Should AI reject an application before a human reads it?

It depends on three things, and the case that raised the question this week fails at least one of them. Times Higher Education reported on 25 September 2026 that in a UKRI-funded cybersecurity call, machines removed about half of 179 proposals before any human read them, and that the applicants were told only that AI “may” be used. A rejection nobody read is defensible when three conditions hold: a named person can see what was rejected and why, and answers for it; the rejected can find out and contest it; and the material being screened still carries information about the thing being judged. The first two are governance and can be built. The third is the one a new preprint on hiring says is failing without anyone measuring it, because when applicants write with the same tools the screener reads with, the documents stop telling anyone apart, and a harder filter finds less rather than more.

What happened in the UKRI call, as reported#

Frances Jones’s report in Times Higher Education concerns the first phase of CRANE, a pilot funded through the Engineering and Physical Sciences Research Council and hosted at the University of Oxford. The call drew 179 proposals. AI systems scored them against seven criteria, among them relevance, quality, novelty, feasibility and value for money, using several models and then a second pass with different models “to guard against any one model’s idiosyncrasies”, and about half were removed before a human reviewer saw them. John Stott, a senior lecturer in astrophysics at Lancaster University, told the paper: “It’s depressing that I wrote something that no one ever read before it was rejected.” He added that if calls were to be assessed this way, “I think I’d avoid applying.” CRANE’s stated policy, as quoted: “We may use AI tools in initial assessment of proposals, but reviewers will be accountable within CRANE for their reviews.” UKRI said it was “exploring the safe use of AI-assisted assessment” with “appropriate controls and safeguards”, including trials that compare the machine’s conclusions with past decisions. The call document, the paper reported, had told applicants that AI “may” be used in initial assessment.

Condition one: someone answers for the rejection#

CRANE’s policy makes reviewers accountable for their reviews. The ninety or so proposals cut in the first pass had no review in that sense, so the sentence does not reach them. That is the first test for any organisation screening people out by machine, in grants, hiring, admissions or lending: is there a named person who can see the list of the rejected, the reason each was rejected, and the pattern across them, and who would be answerable if the pattern turned out to be wrong. The law on hiring already asks a version of this. Annex III of the EU AI Act places systems used to “analyse and filter job applications” among high-risk uses, with the deployer’s oversight duties in Article 26, and Article 22 of the UK GDPR gives a person the right not to be subject to a decision based solely on automated processing where it has a legal or similarly significant effect on them. What Article 22 does not do, as human at the start sets out, is accept a human who glances at a machine’s list and signs it; the involvement has to come from someone with the authority and competence to change the decision. Grant proposals from institutions sit at the edge of that rule, which is written for decisions about individuals. Whether a rejected researcher is such an individual has not been tested.

Condition two: the rejected can find out and contest it#

Stott knew his proposal had not been read because the process said so. In most hiring the applicant never learns whether a person or a program rejected them, and a screen that cannot be seen cannot be contested. The research’s position on whether AI can be unbiased is that a machine screener removes some human variance and encodes one pattern at a scale no individual reaches, and that the difference that decides the outcome is that the machine can be audited and a human screener cannot. That advantage exists only if someone audits it. UKRI’s proposed safeguard, comparing the machine’s conclusions against past decisions, tests agreement with previous human panels. It would catch a system that diverges from past reviewers. It would not catch a system that reproduces their habits, including the tendency of any scoring process to reward the familiar over the new, which for a call that lists novelty among its criteria is the failure most worth looking for. A contestable process needs a route by which an applicant can ask what the machine saw, and a person able to overrule it, the capability examined on who can override an AI system.

Condition three: the document still tells you something#

This is the condition the week’s other document addresses. Itai Ashlagi, Ramesh Johari, Jon Kleinberg and Anushka Murthy posted a model of hiring markets on 24 September 2026. Their argument is that AI tools “make it easier to find and apply to jobs” and at the same time “reduce how informative those materials are about applicant fit”. When every application is fluent and tailored, the ones that once stood out no longer do, and firms fall back on the one signal the tools cannot manufacture, prior experience. The candidates hurt most are the inexperienced and well matched, who “lose the individualized information that could distinguish them from other inexperienced candidates”. The bottleneck, in their phrase, moves “from submitting applications to obtaining credible evaluation”. It is a theoretical model rather than a measurement, and it says nothing directly about grant proposals. But the mechanism transfers. A machine that screens documents written by machines is grading the tool, and the harder it screens the more it selects for whoever used the tool best. The authors’ predicted market response is an extra stage, an intermediate assessment that produces new evidence before the costly full review, rather than a sharper filter on the old evidence. That is a different design from the one CRANE ran, and the entry-level labour market is moving towards it for the reasons set out on will AI replace entry-level jobs.

What a screener is for, and what it does to the people who read#

A funder or an employer that cannot read everything has always screened. The question is what is lost when the screen becomes a reject rather than a sort. A sort puts the strongest proposals in front of the people who decide and lets them look further down the pile when they choose; the humans keep the practice of reading the whole field, and with it the ability to notice the odd one that the criteria missed. A reject removes the bottom half from view, and the people who decide never see what they are no longer able to judge. Over several rounds that is a change in what the reviewers can do, of the kind the research calls capability debt: the skill of spotting an unusual proposal atrophies because the unusual proposals stop arriving. For hiring, the parallel is the recruiter who no longer knows what an unpolished but able application looks like; that is the reason hiring for judgement puts weight on evidence the candidate produces live rather than in a document.

The decision an organisation controls#

On this site’s argument the measurable risk sits in the handover of a decision to a machine and in what the humans around it can still do afterwards, and the decision the organisation controls is which decisions the machine may make, who can stop each one, what people must remain able to do, and how anyone would know it went wrong. Applied to screening: let a machine sort, and decide in writing whether it may also reject. If it may, name the person who owns the rejected list and reads a sample of it every round. Tell applicants which stage is automated, in the call document and in the rejection. Keep a route to a human who can reverse the decision and the authority for them to do so. Test the screen against outcomes the machine did not produce, not only against last year’s panel. And where the documents being screened are themselves machine-written, stop treating the document as the evidence and add a stage that produces some. Those are decisions to take before the call opens, the point Rules Before Tools makes about every system that acts in an organisation’s name.

What this does not show#

The account of the CRANE call is one newspaper report and the statements it quotes; this page has not seen the call document, the scoring, or the rejected proposals, and cannot say whether the machine’s half was the weaker half. Nothing here shows that the rejected proposals would have been funded, or that human panels would have chosen differently; the history of peer review is not a record of consistency. Ashlagi and colleagues built a model, and a model shows what follows from its assumptions rather than what happened in any market. The legal position on automated grant decisions is untested. The capability argument in the fifth section is the research’s reasoning applied to a new case, not a measurement of any reviewer. What the two documents together do show is that a machine rejection nobody reads raises a governance question about accountability and a harder question about whether the thing being read still means anything, and that the first is being answered while the second is not yet being asked.