If you’ve experimented with tools like ChatGPT, Claude, or Gemini, you’ve probably experienced this: ask the same question twice and you’ll often get two different answers.
This is by design. Gen AI models are probabilistic. They generate answers by predicting the “next most likely word,” with an intentional dose of randomness (temperature, sampling) to make them feel more “human.”
When you use that design principle in recruitment, you’re playing with fire.
Variability = Risk
Imagine using generic AI for:
- Screening CVs for “culture fit”
- Generating interview questions
- Evaluating candidate responses
- Writing job adverts
If the same input produces different outputs depending on the day, you have a trust problem:
- Inconsistent screening = discrimination claims waiting to happen
- Variable interview guides = unfair candidate experiences
- Drifting evaluation criteria = missed talent
- Mixed messaging = damaged employer brand
Hiring decisions are high-stakes. Candidates deserve certainty and fairness. Employers need defensibility. Probabilistic creativity is great for drafting emails or brainstorming headlines. It does not belong where the output affects someone’s career.
Thin wrappers, big risks
What we’re now seeing in the market is a proliferation of “thin wrappers”. Hiring tools that are built quickly on top of open-source AI models. The logic is simple: take a model like Qwen, Mistral, or LLaMA, put a UI around it, and call it a recruitment solution.
The problem? These wrappers inherit all the instability of their foundation models. And worse, they add risk x10:
- Data governance: Who owns the candidate data once it touches that model? How is it being stored, trained, and reused?
- Bias & fairness: If outputs vary day-to-day, bias testing becomes impossible. What are you even testing against?
- Regulatory exposure: Under the EU AI Act, voice or biometric data misused in training is a compliance nightmare. Under GDPR Article 9, it constitutes a breach of sensitive data handling.
- Enterprise readiness: A startup wrapping an open model cannot match the scale, validation, or auditability that global enterprises require.
This is the hidden risk of generic AI in the hiring process. On the surface, it looks sleek, fast, and innovative. Underneath, it’s a house built on sand.
The alternative: A specialised AI system built for measurement, not inference
At Sapia.ai, we’ve taken a very different path. We’ve built a for-purpose AI system designed specifically for hiring, utilising methods published in peer-reviewed journals.
Over the last eight years, we’ve conducted more than 8 million structured, conversational interviews across 50 countries and 20 languages. Every response is scored against validated competencies, ensuring that our assessments are:
- Structured, not random: Every candidate gets the same questions. Every answer is scored against the same criteria.
- Transparent: You can see how scoring works. Every candidate gets personalised feedback.
- Governed: Our FAIR™ framework, ISO 42001 alignment, and regular bias testing mean enterprises can adopt AI responsibly.
- Loved by candidates: 91% completion rates, 9/10 experience ratings.
This isn’t a thin wrapper. It’s an AI system designed from the ground up for hiring, with fairness, science, and trust at its core.
Why should HR leaders care?
The convergence of Talent Acquisition, Talent Management, and Reskilling means the pressure on HR leaders has never been higher. Everyone wants internal mobility, but the default playbook (job boards, CV self-mapping) rarely delivers.
If the tools you adopt today are built on randomness and inference, you’re not just risking a poor candidate experience. You’re risking lawsuits, compliance failures, and reputational damage.
If instead, you invest in measurement, structure, and science, you create a workforce data asset that compounds in value, unlocking hiring intelligence, mobility pathways, and skills development at scale.
Variability is a liability
Generative AI has transformed how we create at pace and at scale. But let’s not confuse creativity with science. Recruitment isn’t about “good enough, most of the time.” It’s about fairness, rigour, and trust.
For those who want to understand more, check out our ebook Understanding Responsible AI in Recruitment.
Frequently Asked Questions (FAQ)
- Why is generic AI risky to use in hiring?
Generic AI models like ChatGPT or Gemini are probabilistic. They generate different answers to the same input. In hiring, that variability creates risk: inconsistent candidate screening, unfair interview questions, and drifting evaluation criteria. For employers, this opens the door to discrimination claims, compliance failures, and reputational damage. - What are “thin wrapper” recruitment tools, and why are they dangerous?
Many new hiring tools are “thin wrappers” built on top of open-source models like LLaMA or Mistral. They inherit instability from their foundational models and introduce additional enterprise risks, including unclear data ownership, a lack of fairness testing, and poor regulatory compliance. In contrast, purpose-built systems like Sapia.ai are validated, explainable, and designed for defensibility. - How does Sapia.ai ensure fairness in AI hiring?
Sapia.ai adheres to the FAIR™ Framework, a global standard that ensures AI-powered hiring is unbiased, explainable, valid, and inclusive. Every candidate receives the same structured questions, and responses are scored against validated criteria. Models undergo continuous bias testing. Independent research shows Sapia.ai reduces the gender gap in hiring by up to 36%. - Does purpose-built AI really help increase diversity in hiring?
Yes. A Monash/Gothenburg University study using Sapia.ai found that when candidates knew AI would assess their application, 30% more women applied, without reducing quality or volume. When paired with AI scoring, evaluators selected men and women equally, closing the gender gap by 36%. - How does AI improve the candidate experience?
Unlike CV screening or psychometric tests, Sapia.ai’s Chat Interview is untimed, text-based, and mobile-first. Candidates can respond in their own time and receive personalised insights, not just a rejection. Across 8M+ interviews, the average candidate satisfaction score is 9.05/10, with over 80% voluntarily leaving feedback. - Is Sapia.ai inclusive for people with disabilities and neurodiverse candidates?
Yes. Traditional hiring (video interviews, timed assessments) can disadvantage people with disabilities. Sapia.ai’s chat format removes those barriers: it’s untimed, compatible with screen readers, and designed for psychological safety. Research shows candidates with disabilities progress through the funnel at the same rate as others, ensuring hiring equity of up to 98%. - How does Sapia.ai protect candidate data?
All data is processed securely through AWS Bedrock, ensuring nothing is shared with or retained by model providers. Sapia.ai does not use demographic data (e.g., gender, age, race) for scoring, and complies with GDPR, ISO 27001, and ISO 42001 standards. Learn more about our security and compliance in our Trust Centre. - How is Sapia.ai different from generic generative AI tools?
Unlike generic GenAI, which is built for creativity, Sapia.ai is built for measurement. Every candidate is assessed against the same validated competencies, creating structure, transparency, and defensibility. This makes it enterprise-ready, compliance-aligned, and trusted by brands like Woolworths, Qantas, BT and Holland & Barrett
About Author
