Every conversation about AI risk focuses on what could go wrong. Almost none of it asks who benefits when it does. That is the question that actually tells you how the technology will develop.

In my technology and AI course, students work through five analytical lenses applied to real technologies and real consequences. The lens that produces the most discomfort, and the most genuinely useful analysis, is the one I call incentives. The incentive lens asks a simple question about any technology, any policy, or any system: who benefits when this goes wrong? Not who is harmed, which is the easier question and the one most ethics frameworks start with. Who benefits.

When an AI hiring tool produces biased results that systematically disadvantage certain candidates, the company using the tool has plausible deniability. The tool made the decision, not the hiring manager. The liability is diffused. The discrimination continues. Who benefits from that ambiguity? Not the candidates. The benefit flows to the organisation that can now automate a decision it was already making and call it objective. This is not cynicism. It is analysis.

The Governance Lag Problem

Research on technology governance consistently documents governance lag: the gap between the deployment of a technology and the development of frameworks capable of managing its consequences. AI is experiencing one of the most pronounced governance lags in the history of consumer technology. In that gap, the incentive question becomes the most practically useful analytical tool available. Regulation has not arrived. Institutional norms are still forming. In the absence of external accountability, the behaviour of any system is determined by the incentives operating on the people who build and deploy it.

The global mood is split down the middle and both halves are growing. The Stanford AI Index 2026 reports 59% of people now say AI products bring more benefits than drawbacks, up from 52% in 2022, while 52% say those products make them nervous, up from 39% (Chapter 9, Figure 9.1.1). Hollow competence is what happens when the first number rises faster than the skill behind it.

A Worked Example: The Screening Tool Nobody Questioned

Minseo is a 34 year old HR manager at a retail chain headquartered in Seoul. Last year her company bought an AI screening tool to sort the 3,000 or so applications it receives in each round of store manager hiring. The vendor promised faster shortlists and less bias. Six months in, a regional director asked her why so few applicants from outside the capital region were reaching interview. Nobody in the company could answer, because nobody had asked the vendor what the model had been trained on.

Run the incentive lens on that and the picture is uncomfortable. The vendor benefits from speed, because speed is what it sells. Minseo’s company benefits from a shortlist that looks objective. The candidates who were filtered out bear the cost, and they never find out why. Nobody with the power to change the tool was also harmed by it, which is exactly the condition under which, as the questions below predict, a tool keeps running.

The Three Questions

In the course I teach, students apply three questions to every AI system they analyse. Who commissioned this tool and what are they optimising for? The gap between the stated purpose and the actual optimisation target is usually where the most significant risks live. Who bears the cost when this tool produces a wrong output? If the answer is a different party from the one that benefits from right outputs, the incentive to improve is weak. Who has the power to change this tool when it is producing harmful results? If the answer is nobody who is also harmed by it, the tool will continue operating regardless of the harm it produces.

None of the right column is exotic. It is what any careful buyer would ask for, and in Korea some of it is now backed by law. The point of the table is that the left column had been true for six months before anyone wrote it down.

If you are in Minseo’s seat, the order matters. Write the left column first, honestly, even if it is embarrassing. Then take the right column into the next vendor meeting as a list, not a conversation, and ask for each item in writing. A vendor that cannot answer the third row is telling you who really controls the tool.

What Korean Law Now Says About AI in Hiring

Two Korean laws now speak to Minseo’s situation. Since 15 March 2024, the amended Personal Information Protection Act has given people rights over automated decisions. The Personal Information Protection Commission explains that where an automated decision affects a person’s rights or obligations, they can ask for an explanation or a review, and where it significantly restricts those rights, they can refuse it. The organisation then has to stop applying the decision or have a person process it again.

The AI Basic Act, in force since 22 January 2026, goes further. A 정책브리핑 explainer lists hiring (채용) alongside healthcare, energy and loan screening as fields where AI can seriously affect people’s lives or rights, and says providers must build a system in which people manage and control the AI. Neither law answers the incentive question for you. Both give you something to point at when you ask it, and as the AI governance gap argues, policy still tends to arrive after the tools.

Why This Matters for Professional Communication

A 2025 report from the OECD on AI in the workplace found that workers in organisations with clearer AI accountability structures reported significantly higher confidence in the tools they used and lower rates of AI related errors reaching consequential stages. Accountability structure is not a technical problem. It is a communication and governance problem. It requires people who can ask the incentive question clearly and insist on an answer. That is a communication skill. It is also, in 2026, a professional responsibility.

The second email is not hostile. It is specific, and specificity is what makes accountability possible. Candidates care about this too. In a Pew Research Center survey of 11,004 U.S. adults in December 2022, Americans opposed AI making final hiring decisions by 71% to 7%, and 66% said they would not want to apply for a job with an employer that uses AI to help make hiring decisions. Asking the vendor these questions protects your candidates and your reputation at the same time.

Before the next tool lands on your desk, try The Offload Line to separate what you can hand to software from what has to stay with a person, and The Long Game to test how your own values hold under pressure. If your work touches hiring, how ATS filters decide before recruiters do shows the same incentive problem at an earlier stage, the AI conversation nobody is having at work covers accountability inside your own team, and the escalation ladder shows how one unexamined shortcut grows.

→ This is one of the topics I speak on for corporate and academic audiences: how to think critically about AI tools using the governance and incentive frameworks from my technology and AI course. The Work With Us page has more on speaking engagements and workshop formats.

---

Update, September 2026: Korea now has an answer to the accountability question, at least on paper. The AI Basic Act, in force since 22 January 2026, classifies AI used for employment decisions as high impact and requires transparency and human oversight, with administrative fines phased in after a one year grace period (U.S. International Trade Administration, Cooley).