How breaking Hiring with AI might finally fix it
The hiring market is locked in a mutually assured misery and resentment loop.
For candidates, finding a job has become a dispiriting endurance test. Openings are scarce—the UK Office for National Statistics (ONS) August 2026 report indicates that overall national vacancies fell to 707,000, marking a five-year low—, hiring processes drag across five or more stages, and ghosting has become a standard at all of them. Applicants have zero visibility into how they’re evaluated, leaving them feeling commodified and disempowered.
For recruiters, the problem is the signal-to-noise ratio. Open job ads are flooded within hours by thousands of applications, many of them automated and irrelevant. Finding qualified talent in a sea of synthetic CVs is a near impossible task, while AI-assisted cheating has turned take-home tests and live interviews into a mockery of candidate evaluation.
Instead of alleviating pain in this situation, AI has mostly brought more chaos. Recruitment is degenerating into an escalating battle of automation:
Recruiters, overwhelmed by the volume of applications, implement AI CV screening, leading candidates to use AI agents to blast hundreds of tailored applications to all the ads they can find, pushing recruiters to add bespoke screening questions to introduce friction, making candidates use LLM wrappers to auto-generate answers, forcing recruiters to insert hidden adversarial prompts in job descriptions, angering candidates who retaliate with white-text prompt in their CVs to get AI screening to move their application forward, resulting in recruiters overwhelmed by the volume of applications…
This cycle produces friction for both sides and does not generate much signal. Hiring managers delegate the hard work of assessment to software, candidates delegate their authenticity to bots, and genuine talent is lost in the middle.
Does AI have a place in hiring?
AI brings a lot of promises for hiring: streamline the process, save time and effort for talent managers and hiring managers, reduce the cost of hiring, remove location and time constraints (24/7 running pipeline), improve the decision making process and its fairness.
Sadly, it’s also introducing severe systemic vulnerabilities.
Despite AI’s potential to make decisions more objective by removing human judgement and biases, it’s actually creating its own fairness issues:
- Historical bias encoding: AI systems are trained on past hiring data, which embeds historical disparities directly in their weights and leads them to make decisions based on systemic patterns of discrimination, e.g. Meta’s and LinkedIn’s algorithmic ad delivery targeting STEM high-paying jobs at men and lower-wage roles to women and minority candidates, Amazon abandoning its AI recruitment tool after audits revealed it systematically downgraded CVs containing the word “women’s” (e.g. “women’s rugby team”).
- Proxy discrimination: Simply stripping protected attributes (gender, race, age…) from applications fails because models learn to identify proxy signals. Algorithms learn to optimise for indicators such as uninterrupted career paths , specific zip codes or the name of private education institutions, penalising caregivers, non-traditional backgrounds, and minority communities.
- Asynchronous video assessments (HireVue, Pymetrics): Critics and demands for strict auditing have been aimed at AI interviewing tools. Their analysis of facial micro-expressions, vocal cadence or behaviour metrics disproportionately penalise non-native English speakers and neurodiverse candidates whose communication styles fall outside of the model’s training distribution.
AI in hiring also brings concerns about legal compliance. The ingestion of sensitive candidate profiles into third party AI APIs, without explicit consent, transparent logic, or a clear right to challenge the outcome, creates major regulatory exposure under frameworks like the EU General Data Protection Regulation (GDPR) and the EU AI Act, which formally categorises recruitment and worker evaluation systems as “high risk”.
Yet existing legal protections remain fundamentally mismatched with how algorithms actually discriminate. Traditional employment law was written to catch direct discrimination, i.e. the explicit, intentional exclusion of protected classes. It has almost no mechanism to police statistical proxy bias or the automated reinforcement of historical disparities.
Weirdly, privacy regulations—like GDPR which mandates data minimisation, lawful processing, and limits on automated profiling—have become our best defence against algorithmic bias.
Back to the question…
Where, if anywhere, does AI belong in Hiring?
Companies need to draw a strict boundary between what AI can help with and what humans must own.
High-risk use of AI => restrict or eliminate:
- Automated CV scoring / ranking / filtering: LLMs can’t read between the lines, they might pay too much attention to the wrong signals, and they’re vulnerable to buzzword targeting. There’s a high risk to dismiss great candidates and let through mediocre ones.
- Asynchronous AI video analysis: Evaluating facial micro-expressions or vocal pitch is pseudo-scientific woo-woo that destroys candidate trust and discriminates against diverse profiles.
- Algorithmic sourcing and ad targeting: Delegating ad distribution to self-optimising platforms silently replicates historical bias
- Any kind of automated scoring or recommendation based on application questions, CV, interview notes, candidate data… AI models do not provide sufficient consistency in decision-making nor transparency. They build their own proxy goals that might not align at all with recruitment criteria and might quietly discriminate against minorities and exclude great candidates.
High value use of AI:
- Supervised factual extraction: parsing employment timelines, verifiable certifications, and quoted achievements into structured recruiter-facing summaries, without generating any advice or ranking. The challenge here is to create focus in the extraction (e.g. ask for specific signals) without imparting bias (e.g. causing the LLM to make judgement about the relevance of signals).
- Protection against bots: detecting automated application spam and candidate LLM cheating to increase signal-to-noise ratio. The former is easier than the latter, although some progress is being made on identifying LLM-generated text. This should be implemented with caution to avoid excluding candidates due to false positives.
- Reducing administrative overhead: One of the biggest and unnecessary pain in hiring is all the admin around it, like scheduling interviews with busy senior hiring managers or multi-party assessment panels. That’s where AI can help the most: scheduling, sending transparent pipeline updates to stakeholders, automating non-critical comms to avoid ghosting, etc…
Rethinking human assessment
Hiring is a high-stake strategic bet on a company’s future competitiveness. It’s not an administrative chore to be automated away. When done right, it balances four goals: securing operational capability, building long-term organisational agility, nourishing the business’ brand with candidates for future hiring needs, and providing fair and equal opportunity. The latter is not only an altruistic goal, it’s also a selfish one: you want to get more unique talent playing for your side and building your moat.
Technology cannot compensate for flawed evaluation design. If organisations want to get out from under a mountain of irrelevant applications and improve their chances to find the best fit for a role, they’ve got no choice but to invest in improving their evaluation processes.
Goals and decision-making process first
A structured hiring process starts with clear decision mechanics before anything else. What concrete competency signal is being evaluated at each specific touchpoint? What criteria defines an objective "pass" versus a rejection? Who needs to be involved in the decision? How are scores weighted, calibrated, and documented across different interviewers?
Without this baseline clarity, adding interview stages does not increase predictive confidence. If talent teams and hiring managers do not know what signals they are measuring, a 7-stage process provides no more rigor than throwing a stack of CVs down a stairwell and hiring whichever lands on the top step.
The hiring process establishes the relationship with a future employee. If an organisation demands custom assessments, take-home work, multi-person panel presentations, in-person culture fit checks, and more, it owes applicants transparent timelines and evaluation criteria, as well as actionable feedback at all stages of the process past CV screening.
Potential-based hiring
If I was hiring for a role today, I’d prioritise learning potential and velocity over past experience box-checking without hesitation. Hiring for a 100% experience match admittedly gives assurance that candidates have been able to do the job in some form before, but it could also select for people who are more comfortable in reusing the same past playbooks than trail-blazing in an ambiguous environment. Evaluating for core reasoning, curiosity and learning aptitude is a far superior bet in some cases to find adaptable operators who can be quickly “molded” to specific business needs and evolve with a company’s ever-changing challenges.
Whenever my teams or I have made bets on less experienced but more motivated and resourceful candidates, the return on investment did not disappoint.
Fixing the probation safety net
No interview process provides absolute certainty. In the end, you only truly know how an employee performs once they’re on the job.
In theory, the probation period exists to catch these misfires. In practice, it has degenerated into a bureaucratic rubber stamp. Companies routinely let underperforming hires roll into permanent contracts simply because nobody really paid attention during probation or made the effort to do a rigorous evaluation or wanted to have an uncomfortable conversation.
However, treating probation as an extended audition process is an expensive and ethically flawed trap. Deciding to let a new hire go at the end of their probation can inflict substantial damage: financial sunk cost (recruitment fees, onboarding overhead, probation salary), senior contributors time waste (interviewing, mentoring, pairing, reviewing early deliverables…), lost hiring momentum.
There are ways that probation could be turned into a real safety net to mitigate false positives in hiring:
- Making it an active and documented confirmation process rather than a passive “nothing to flag” roll-over
- Replacing the vague 3 or 6 month review with a structured 30 (tooling and domain onboarding), 60 (independent contribution) and 90 day (fully up to speed) timeline.
- Maintaining an active “silver-medalist” pipeline, treating strong runner-up candidates with radical transparency and regular communication. If an early placement misfires within the first month, the next best candidate might still be available (and no one will begrudge you for choosing them second, this is not dating!)
In any case, probation is neither an extra round of interviewing nor an excuse for bad upfront evaluation.
The problem with networking
As the public application processes collapse under the weight of AI spam, industry commentary often falls back on an alternative: networking.
While referrals undeniably fill vacancies, treating networking as a systemic solution is an abdication of responsibility. As a job seeker, relying on backchannels is an inherited privilege. It favours candidates with preexisting social capital, geographic proximity and cultural alignment with people in positions to pull them into their organisations. It leaves many deserving qualified people stranded. If a company claims to care about meritocracy and finding the best fit for a role, its hiring pipeline must be functional, transparent and fair, not driven by connections behind closed doors.
High-signal human evaluation
What does good assessment look like in practice in the age of AI?
Take-home tests were a good way to assess candidates’ thinking process and their ability to do the job, but AI is diminishing their reliability. How do you know they haven’t used AI to complete the task? What if they have, does it matter? They’ll have access to AI in their job, after all… What are we assessing anyway?
A higher-signal alternative is live problem solving. Ten years ago, this was the gold standard: give a candidate a scenario and some time to think about it on their own, then work through it in front of a panel or in a pairing session. These tests didn’t check for polished and perfect answers but for critical thinking. Does the candidate ask questions? Do they flag assumptions? Explore options? Etc…
Yes, this requires availability from senior team members. But isn’t it worth the investment if it saves you the effort to review other noisy stages and allows you to make a better hiring decision?
In the same spirit, I would argue that stages can be condensed into 3 rounds: screening, hiring manager interview exploring past experience and motivation, technical/practical interview (+culture fit, also for the benefit of the candidate). The rise of remote hiring has made it too easy to tack on “just one more 45-min Zoom call”, destroying momentum and motivation. I would consolidate the final evaluation into a same-day stage, and if geography allows, conduct it in person. It forces the hiring team to calibrate the first stage to progress strong candidates to the second stage, and makes it easier to come to a definitive decision rather than dragging the process out for weeks.
—
AI didn’t exactly break Hiring, but it came at the perfect time to make it worse for everyone. But that’s not inescapable! Companies have two levers to make the experience and outcomes better: shifting AI support to automation-safe steps and rethinking how candidates are assessed, with clear decision mechanisms, effective tests, and a tight and transparent process.
References
Jansen P, et al. Fairness, AI & recruitment. Computer Law & Security Review. 2024; 53:105953.
Al-Shayea Q, et al. Bias in AI-driven HRM systems: Investigating discrimination risks embedded in AI recruitment tools and HR analytics. Results in Engineering. 2025; 25:103811.
Tilmes N. Disability, fairness, and algorithmic bias in AI recruitment. Ethics and Information Technology. 2022; 24(2):21.
Dastin J. Amazon scrapped 'sexist AI' tool. BBC News. 2018 Oct 10.
Lufkin B. AI hiring tools may be filtering out the best job applicants. BBC Worklife. 2024 Feb 14.
Future of Life Institute / Artificial Intelligence Act. What the Act means for staffing businesses. EU AI Act Explorer. 2024.
European Parliament, Council of the European Union. Article 6: Classification rules for high-risk AI systems. EU AI Act. 2024.
BBC News. Job vacancies at five-year low as smaller firms scale back recruitment. BBC News. 2024.
Office for National Statistics. Vacancies and jobs in the UK: August 2026. ONS Statistical Bulletin. 2026 Aug.