The screening system that filters out good people, and nobody finds out

HR usually measures an automated screening system by time saved and applications not read.

Nobody measures the number of well-suited candidates filtered out for writing a CV in a shape the system was not trained on.

That cost is perfectly invisible, because the people wrongly rejected never learn why and never come back to say so. The organisation sees the time saved and never sees the opportunity lost.

Where the bias comes from

A screening model learns from past hiring data, which means it learns who the organisation has hired — not who performed well.

If most past hires came from the same handful of universities, the model scores candidates from those universities higher, without anyone writing that rule.

Harder still: removing the university from the data does not fix it. Correlated signals remain — writing style, vocabulary, internships at a particular set of firms — and a model will find its way back to the same thing if the data allows it.

And employment gaps, which commonly come from parental leave, caring responsibilities or illness, get read as a negative signal automatically unless somebody designs against it.

What to give the system

Administrative work requiring no judgement. Extracting details from CVs, scheduling interviews, sending status emails, flagging applications sitting too long. This consumes a great deal of HR time and carries no bias risk.

Genuinely objective requirements. A professional licence the role legally requires, or a clearly specified language level, can be checked automatically and fairly.

Transparent ordering. If you rank, you must be able to say what the ranking is based on, and the recruiter must always be able to see every application rather than only the top ten.

A reply to every applicant. Not a technology question at all, but systems make it practical — and it is what candidates remember about your organisation longest.

What not to give it

Automatic rejection on a model's score.

Video interview analysis inferring personality from expression and tone — a technology whose scientific basis is far weaker than vendors present, and which discriminates directly against disabled candidates.

Culture-fit scoring, which in practice measures similarity to the people already there.

The audit to run

If you use screening, check periodically whether the proportion of candidates passing, broken down by gender, age and educational background, differs materially from the proportion applying.

This is not hard to do and should be done by someone who does not own the system. Without it, a model drifts somewhere nobody intended, and nobody learns until it is late.

Applicant data is personal data under the law. Retention periods and the right to erasure need to be settled at the start.

What GIPSIC does here

We build internal enterprise systems and systems handling personal data under Thailand's data protection requirements.

On work of this kind we usually propose a narrower scope than the client arrived with: make the process faster and more transparent for both sides, without letting the system decide on anyone's behalf. The trade-off for automating this particular decision is not worth the time it saves.

If you would like to talk about your hiring process, get in touch.