How to screen when every resume reads perfectly
The problem is not that applications got worse. It is that they all got equally good, which is the same thing as them carrying no information.
A recruiter reading applications in 2026 is doing a task that quietly stopped working. The document in front of them used to be weak evidence of ability, produced at some cost by the candidate. It is now produced at no cost, to a uniform standard, by everybody. The distribution of resume quality collapsed, and the thing that collapsed was not quality. It was variance, and variance was the entire information content.
This matters because most screening advice still assumes the old distribution. Read closely. Look for specifics. Watch for vague verbs. All of that was good advice when a specific, well-structured application implied a candidate who understood the work. It implies nothing now, because specificity is generated.
Reading harder cannot work, and it is worth being precise about why
If two candidates produce identical documents, no amount of attention applied to those documents will separate them. The information is not hidden in there, waiting for a better reader. It is absent. Every hour spent on closer reading is spent recovering something that is not present, and the people who are best at gaming the format are the ones it rewards.
The same logic disposes of detection. Suppose you could perfectly identify which applications were machine-assisted. You would have sorted candidates by tool use, which correlates with neither ability nor honesty, and you would have systematically penalised applicants writing in a second language. You would have answered a question nobody asked.
The rule: ask for something that cannot be prepared in advance
A screen carries information when passing it requires something the candidate produces under conditions you control. That gives three usable families, in rough order of cost to you.
- A fresh scenario question, generated per candidate, answered in a bounded time. Nothing to search for, nothing worth passing around afterwards, and the scoring can be blunt because you are not grading prose.
- A short work sample drawn from the actual job. Twenty to thirty minutes, obviously relevant, and reviewed by whoever will manage the person.
- A verifiable artifact plus one specific question about it. A repository, a deployed thing, a document. The artifact is the claim and the question is the check, and the question is the part most processes skip.
What these share is that the cost of faking them scales with the work itself. That is the property you are buying.
Put it before the expensive hours, not after
Most funnels already verify. They do it in references, at the end, after four people have spent an hour each. By then the stack has been ordered by document quality and the expensive hours have been allocated on the strength of the one signal that no longer carries information.
Invert it. A cheap evidence step in front of the loop reorders the stack before anyone senior reads anything, which is the whole point. The stage order matters more than the stage quality here: a blunt check in the right place beats a sophisticated one in the wrong place.
- Application: collect, do not rank. Treat the document as contact details and consent.
- Evidence, automated: a fresh question or a short sample. This is the stage that replaces resume reading, and it is the stage most funnels do not have.
- Human screen: thirty minutes, on what the evidence stage surfaced. You are now interviewing a shortlist that was ordered by something real.
- Loop: your existing process, on a much smaller and better-sorted set.
What this costs, honestly
It costs you a conversion step. Any additional stage loses candidates, and some of the ones you lose will be good. The trade you are making is a smaller top of funnel against a stack that is ordered by ability rather than by formatting, and that trade is only worth it if the evidence stage is genuinely short. Under about twenty minutes it reads as a reasonable ask. Over an hour, unpaid, it reads as free labour, and the candidates with options decline first, which inverts the filter you were trying to build.
It also costs you the comfort of a familiar artifact. Hiring managers like resumes because they are skimmable and because they have always been there. Replacing that with a score and a transcript is a change in how decisions get defended internally, and that is a real cost that is usually underestimated.
The market context, from our own data
One number from our corpus is worth having in mind while you design this. We hold 9,188 live postings, and between 24% to 31% of them state what the job pays. If your screen asks candidates for real work, and your advert will not tell them what the job pays, you are asking for evidence while declining to provide any. Candidates notice the asymmetry, and the strongest ones are the most willing to act on it.
Where we stand in this
We build one of the things described above, so treat the rest of this paragraph as interested. insiderOne generates fresh scenario questions per candidate, grades them for reasoning rather than vocabulary, and signs the result so it can be checked by an employer who was not in the room. The useful part of this post is the shape of the funnel, and that shape holds whether you assemble it from our tools, somebody else's, or a shared document and a calendar invite.
The thing not to do is keep the old funnel and read harder. That is the one option that is certain not to work, because the information it depends on is no longer in the document.
Frequently asked
How do you screen candidates when AI writes every application?
By changing what you ask for rather than how hard you read. A screen works when a strong candidate can pass it and a weak one cannot, which means it has to depend on something the applicant produces under conditions you set: a fresh question they could not have prepared, a short work sample tied to the actual job, or a verifiable artifact. Anything the candidate can prepare in advance and at no cost is now free for everyone to produce.
Are AI detectors a good way to filter applications?
No, for two reasons that are both worse than the false positives. They punish the wrong thing, since a careful non-native speaker who used a tool to fix their grammar is not a fraud, and they are an arms race you do not control. Detecting the tool tells you nothing about whether the person can do the work, which is the only question you actually have.
Where should verification sit in a hiring funnel?
Before your expensive human hours, not after. Most funnels verify at the end, during references, which means the costly interview loop has already been spent on a stack ordered by document quality. Moving a short evidence step to the front reorders the stack before anyone senior reads anything.
Does asking for a work sample reduce the applicant pool?
Yes, and that is most of the value, as long as the sample is short and paid attention to. A task that takes twenty minutes and is obviously connected to the job filters for interest as well as ability. A four hour unpaid project filters for desperation and free time, and the strongest candidates are the first to decline it.