If a company uses AI to help recruit, who should a rejected applicant ask for an explanation?
My concern is not to return every decision to manual work. As recommendations, scoring, and execution become easier to delegate to systems, have we made authority and responsibility equally clear?
The original essay was written on July 12, 2025, using 2027 as a point for reflection. This revision retains that outlook without making a particular year a deadline for AI to take over every decision, or treating an imagined scenario as a survey of adoption that has already occurred.
TL;DR: When AI participates in decisions, first distinguish organizing information, recommending, scoring, and authorized execution. A person’s ability to give reasons does not establish their faithfulness; AI explainability also varies. Bias can arise from data, objectives, and use processes, without necessarily being amplified every time. A hypothetical hiring process illustrates four checks: reasons, differences in impact, authority, and handling objections. These are my governance suggestions, not a prediction about 2027, a legal checklist, or a guarantee of fairness.
1. Specify which step AI participates in
Suppose a company uses AI to organize résumés and rank applicants, then lets recruiting staff choose whom to interview. That differs from a system sending rejection notices without anyone checking for errors.
The same label, “AI recruitment,” can describe very different arrangements of authority. This is an illustrative scenario, not an account of a named company or a study of outcomes.
Likewise, the presence of AI tools in banking, medicine, or justice does not establish that AI alone decides loans, treatments, and judgments. Discussing an actual system requires checking its purpose, authorization, professional procedures, and applicable rules. This essay has not performed those individual evaluations.
I would first ask: who performs each step, which steps can be stopped, and who is affected? Map the process before deciding what to inspect.
2. Giving reasons does not establish that the reasons are reliable
A recruiter’s explanation of a ranking does not necessarily reflect the actual judgment process faithfully. Conversely, AI explainability varies by method, purpose, and audience; it cannot all be described as completely inexplicable.
NIST’s 2021 explainable-AI document distinguishes supplying an explanation, its meaningfulness to the recipient, and its accuracy. It also discusses the limits of people’s explanations of their own decisions.[1]
I would therefore not settle for a plausible paragraph. In the hiring example, I would check which information was used, whether ranking criteria can be verified, and whether the explanation matches the actual process.
That does not require every applicant to understand every model parameter. It means helping affected people understand relevant reasons and know what remains unclear. The appropriate level still depends on context.
3. Fairness requires more than an absence of discriminatory intent
AI expressing no discriminatory intent does not make an outcome fair. Nor does a person’s sense of being fair constitute evaluation evidence.
NIST’s 2023 AI risk framework considers broader systemic, computational/statistical, and human-cognitive sources of bias.[2] This leads me to ask not only about training data, but also sampling, labeling, scoring objectives, and how people use a ranking.
For example, a hypothetical résumé ranking that treats lifestyle clues unrelated to the job as indicators of ability deserves inspection. This illustrates a risk; it is not a finding that some café behavior predicts ability.
A risk does not mean bias is necessarily amplified every time. We need to examine differences within a specified use and population, whether those differences have a reasonable basis, and whether corrections improve matters. Without hiring data, this essay cannot declare any system fair or unfair.
4. Locate power concerns in arrangements we can change
I remain concerned that organizations can lose the ability to inspect and change processes when they depend on a few suppliers, do not understand data flows, cannot retrieve records, or cannot switch tools. This is a conditional governance concern, not a law that only giants can develop or deploy AI.
Training a model, adapting an existing one, and adopting a ready-made service are different activities. Their resource requirements cannot be collapsed into a single threshold. This essay estimates neither supplier market shares, training costs, nor market concentration, and does not infer that four particular companies therefore know everything about everyone.
In the hiring example, I would inspect contracts and implementation: who can read résumés, who keeps records, whether purposes can be restricted, how data is handled after discontinuation, and who can correct ranking errors.
Those questions need concrete answers. Neither “AI is undemocratic” nor “technology has no remaining challenges” substitutes for examining a system and its governance.
5. Human participation need not depend on denying tool capabilities
AI can compare options and may help propose objectives or identify problems. Human responsibility need not rest on claiming that tools can never perform a certain kind of thinking.
I care about who is authorized to adopt an objective, who bears the consequences, and whether affected people can object. A model score alone cannot explain those arrangements.
The OECD AI Principles discuss accountability and traceability according to roles and context, and information that enables adversely affected people to challenge outputs.[3] These provide governance directions, not a completed analysis of local law or a requirement that one person manually recalculate every result.
Human participation can include teams setting boundaries together, keeping records, and establishing stopping mechanisms for serious errors. Not everyone needs to become a model expert, and using AI does not remove anyone’s claim to respect.
6. Start with four checks rather than waiting for 2027
Before using AI to assist organizational decisions, I would ask four questions:
- Can the reasons be verified, and does the explanation match the actual process?
- Have impacts on different people been examined, and which information is still missing?
- Who can authorize, stop, or correct the process, and who is responsible for its consequences?
- How can affected people object, and who responds and records the handling of those objections?
These are practical questions I have organized, not four universal legal obligations or a guarantee that asking them produces fairness and safety.
The year 2027 can encourage a longer view, but it should not become a story that lets us skip present checks. Rather than predicting when AI will decide everything for everyone, I want organizations to explain now what they use, what they authorize, and how they take responsibility when things go wrong.
Sources and reading limits (checked through November 4, 2026)
- NIST, Four Principles of Explainable Artificial Intelligence, September 2021, §2 principles and §8 limits of human explanations. Used to distinguish purposes and accuracy of explanations, not as test results for this example.
- NIST, AI Risk Management Framework 1.0, January 2023, §3.7 bias and relevant GOVERN/MAP passages. A risk framework, not a fairness evaluation of a particular hiring system or new legal obligations created by this essay.
- OECD, official AI Principles page, transparency, oversight, and accountability passages in the principles revised in 2024. Used at the level of principles, without inferring local obligations or compliance of an individual case.
Revision note: This version retains 2027 as the author’s outlook and withdraws assertions that AI necessarily enters every decision, people are inherently transparent, bias is invariably amplified, or four giants know everything. Nor does it turn the original into a summary of the AI2027 report. All examples are hypothetical; no cases, statistics, or outcomes have been added.


💬 Comments
Loading...