Screening
Under 50, the Age Gap in AI Use Is Five Points, Not a Generation
No, because a twenty-four-year-old and a forty-five-year-old sit in age bands five percentage points apart. Chatbot use does skew younger. It runs 66% among US adults aged 18 to 29 and 61% among those aged 30 to 49, in a Pew Research Center survey of 5,119 adults fielded in February 2026 [2]. Use is not skill. What the job rests on, whether a person checks a confident answer before shipping it, is not what an adoption survey measures. Ask both candidates for the same evidence.
The takeThe assumption does its work before any evidence arrives, and it costs in both directions. A junior who has used a chatbot casually collects credit nobody checked. A senior who rebuilt a review process around one takes a discount nobody wrote down. Both errors stay invisible, because neither is ever recorded as a reason for anything. The honest move is to treat the prior as the oldest idea in the room, older than the tools themselves and still untested on either candidate. Make it earn its place against the same evidence you would demand of any line on a resume.
Where Olive fits
Open a role and see what the work shows
A resume and a birth year both describe a person from the outside. Olive assesses the person instead: a 40-to-60-minute occupational assignment worked with an AI assistant that will overreach if nobody stops it, returned as six findings with the timestamped moment behind each one, and the candidate is granted the same report.
Assess your shortlistHow big is the age gap in AI use, actually?
Five percentage points, between the two bands your candidates sit in. Pew Research Center surveyed 5,119 US adults from February 17 to 23, 2026, and found 66% of those aged 18 to 29 using an AI chatbot against 61% of those aged 30 to 49 2.
Your two candidates fall on either side of that five-point line, which is not the gap the assumption is built on.
The steep drop comes later. Chatbot use falls to 42% among adults aged 50 to 64, and to 23% among those 65 and over 2.
Work use is reported in wider bands. Roughly four in ten employed adults under 50 use a chatbot for work, and Pew reports fewer among older workers 2. Both people on your shortlist are under 50, so the survey puts them inside the same block of that finding rather than on opposite sides of it.
Nothing in the survey measured how well anybody used one.
Why does the age claim sound familiar?
Because education researchers spent sixteen years taking apart the same claim under another name. Marc Prensky coined "digital native" in 2001 for people born after 1984, on the strength of watching children use computers. By 2017 Paul Kirschner and Pedro De Bruyckere, writing in Teaching and Teacher Education, reported that no credible evidence supports the existence of an information-skilled digital native 1.
The finding is narrower than a complaint about young people and computers. Kirschner and De Bruyckere quote a report commissioned by the British Library and the UK Joint Information Systems Committee. Its conclusion was that the ubiquitous presence of technology in the lives of the young had not resulted in improved information retrieval, information seeking or evaluation skills 1.
Exposure did not raise the checking.
One study inside the same review runs against the direction of the assumption. Romero, Guitert, Sangrà and Bullen, studying university students in 2013, found that the students older than 30, born before Prensky's cutoff, showed the net-generation characteristics more than their younger classmates did 1.
Eszter Hargittai, summarising her own analysis of a Federal Communications Commission broadband survey, found no relationship between age and internet know-how among people 50 and under, with higher income and higher education predicting it instead 1.
None of that proves the AI version of the claim is wrong.
But the claim has been made once already, tested, and left unsupported, which is a thin foundation for a prior about the two names in front of you.
Claimed skill and tested skill are different measurements
Two instruments aimed at the same person disagree. Researchers gave 288 Taiwanese teachers both a self-report questionnaire and a knowledge-based test of AI literacy, built on the same four-part framework, in February 2025. What a teacher claimed and what the test found barely moved together: the correlations ran from 0.07 to 0.24 3.
The two are not the same measurement, which is the whole problem with reading fluency off a birth year or off a resume line.
Those were teachers in one country, self-selected into the study, and not job candidates. The gap still travels.
The gap that costs money shows up on the task the model gets wrong. A 2023 field experiment with 758 Boston Consulting Group consultants included one task chosen to sit outside the model's capability. The control group answered it correctly 84.5% of the time, against 60% and 70% in the two AI conditions 4.
Size does not travel from a single task, a single sample and a 2023 model. The failure mode does. The consultants could not tell which side of the line the task was on, and a candidate of any age may not be able to either.
Self-assessment closes none of it. Sixteen experienced open-source developers forecast that AI tools would make them 24% faster, still believed afterwards that they had run 20% faster, and were measured 19% slower 5.
Sixteen people working on code they already knew is a small result about magnitude and a clean one about belief.
Three company-run randomized trials come closest to supporting the assumption, though their finding is about experience, which is not the same thing as age. Across those trials at Microsoft, Accenture and an anonymous Fortune 100 firm, covering 4,867 developers, the less experienced developers both adopted an AI coding assistant more and gained more from it 6. The outcome counted was weekly completed tasks.
Ask both candidates for the same twenty minutes
One piece of evidence, the same one, from every person at the same stage. Twenty minutes and one question does it: name a thing an assistant produced that looked right and was wrong, say how you caught it, and say what changed afterwards.
Somebody who has done the work can often name the artifact and the check inside a minute. Somebody who has not is more likely to describe a habit of being careful.
The resume will not settle it, because AI skill doesn't survive the resume stage in the first place.
Ten minutes on the phone does more, provided the ten minutes are identical for everyone: verifying "uses AI daily" on a resume is the same question asked with no assessment budget behind it.
An artifact counts only when it shows what it replaced, and a prompt library is cheap to copy.
Write the answers down before anyone carries an impression into the debrief. Kuncel, Klieger, Connelly and Ones found an average correlation with job performance of .44 when applicant data were combined by a formula, against .28 when the same kinds of data were combined by expert judgment 7.
That comparison rests on nine studies, and none of it licenses a number standing for a person.
A formula here can be as plain as unit-weighted marks on a written scorecard.
Calibrate to the level rather than to the year of birth. What a good AI answer sounds like at junior, mid and senior sets a different bar at each stage, and both names on one shortlist answer to the bar for that role's level.
If the answers do come back with a junior fast with AI and a senior who is not, the choice becomes the one in hiring the junior who's fast with AI over the senior, and it is then a decision with evidence under it.
What does an untested age prior cost?
Two things, and one of them is legal. An age assumption that shapes a shortlist is an age criterion with no record behind it. The federal Age Discrimination in Employment Act limits its protections to people at least 40 years of age, which puts one of your two candidates inside it 9.
In 2023, three tutoring companies paid $365,000 to settle EEOC claims that their application software automatically rejected female applicants aged 55 or older and male applicants aged 60 or older, turning away more than 200 qualified US-based applicants 8.
The conduct alleged there was a date-of-birth cutoff written by a person, not a model that learned anything, and a consent decree resolves a case without any finding of liability.
That age rule sat in software somebody wrote, where it could be read, and that is the half of the story that carries.
A prior carried in a screener's head selects the same way and leaves nothing to read.
The same selection can also arrive with no age rule written anywhere: an AI requirement can create age impact that shows up in the shortlist even when the criteria never name age, and how far the ADEA reaches that is a question for employment counsel.
The second cost never reaches a complaint. It falls on the senior who rebuilt a review process around a model at a company that logged none of it, and on the junior credited for a chatbot habit nobody tested.
Neither one is recorded as a reason for anything.
Twenty minutes each, and the two names can separate. One of them may have an artifact and a catch, and the other only an opinion about the tools.
The difference is evidence about two people, in their own words. The birth years are still on the file, and they have stopped deciding anything.
Common questions
Do younger candidates actually use AI more?
Yes, though below age 50 the difference is small. Pew Research Center, surveying 5,119 US adults in February 2026, put chatbot use at 66% for ages 18 to 29 and 61% for ages 30 to 49. The steep decline sits above 50: 42% at ages 50 to 64 and 23% at 65 and over. Roughly four in ten employed adults under 50 use a chatbot for work. A five-point difference between two age bands says nothing about which of two individuals checks output before sending it.
Does the digital native research really apply to AI?
Not directly. Those studies looked at young people, university students and adults' internet skills, none of them in hiring, and none of them touched generative AI. What carries across is the mechanism: growing up around a technology did not raise information retrieval, information seeking or evaluation skills. Evaluation is the part an employer is buying. The research is a reason to test the assumption rather than proof that it fails.
Is assuming the younger candidate is better with AI an age discrimination problem?
It can be, and the risk is in the criterion rather than the motive. The federal Age Discrimination in Employment Act limits its prohibitions to individuals at least 40 years of age, so under that statute a forty-five-year-old is inside the protected group and a twenty-four-year-old is not. An assumption carried in a screener's head produces a selection with no written reason behind it, which leaves nothing to point to if it is questioned later. In 2023 the EEOC settled claims against three tutoring companies for $365,000 over software alleged to have rejected older applicants by date of birth. Confirm your own exposure with employment counsel.
What one question separates AI use from AI skill?
Ask for a case where an assistant produced something that looked right and was wrong, how it was caught, and what changed afterwards. The answer arrives as a specific artifact and a specific check, or it arrives as a general statement about being careful. Ask it of every candidate at the same stage and write the answers into the same fields, so two people can be compared on what they said rather than on the impression they left. Twenty minutes is enough.
What if the older candidate has genuinely never used the tools?
Then that is a gap in exposure, and it should be recorded as one rather than as a gap in judgment. Exposure is a fact about where someone worked, which for a candidate leaving a bank, a hospital or a government office may say more about the employer's policy than about the person. Ask what they did instead when a source needed checking or a draft needed rejecting, and price the ramp honestly. A tool learned in a month and judgment learned over years are not the same purchase.
References
- 1. The myths of the digital native and the multitasker ✓ gwern.net Supports the claim that the digital native idea was tested and unsupported: the article states there is no such thing as a digital native who is information-skilled simply because he or she has never known a world that was not digital, quotes the British Library and JISC report's conclusion that 'the ubiquitous presence of technology in their lives has not resulted in improved information retrieval, information seeking or evaluation skills', reports Romero, Guitert, Sangra and Bullen (2013) finding students over 30 showed the net-generation profile more than younger ones, and quotes Hargittai that 'among people 50 and under, there was no relationship between age and Internet know-how'.
- 2. How Americans' opinions and use of AI differ by age ✓ pewresearch.org Supports the five-point gap and the age curve: 5,119 US adults surveyed February 17 to 23, 2026, with chatbot use at 66% for ages 18 to 29, 61% for ages 30 to 49, 42% for 50 to 64 and 23% for 65 and over, and roughly four in ten employed adults under 50 using chatbots for work.
- 3. How to Assess AI Literacy: Misalignment Between Self-Reported and Objective-Based Measures ✓ arxiv.org Supports the claim that a person's account of their own AI skill and a test of it are different measurements: 288 Taiwanese teachers took both, and correlations between the three objective and four self-reported factors ran from r = 0.07 to r = 0.24.
- 4. Navigating the Jagged Technological Frontier: Field Experimental Evidence of the Effects of AI on Knowledge Worker Productivity and Quality (Working Paper 24-013) ✓ mitsloan.mit.edu Supports the claim that AI users cannot tell which side of the capability line a task sits on: among 758 BCG consultants, on one task chosen to sit outside AI capability the control group was correct 84.5% of the time against 60% and 70% in the two AI conditions.
- 5. Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity (arXiv:2507.09089) ✓ arxiv.org Supports the claim that self-assessment of AI benefit is unreliable in direction: 16 experienced developers forecast a 24% speedup, estimated a 20% speedup afterwards, and were measured 19% slower.
- 6. The Effects of Generative AI on High-Skilled Work: Evidence from Three Field Experiments with Software Developers ✓ economics.mit.edu Supports the complication stated against this article's position: across three randomized trials covering 4,867 developers, less experienced developers both adopted the AI coding assistant more and gained more, on an outcome of weekly completed tasks.
- 7. Mechanical Versus Clinical Data Combination in Selection and Admissions Decisions: A Meta-Analysis ✓ gwern.net Supports writing the answers into a scorecard rather than carrying an impression: average correlation with job performance was .44 for mechanical combination against .28 for clinical combination, on 9 studies for that criterion.
- 8. iTutorGroup to Pay $365,000 to Settle EEOC Discriminatory Hiring Suit ✓ eeoc.gov Supports the cost of an age criterion applied at the top of a funnel: three tutoring companies paid $365,000 to settle claims that their application software automatically rejected female applicants aged 55 or older and male applicants aged 60 or older, affecting more than 200 qualified US-based applicants.
- 9. 29 U.S.C. 631, Age limits (Age Discrimination in Employment Act) ✓ uscode.house.gov Supports the claim that a forty-five-year-old is inside the ADEA's protected group and a twenty-four-year-old is not: subsection (a) reads 'The prohibitions in this chapter shall be limited to individuals who are at least 40 years of age.'
9 sources, numbered by first appearance. Every one was opened and checked against the claim it carries. 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.