자동화된 채용 심사에 관한 세미나

Professor Siddiqui: Last week we looked at how automated tools sort job applications. This week I'd like your reactions to the case study. Marco, you've read it. Was the company wrong to use the system?

Marco: Not necessarily wrong, though I'd say careless. The company claimed the software was objective because it had no opinions. But it had learned from ten years of past hiring decisions, and if those decisions favoured a certain kind of candidate, the system simply reproduced the pattern, only faster.

Lin: I'd push back a little. Human recruiters are hardly neutral. One study in the reading found that people reviewing applications spent an average of seven seconds on each one, which suggests plenty of snap judgements. At least a program can be audited.

Professor Siddiqui: An important distinction. Can it be audited in practice?

Lin: That depends on whether anyone is allowed to look inside. In this case the supplier refused, saying the method was a trade secret, so the company couldn't explain its own rejections.

Marco: Which seems to me the real scandal. Had the company been able to explain a decision, a rejected applicant could have challenged it. As it was, people were turned away by a process nobody could describe.

Professor Siddiqui: So the objection is not to the technology as such, but to its opacity.

Lin: Mostly, yes, though I'd add that efficiency is being mistaken for fairness. Ten thousand applications in a week is a cost problem, and the tool solves the cost problem. Whether it's fair is a separate matter that nobody was obliged to answer.

Marco: I'm not sure I'd go quite that far. Some firms do test their tools carefully. The difficulty is that they test for the biases they already know about.

Professor Siddiqui: Which brings us to your assignment. I'd like each of you to draft a one-page policy for a small organisation that wants to use such a tool. I don't want a ban, and I don't want blind enthusiasm. What safeguards would you require? Think about who should be able to see the results, how often the system should be reviewed, and what an applicant is entitled to ask.

Lin: Can the policy include a human decision at the end?

Professor Siddiqui: Certainly, but be specific about what that person actually does. A human who merely clicks approve on a machine's recommendation is not much of a safeguard.