By Lee Flanagan
✨ AI Summary:
- Label AI screening tools accurately as filtering, not interviewing—misrepresenting their function sets unrealistic candidate expectations and obscures their actual role in the hiring pipeline.
- Vendor claims of “bias-free” AI require published methodology and independent verification; unsubstantiated assertions about fairness can mask inherited historical biases in training data.
- Contradictory feedback from chat bots reflects their design purpose (high-volume pattern-matching) not flaws—treating bot outputs with the same evidentiary weight as human-led conversations moves bias risk earlier, not forward.
- Reserve hiring decisions for structured human conversations after screening; automated tools legitimately narrow large applicant pools, but the judgment to hire must come from evidence-based interviewing, not algorithmic recommendation alone.
When Sydney Morning Herald reporter Millie Muroi applied for a “holiday casual” role at Kmart, the retailer’s Sapia.ai screening tool told her two contradictory things in the same feedback email. “Putting the needs of the group ahead of your own personal gain seems to come naturally to you,” it read, before warning her to “be careful to not dominate group conversations and projects.” She was praised for selflessness and cautioned against dominance in the same five-question chat exchange.
The Feedback That Contradicted Itself
In our view, that is not a flaw in the technology. It is the predictable output of a tool built to sort applicants fast enough for a national retailer’s hiring calendar, not to interview them. Call that an interview and you set an expectation the tool cannot meet. Call it screening, automated and fast, and the contradictory feedback stops looking like a glitch. It looks like exactly what volume filtering produces: pattern-matched, directionally useful, no substitute for judgment about an actual person.
More Than 600,000 Applicants a Year Is a Filtering Problem
Kmart’s tool does not run one interview. It runs triage across an applicant pool that tops 600,000 people a year. Sapia.ai founder and chief executive Barb Hyman says the platform has assessed more than 10 million people since launching in Melbourne in 2018 and now serves 35 per cent of ASX100 companies, largely consumer brands. Woolworths, Australia’s biggest employer, uses Sapia.ai for the early stages of its recruitment process, as do Bunnings and Qantas.
Running triage across 600,000 applicants a year is sorting at industrial scale, a legitimate thing to build software for, but it is not interviewing. Yet the email Kmart sends applicants insists “there are no right or wrong answers.” A recommendation employers follow nine times out of ten disagrees. Kmart declined to comment. The company’s chief people and corporate affairs officer, Tristram Gray, previously told The Australian that applicants rated the experience highly because it was “no longer impersonal.” That gap between description and function is where the naming problem starts.
The ‘Bias-Free’ Claim Nobody Outside Sapia.ai Can Test
Hyman says Sapia.ai reduces unconscious bias and tests frequently for it across race, age and disability. That is a vendor’s claim about its own product, made without any published methodology, and deserves to be treated as exactly that. Hyman also says about 90 per cent of candidates the tool recommends to a firm are ultimately hired, a hiring rate separate from how often employers accept the recommendation. The Australian HR Institute’s Sarah McCann-Bartlett offers the more useful caution. “There’s also a lot of research on bias and discrimination in hiring,” she says. “If an AI system is trained on historical data that reflects past biases or discrimination, it may replicate that.”
Whether Kmart’s tool clears that bar is not something an applicant, a journalist or a talent acquisition leader outside Sapia.ai can verify. That unverifiable gap, not any single failure, is the real story here. Experts and politicians quoted in the reporting are already warning that AI hiring tools risk entrenching bias and locking out unorthodox candidates. “Bias-free” is not a claim a vendor gets to make about itself without showing the work.
The Diversity Number Cuts Both Ways
Here is the strongest evidence against treating this purely as a naming problem. Kmart reports that First Nations people now account for 8.25 per cent of its hires, well above the 3.2 per cent share Aboriginal and Torres Strait Islander people held of the population at the 2021 census. Asked about that increase, Hyman said the tool likely made the process more comfortable. “Given the disadvantage they face, they might just feel more comfortable with this way of hiring,” she said, adding that candidates “might assume they’ll be discriminated against in a face-to-face interview, but not in a blind interview.”
That is Hyman’s speculation about her own product, not a tested finding. It is a fair argument that screening can be more accessible than a face-to-face gatekeeper round. It is not an argument that a five-question chat bot has assessed who someone actually is well enough to stand in for a hiring decision. A tool can widen the funnel at the top and still be the wrong instrument for the judgment that happens further down it. Kmart’s diversity number only supports the first claim.
Where the Line Actually Belongs
In our view, hiring teams running screening tools like this at scale rarely have a fairness problem at the screening stage itself. They have a labelling problem. Candidates are told they are being interviewed when they are being filtered. The people running the process internally start treating a chat bot’s personality summary as if it carries the same evidential weight as a structured, human-led conversation. It does not, and we think that gap is what turns a volume filtering tool into the bias risk already flagged here.
The fix is not to abandon automated screening for high-volume roles. Sorting 600,000 applicants by hand is not realistic for any retailer. The fix is to be explicit, with candidates and hiring managers, about which stage this actually is. Screening narrows the pool. Interviewing is the human, evidence-based conversation that decides who gets the job, regardless of how capable the sorting software upstream becomes. Ask yourself which one your process is actually running before you call it either. The moment a business lets a chat bot’s output stand in for that conversation, it has not modernised its hiring. In our view, it has moved the bias problem one step earlier and called it progress.
Original reporting: SMH.com.au.
Frequently asked questions
Is Kmart’s Sapia.ai chat tool actually an interview or a screening step?
It functions as high-volume screening, five scripted questions that triage part of Kmart’s 600,000-plus annual applicants. Employers adopt the tool’s recommendation about 90 per cent of the time, and Sapia.ai founder Barb Hyman separately says roughly 90 per cent of recommended candidates are ultimately hired. That is a filtering function, not the human, evidence-based conversation that should decide who gets the job.
Does Sapia.ai’s ‘bias-free’ claim hold up to scrutiny?
No published methodology accompanies Sapia.ai’s bias claims in the reporting, so it stands as the vendor’s own assertion about its own product. The Australian HR Institute’s Sarah McCann-Bartlett notes that AI trained on historical hiring data may replicate past bias, which is the open question this claim leaves unanswered.
Why did Kmart’s First Nations hiring rate rise after adopting AI screening?
Kmart reports First Nations representation of 8.25 per cent among hires, above the 3.2 per cent population share recorded at the 2021 census. Sapia.ai’s Barb Hyman says the tool likely made the process more comfortable given the disadvantage First Nations candidates face, and that they might assume they would be discriminated against in a face-to-face interview but not a blind one. That explanation is her own speculation rather than a tested finding.
Should high-volume employers drop AI chat screening tools altogether?
No. Filtering hundreds of thousands of applicants by hand is not realistic for a retailer the size of Kmart, and automated screening has a legitimate role narrowing that pool. The tool needs to be labelled accurately as screening, with real interviewing judgment reserved for the decision that determines who is actually hired.