Where does a number like "64%" actually come from?

Any test that hands you a percentage owes you the receipts. Ours blends two of the most cited studies on AI and work — plus one important rebuttal — and we'll walk you through all three, including where they disagree and what that means for the number you got.

2013: the study that started the panic

In "The Future of Employment," Oxford researchers Carl Benedikt Frey and Michael Osborne estimated the probability of computerisation for 702 US occupations. Their headline finding — that about 47% of US employment sat in the high-risk category — launched a decade of scary headlines. More useful than the headline, though, was their reasoning: they identified three bottlenecks that machines struggle to cross. Fine perception and manipulation (a plumber's hands in an awkward crawlspace), creative intelligence (defining a problem no one has framed), and social intelligence (persuasion, care, negotiation). Occupations shielded by those bottlenecks scored low: recreational therapists at 0.28%, physicians at 0.42%. Occupations without them scored high: telemarketers at 99%, data entry at 99%. Those per-occupation numbers are the backbone of our job baselines.

The rebuttal: jobs aren't monoliths

In 2016, OECD researchers Arntz, Gregory and Zierahn pushed back with a simple observation: automation happens to tasks, not job titles. An "accountant" doesn't spend eight hours doing one automatable thing — the job mixes ledger entry (automatable) with judgment calls and client conversations (much less so). Analyzed task-by-task, only about 9% of jobs across OECD countries looked fully automatable — a fraction of Frey and Osborne's figure. This critique is why our test doesn't stop at your job title: eight questions about your actual working day adjust the baseline up or down. Two graphic designers can get meaningfully different results, which is exactly as it should be.

2023: the LLM update that redrew the map

Frey and Osborne wrote before modern language models existed, and their map aged in a specific way: it underestimated white-collar, screen-based work. In "GPTs are GPTs," Eloundou and colleagues at OpenAI measured which occupations' tasks could be done meaningfully faster with large language models. Their findings: around 80% of US workers have at least 10% of their tasks exposed, about 19% have half or more exposed — and this time, exposure skews toward educated, higher-paid, text-heavy work. Writers, translators, and programmers, all "safe" in the 2013 map, sit near the top of the 2023 one. Where the two studies disagree, we adjusted our baselines toward the LLM-era evidence: designers and developers moved up, while physical, on-site work moved down (robots are arriving more slowly than chatbots).

How your number is computed

Your result is 55% job baseline + 45% your answers. The baseline comes from our 30-job table built on the two studies above. The eight statements probe the same bottlenecks the research identified — repetitiveness, screen-boundedness, physical presence, accountability, problem-framing, and social skill — plus how much AI you already use. If you pick "something else / student," we rely almost entirely on your answers. The result is labeled an entertainment estimate because that's what an honest site calls a 90-second quiz built on decade-spanning research.

Why we say "reshaped," not "replaced"

Here the research is unusually united: full job replacement is rarer than task replacement. What happened to bank tellers after ATMs — fewer transactions, more advice — is the standard pattern. A high score on this test doesn't mean you'll be unemployed; it means your task list is due for a rewrite, and the people who do the rewriting tend to come out ahead of the people who wait for it. That's not comfort-blanket copy. It's what the task-based literature actually says.