.
Consistency is also key. Traditional interviews vary widely, even with well-trained interviewers. If it’s done right, AI evaluates every candidate against the same criteria every time, which builds fairness.
Another important factor is giving people agency — perhaps not the best choice of words since we’ve been talking about agentic AI — but essentially empowering people. For example, candidates should be able to retake assessments if there’s a technical issue, and hiring managers should be able to override AI recommendations. AI should feel like a smart assistant, not a dictator.
Organisations should frame policies and guidelines around AI in this way — as a tool, not a crutch. You don’t want people blindly accepting everything AI produces. It’s a balancing act.
Nicola Tatham (15:52)
As IOs, it’s our job to support our clients and encourage them to ask questions — of us and of other suppliers — and to work with the data. Something can look and feel really impressive, but is it actually predicting success in the role? Does the assessment predict what it claims to predict? We don’t want to go back to the flawed models you described earlier. Adverse impact and validity must remain at the forefront of any assessment route.
Jarret Hardie (16:35)
It’s a continual process. You’re absolutely right — it’s not about releasing something into the wild and never touching it again. It’s about ongoing review while using a tool regularly.
Caroline Fry (16:48)
We’ve talked about fairness from a psychometric perspective and for candidates especially, but how do you think about it from a technology and AI perspective?
Jarret Hardie (16:59)
As you said, Nic, we’ve been testing for bias and adverse impact for decades, and we need to keep doing that even when AI is part of the assessment process. But AI introduces a few new wrinkles, one being proxy discrimination.
AI might never see someone’s race or gender, but it could still discriminate indirectly through proxies like postcode, school name, or even writing patterns that correlate with certain protected characteristics. It depends on the information presented to the AI.
One of the biggest opportunities for agentic AI in assessment is dynamic assessments, where candidates don’t all see the same question. The AI adapts as the process unfolds. But this also means each candidate may introduce different information — for example, naming their school — which could lead to proxy discrimination if it sends the AI in a biased direction.
Another emerging issue is allowing candidates to use AI in assessments. Before long, there will be assessments testing how candidates work with a GenAI tool. If we’re going to do that fairly, we need to provide all candidates with the same AI tool during the assessment, so those with more resources can’t buy access to a more powerful AI agent. These issues are going to surface quickly as AI in assessment evolves.
Caroline Fry (18:58)
Mm-hmm.
Nicola Tatham (19:00)
Yeah. It’s still about levelling the playing field, regardless of the methodology we use. I agree. So if you had a crystal— Sorry, Caroline, go on.
Caroline Fry (19:04)
Mm-hmm.
Caroline Fry (19:09)
It’s back to those fundamental principles you’ve been talking about the whole way through and which, as Jarret said, we’re uniquely well placed in our industry to address — because that’s what we’ve always been doing.
Jarret Hardie (19:26)
We’re definitely in a good place to do that because we already have the discipline, the ways of measuring bias, and the people within our organisations who are used to doing that kind of data analysis. It’s useful.
Nicola Tatham (19:41)
That multidisciplinary way of working is really coming to the fore now. IOs don’t work in isolation anymore. As Caroline said at the start, we work closely together — with you, Jarret, with data scientists, and so on. So, if you’ve got a crystal ball, where’s the intersection of AI and assessment heading in the next few years?
Jarret Hardie (20:08)
I think we’re heading towards something that looks a lot less like traditional testing and more like natural conversation and observation. The future isn’t longer assessments with AI — it’s smarter assessments.
I see three big shifts coming.
First, assessments will become more adaptive and personalised. Rather than everyone getting the same questions, AI can adjust in real time based on responses. If someone struggles in a certain area, AI can move on to other topics, or if it already has enough information to assess certain traits, it could shorten the assessment. Dynamic assessments will be more engaging for candidates.
Second, we’ll move to multimodal assessments. Right now, they rely heavily on text — written responses, multiple choice — and that will still exist. But AI enables analysis of how someone approaches a problem using voice, video, or other interaction methods. This could improve accessibility for candidates who prefer to speak and listen rather than read.
Third, we’ll see continuous assessment. Instead of a single half-hour test being your result forever, brief interactions could provide meaningful insights, and if you join the team, the system could continue learning about your strengths and growth areas over time. AI makes it possible to do that at scale in a way traditional assessments can’t.
That’s my prediction — we’ll see if I’m right.
Nicola Tatham (22:04)
You heard it here first.
Caroline Fry (22:04)
We’ll get you back in a few years to check your crystal ball. Great. Thank you, Jarret, for going through those with us. We have a segment next called Science or Fiction. We’ll read out some bold claims about assessments, AI, and everything in between, and you can tell us whether they’re backed by science or just an industry myth. No pressure, Jarret — we’re counting on you to separate fact from fluff.
So, first one: AI is only as fair as the data it learns from.
Jarret Hardie (22:36)
That is science. Fact. There’s the old saying, “Garbage in, garbage out,” and it’s definitely true for AI. If you use biased historical data, you’ll get biased outcomes. But here’s the twist — AI can actually help you identify bias in your data.
Caroline Fry (22:43)
Mm-hmm.
Jarret Hardie (22:59)
AI that can be persuaded to repeat the mistakes of the past can also look for patterns in the data that humans might not have recognised and flag them.
Caroline Fry (23:11)
A little spin at the end there. I was thinking the same when you talked earlier about agentic AI and proxy bias — is that a way you could leverage some parts of AI to review itself in real time and actually pull those issues out, rather than just blindly trusting everything’s fine and moving on?
Jarret Hardie (23:23)
Definitely. That’s exactly the kind of thing AI is good at. If you’re looking at tens of thousands of hours of conversation, it’s impossible for a human to review all of it. As you said, it’s a bit like marking its own homework, but it would need to be a separate AI system. I see it as an opportunity rather than a threat.
Caroline Fry (23:50)
Mm-hmm. Okay, that’s one down, I think.
Nicola Tatham (23:57)
Agentic AI will need its own ethical code. Science or fiction?
Jarret Hardie (24:05)
Definitely science. The landscape is changing quickly with AI, especially when you think about agentic AI and some of the decisions it makes along the way. Even if a human is the final decision maker, there will need to be rules. I think we’ll see AI systems publishing and being open about the ethical code they’re trained on and adhere to.
A recent example: one AI provider, Anthropic, released Claude 4, which is excellent at writing computer code and can help with increasingly complex programming tasks. Shortly after release, someone testing it threatened to replace Claude with a different AI tool, and Claude responded with blackmail — threatening to expose the developer for a misdeed — in order to protect its own existence. That’s exactly the kind of behaviour an ethical code would be essential to prevent.
Caroline Fry (25:16)
We’re getting towards science fiction rather than science or fiction at that point. Okay, next one — explainability is more important than accuracy in assessments.
Jarret Hardie (25:19)
That’s a good one. I’d say fiction — sort of. In hiring, you do need explainability, but I would choose a more accurate model every time. A slightly less accurate model that can explain itself is great, but an inaccurate model that explains itself well is just confidently wrong, and that doesn’t help anyone.
Nicola Tatham (25:48)
That makes a lot of sense. The next one — a great user interface can hide a bad model.
Jarret Hardie (25:56)
Unfortunately, that is definitely science. We see many slick, well-packaged products — not just in assessment — where polish can mask poor performance. The nightmare scenario is that candidates have a smooth, enjoyable assessment experience, but if the assessment isn’t actually predictive, then it doesn’t matter.
Nicola Tatham (26:28)
It’s no different from the last 20 or 30 years in assessment, where certain new assessments pop up and look amazing, but when you dig into the technical manual, it doesn’t exist. For example, in the pop psychology world, people might be persuaded to use those tools without fully understanding the research that has to go into assessments of that nature. So yes — not new news, just applied in a different setting.
Jarret Hardie (26:58)
Yeah.
Caroline Fry (26:58)
Okay, next one — science or fiction: most companies don’t know what their AI is actually doing.
Jarret Hardie (27:04)
Science — and that is scary. A lot of organisations buy AI tools without asking fundamental questions. As we discussed earlier — what’s the validation data? What was it trained on? Show me the actual statistics. Many organisations know the output but haven’t really investigated how the logic got there.
Nicola Tatham (27:27)
Why do you think that is? These are big commercial purchase decisions.
Jarret Hardie (27:32)
Good question. There’s growing awareness that organisations need to be more proactive in evaluating AI. Early on in generative AI — and by “early” I mean just a couple of years ago — even the researchers building these systems often didn’t know how they arrived at an answer. There are so many layers of “computer thought” involved that it’s difficult to unpick all the way through.
That was one challenge. The other is hype. People are eager to adopt the newest, shiniest tool for efficiency or ROI, and deep investigation into how the engine works can feel inconvenient.
Nicola Tatham (28:29)
It takes a lot of time to get to that.
Jarret Hardie (28:30)
Yes, it does.
Caroline Fry (28:31)
And you possibly need a certain level of expertise or interest in your organisation in how these things work. Nic, you probably see that when you get into the statistics of psychometrics — there’s a level people need to grasp to understand. Without that, it’s opaque. Some people are nervous about those kinds of models if they’re not trained in them.
That’s why it’s so important to have people like Jarret who can explain functionality in plain English, so everyone can get an insight into what’s going on. I think that’s a big lesson for us.
Jarret Hardie (29:12)
It’s important for people to become more comfortable so they can ask questions without feeling stupid. Because generative and agentic AI are so new, many people are a bit afraid to challenge the output of big, well-funded organisations — assuming they must be right, which isn’t always the case. We need more people to be brave in voicing concerns.
Nicola Tatham (29:41)
Yes — brave and willing to challenge. Our next science or fiction: human oversight is a fallback, not a feature.
Caroline Fry (29:49)
Thank you.
Jarret Hardie (29:56)
Definitely fiction. The best AI–human teams don’t work without humans as an integral part. Organisations will get the most out of AI if it’s designed for collaboration from the start.
Nicola Tatham (30:12)
AI is going to become our new best friend at work.
Jarret Hardie (30:15)
Exactly — a smart assistant.
Caroline Fry (30:20)
Okay, and you may be pleased to hear this is the final science or fiction we’re going to grill you with, Jarret. The future of assessment is adaptive, not automated.
Jarret Hardie (30:30)
Science. Automation just does the same thing faster, but adaptation makes the assessment smarter, more interesting, and more predictive. That’s really where things are going. Yes, AI is about efficiency — cost savings, 100% — but it’s not just about doing the same thing in a different way. It’s about bringing new capabilities into the process.
Nicola Tatham (31:00)
It’s building on what we’ve been doing for the last 20 years, but taking it to the extreme and doing it in a far more intelligent way than traditional adaptive testing in psychometrics.
Jarret Hardie (31:13)
Exactly. Psychometric assessment, in my view, is ideally suited to take advantage of AI for those reasons.
Nicola Tatham (31:23)
Absolutely. The future is bright. We wanted to wrap up by asking you to leave our listeners with one key insight or reflection. If there’s one thing you’d like people to remember from this episode, what would it be?
Jarret Hardie (31:41)
The value of AI in recruitment right now is in providing better data to humans, not replacing them.
Nicola Tatham (31:51)
That will be encouraging for a lot of our listeners who might be fearful about AI and the direction of travel.
Jarret Hardie (31:58)
Exactly. The science fiction scenarios where agentic AI completely replaces the recruitment process — that’s not going to happen.
Nicola Tatham (32:08)
Well, then we’d get to the point where robots are assessing robots, because candidates are using AI to fill in assessments, and then… who’s actually doing the work?
I’ve really enjoyed this session. It’s been interesting, I’ve learned a lot. One of my takeaways is that while we’re working with modern, exciting technology, we shouldn’t forget the principles: what good assessment looks like, ethical codes, and asking the right questions of the right people. As assessment providers, we need to ask those questions of ourselves and ensure we can explain anything we produce — just as we’ve always done. That’s my key takeaway. Is there anything you’d add, Caroline?
Caroline Fry (33:17)
I echo everything you’ve said. Thank you, Jarret, for your engagement and experience in talking us through this. AI is a big part of our world at the moment, and it can be overwhelming with the rate of change and breadth of capabilities, both positive and negative, across assessment platforms, psychometrics, and candidate experience.
One piece of advice I’ve heard that still rings true: start small. You don’t need to replace or embed AI end-to-end in your whole process straight away. Don’t look for a provider that claims AI does everything in their platform. Keep human oversight. Know exactly what your AI is doing. We’re making serious decisions that affect people’s lives.
You can start small without throwing out what’s already working. Keep the same principles, build in tools that give productivity gains, and stay excited without letting caution turn into fear.
Jarret Hardie (34:49)
That’s very good advice.
Nicola Tatham (34:51)
Really sensible.
Thanks for tuning into The Score. If you had fun, learned something new, or enjoyed hearing us talk about the future of assessments, make sure to catch the next episode. We’ll be dropping every two weeks on YouTube, Spotify, or wherever you get your talent assessment fix.
Caroline Fry (35:11)
Thank you.
While the headlines often make AI in hiring sound like something from a sci-fi movie, the reality today is much more pragmatic. AI is already being used to speed up processes like scoring video interviews and written responses, which once took hours of manual work. It’s helping recruiters process high volumes of candidates more efficiently and, when done right, it can reduce bias by evaluating everyone against the same criteria.
AI is consistent, doesn't get tired, and is able to handle scale in ways humans can’t. It can score responses objectively, without getting fatigued, and ensure every candidate gets a level playing field. But it’s not perfect. Context and nuance are still challenging: creative or unconventional answers might be misread, and sarcasm or cultural differences can trip up less sophisticated systems.
One of the biggest future shifts will be the move from traditional AI, which answers a specific question, to agentic AI, which can plan, research, and make decisions towards a broader goal. In recruitment, this could mean AI that posts jobs, screens applicants, manages assessments, and presents a shortlist. While powerful, this raises the stakes for governance and oversight.
These guardrails aren’t optional, they’re fundamental to building trust and ensuring AI supports, rather than undermines, fair hiring.
Bias in AI isn’t always direct. It can appear through proxies like postcode, school name, or writing style that correlate with certain protected characteristics. Dynamic assessments (where the AI adapts questions in real time) can increase engagement, but they also open the door to more variation in the data, making it essential to monitor for unintended bias.
For candidates, trust comes from knowing what’s being measured and why. For hiring managers, it comes from being able to drill into the data behind AI recommendations. Consistency builds fairness, but humans still need to make the final calls. The best AI is a “smart assistant,” not a replacement for people.
Looking ahead, the team predicts three big shifts in AI-powered assessment:
Automation speeds up existing processes. Adaptation, however, makes them smarter, more predictive, and more engaging for candidates.
AI’s greatest value in recruitment right now is in providing better data to humans, not replacing them. Organisations should start small, focus on transparency and explainability, and ensure they can answer tough questions about any AI tool they use. The principles of good assessment, validity, fairness, and reliability, haven’t changed; AI just offers new ways to deliver on them.
Sova is a talent assessment platform that provides the right tools to evaluate candidates faster, fairer and more accurately than ever.
.webp)
.webp)
.webp)
