Screening
Three Things on an Application You Can Check, and One You Never Will
Three claims on an application have records behind them: employment history with titles and dates, education, and licensure or credentials. Check those on everyone who reaches an offer, through the issuer or the former employer, rather than on the candidates who feel off. Authorship has no record and no instrument, so budget nothing to working out who wrote the document. The claim you care about most, that the person can do the work, is not a verification question at all.
The takeBackground-screening vendors are selling the wrong fear. The pitch is fabricated identities and generated candidates, priced as detection, and it arrives while the dull checkable claims sit unchecked at the bottom of the offer process. A team paying for authorship analysis before it runs employment verification on every hire has bought the interesting problem and left the tractable one open. The unglamorous version, a registry lookup and two phone calls, settles more than any detector on the market can.
Where Olive fits
Open a role and see what the work shows
Records answer the first three questions and no document answers the fourth, which is where Olive sits: an employer opens a role, the candidate works through a 40-to-60-minute occupational assignment with an AI assistant, and a human reviewer writes six findings, each anchored to a timestamped moment in the session. The candidate is granted the same report, free, on every tier.
Rank your shortlistWhich Claims Have a Record Behind Them?
Three of them: employment history with titles and dates, education, and licensure or credentials. Each has an issuer, a registry line or a payroll record standing behind it, and each can be settled in a lookup or a phone call. Those are also the claims a falsification would have to touch to change your decision, because they are the ones the offer rests on.
How each one is actually settled:
- Employment and dates. The former employer's HR line, or the payroll-verification service they use. Expect title, dates and eligibility for rehire and nothing else, which is enough for the claim being checked.
- Licensure. The issuing board's public register, searched by number. Ask for the number on the application rather than a yes-or-no field, so the check takes seconds instead of an email thread.
- Education. The registrar or the national clearing house the institution reports to. Slower and worth it only where the credential is a genuine requirement of the role.
Two cautions travel with all three. A discrepancy is a question rather than a verdict: titles get inflated by the company that awarded them, dates move when a contract converts, and a name changes for ordinary reasons. And in the United States, where a third-party screening firm performs the check, the federal Fair Credit Reporting Act of 1970 attaches: 15 U.S.C. 1681b(b) requires a clear and conspicuous written disclosure in a document consisting solely of that disclosure, the applicant's written authorisation before the report is obtained, and a copy of the report before adverse action 1. Other jurisdictions have their own rules, so have counsel classify the arrangement before the first call.
What Settles the Claim You Actually Care About?
Nothing on the application does. Whether this person can do the work is not a claim with a record behind it, so it is not a verification problem at all. It is an evidence problem, and the thing that settles it is watching the work happen under conditions close to the job. A job-shaped exercise is one instrument for that, and it holds up about as well as the strongest interview evidence.
The familiar number for work samples is stale. The often-quoted .54 traces to a 1974 review and its own lineage abandoned it; the current estimate is .33, and the paper making that correction argues it sits alongside structured interviews rather than beneath them, since the credibility intervals overlap heavily 5. Most of the underlying studies also tested people already doing the job, and none of them tested unpaid multi-hour take-homes. Read it as a reason to include a job-shaped step.
This is where the specifically modern claim lands too. A resume line reading AI-proficient has no issuer, no registry and no consequence attached, which puts it in the same class as a skill slider: verifying an AI-proficiency claim without taking someone's word for it means creating the evidence yourself. Same for a portfolio piece with no visible author, where working out who actually built the project is answered by a conversation about the decisions inside it.
Check the Same Things on Every Candidate
Run the same checks on everyone who reaches the same stage, in the same order, recorded the same way. Consistency is what makes a verification step defensible, and it is what keeps the step from turning into a hunch applied to whoever felt off that week. It is also cheaper, since a standing checklist at one stage costs less than an investigation triggered by a feeling.
A workable standard, run at offer rather than at application:
- One page per hire, listing what was checked, by whom, on what date, and what the source said.
- The same three checks for every finalist in a role, with any exception written down and approved by a named person.
- A discrepancy protocol: ask the candidate before drawing a conclusion, in writing, and record the answer next to the finding.
- Nothing spent on authorship, and no informal substitute for it either, since a hallway conversation about whether a document reads as generated is the same unreliable judgment without the paperwork.
What this leaves is a clean division of labour. Records answer the record questions. Work answers the capability question. Conversation answers everything about judgment, which includes the AI question most teams are trying to sneak into the document check: reference-checking for judgment when the reference just says they were great is a better use of the same hour. The document itself has changed meaning under all of this, which is worth reading directly, because a resume now works as a claims index rather than a writing sample.
Common questions
Do you need permission to verify employment history?
Get written authorisation as part of the offer stage, and contact only the employers the candidate listed. Contacting a current employer without explicit permission can cost someone their job before yours is confirmed, which is the one avoidable harm in this whole step. In the United States, where a third-party screening firm runs the check, the Fair Credit Reporting Act adds a standalone written disclosure and the applicant's written authorisation on top 1, so route it through the process your counsel already approved.
What if the former employer only confirms dates and title?
That is the normal answer and it is enough. The claim being checked is that the person held that role over that period, and a dates-and-title confirmation settles it. Performance opinions are what references are for, and they are a separate conversation with different rules. Treat a refusal to say more as company policy rather than as a signal about the candidate.
Is a date discrepancy a reason to withdraw an offer?
Not by itself. Ask first, in writing, and record the answer: contract-to-permanent conversions, acquisitions, parental leave and agency placements all produce dates that do not match a payroll system. What matters is whether the explanation holds together and whether the discrepancy touched something that changed your decision. A withdrawn offer over a two-month gap nobody asked about is an expensive way to learn this.
Should you verify AI-proficiency claims on a resume?
There is nothing to verify against, so the question is really whether to create evidence. No registry issues AI proficiency, vendor certificates cover their own products, and self-reported fluency is a claim about confidence. If the capability matters for the role, put a short piece of real work in the process where someone can see the judgment, and stop treating the resume line as something that could be true or false.
How much should a team spend on authorship analysis?
Nothing, and the money saved covers the checks that do work. No tested tool is reliable enough to carry a decision about a person, the false positives land hardest on people writing in a second language 6, and a wrong call means accusing a candidate on evidence that would not survive being questioned. Spend the budget on employment verification for every hire and one job-shaped step in the process.
References
- 1. 15 U.S.C. 1681b(b) - Conditions for furnishing and using consumer reports for employment purposes uscode.house.gov Supports the United States duties that attach when a third-party screening firm performs the check: a clear and conspicuous disclosure in a document consisting solely of the disclosure, the consumer's written authorisation before the report is procured, and a copy of the report before adverse action.
- 2. Testing of Detection Tools for AI-Generated Text arxiv.org Supports the claim that authorship cannot be checked with a tool: 12 public detectors plus two commercial systems were found neither accurate nor reliable, with a bias toward calling text human-written.
- 3. Monitoring AI-Modified Content at Scale: A Case Study on the Impact of ChatGPT on AI Conference Peer Reviews arxiv.org Supports the distinction between a corpus estimate and an individual verdict: 6.5% to 16.9% of submitted review text was estimated as substantially modified, with the authors noting the trend may be too subtle to detect on a single document.
- 4. Deepfake-Eval-2024: A Multi-Modal In-the-Wild Benchmark of Deepfakes Circulated in 2024 arxiv.org Supports the claim that vendor accuracy figures are benchmark figures: against real circulating material, open-source detectors lost 50% of AUC for video, 48% for audio and 45% for image models. Cited for identity claims only, not for written documents.
- 5. Revisiting Meta-Analytic Estimates of Validity in Personnel Selection: Addressing Systematic Overcorrection for Restriction of Range gwern.net Supports the corrected work-sample estimate of .33 rather than the widely quoted .54, and the point that it sits alongside structured interviews rather than beneath them.
- 6. GPT detectors are biased against non-native English writers pmc.ncbi.nlm.nih.gov Supports the claim that detector false positives land hardest on people writing in a second language: seven detectors over 91 human-written TOEFL essays averaged a 61.3% false-positive rate.
6 sources, numbered by first appearance. How Olive sources claims
General guidance for hiring teams. What works at one company and one volume may not transfer to yours.
Olive assesses how a person works with AI. It does not detect AI-written documents, and it never produces a score, a ranking, or a match percentage for a person. Candidates read the same report the employer reads.