Latest
Artificial intelligence is transforming how companies hire. 87% of organizations are reportedly using some form of AI-powered recruitment technology. But recent headlines remind us that not all AI is created equal. In fact, if AI isn’t designed carefully, it can replicate and even worsen the very biases it's supposed to eliminate.
That tension is at the heart of recent lawsuits filed against Workday, a leading applicant tracking system. Plaintiffs allege that Workday’s AI-driven screening tools discriminate on the basis of age, gender, and disability. While the lawsuits are still ongoing and the claims remain unproven in court, they raise a difficult question: are the tools we use helping us make fairer decisions—or just scaling bias behind a digital curtain?
What happens when AI mirrors the status quo?
AI systems are only as good as the data they’re trained on and the design choices made during development. Many commercial hiring tools rely on machine learning models trained on historical hiring data—data that often reflects the inequities of past decisions. If women were historically passed over for engineering roles, or older candidates were screened out before interviews, then an AI trained on those outcomes will learn to do the same. Amazon made headlines nearly a decade ago when it had to phase out an AI powered hiring tool that it trained on the company’s internal hiring data after it was found to be prejudicial towards women.
In other words: bias in, bias out.
Worse, some tools fall prey to what's known as proxy bias. Proxy bias occurs when algorithms use non-demographic signals—like names of schools, previous employers, or even zip codes—as stand-ins for race, gender, or socioeconomic status. These seemingly neutral details can introduce discrimination into the process, especially if the model was never designed to detect and counteract such patterns. As a result a resume parsing app might never see a candidate’s age or gender but can still learn to devalue resumes that don't “look like” past hires or historical hiring data.
What ethical AI should look like
For AI to be part of the solution, not just compound existing problems, it must be built with fairness and transparency from the start. At CLARA, that principle is foundational.
Our system doesn’t just tack on bias mitigation as an afterthought. It’s designed around it.
Curated, bias-reduced training data. The models that power CLARA were trained on data specifically structured to mitigate bias and prioritize equitable assessments of skill and fit—not pedigree or pattern-matching.
Built-in de-identification tools. When users opt in, CLARA anonymizes key fields—such as names, schools, and employers—before any model sees the application. That means the system evaluates each candidate on what they can do, not where they come from.
Models that focus on capability, not conformity. Instead of looking for the most "typical" candidate, CLARA’s neuro-symbolic AI is designed to evaluate a broader range of signals and surface talent that may otherwise go overlooked. In other words, these models are built to filter talent in, not out.
Moving beyond blind optimism
As AI adoption accelerates, there’s growing urgency to ensure that the tools we rely on are ethical, accountable, and truly aligned with our ethical hiring practices. It’s not enough to assume that automation will improve fairness just because it feels more objective than human decision-making. Left unchecked, AI can quietly entrench old biases in new ways, and at scale.
That’s why due diligence matters. Leaders need to ask hard questions of their vendors:
What kind of data was this AI trained on?
What safeguards are in place to prevent discrimination?
Is this tool surfacing new types of candidates—or simply reinforcing the status quo?