Whether you can tell a passage written by a person from one written by a model — and, from the confidence you attach to each answer, whether your belief that you can is justified.
The apparatus was wrong. The interview turned out to be right.
Philip K. Dick built the Voight-Kampff in Do Androids Dream of Electric Sheep? in 1968: a few
provocative questions, a machine watching your breathing, your pulse, your blush and your pupils, and a
verdict about what you are. The machine half is a polygraph, and the National Research Council's 2003
review is blunt about what a polygraph can do — the states it measures arise without deception too.
The conversational half has now been run properly, and it failed in the direction the film did not
anticipate. In a pre-registered three-party Turing test, GPT-4.5 given a humanlike persona prompt was
judged to be the human 73% of the time — more often than the actual human. The same model with no
persona prompt was picked out as a machine more often than chance. The prompt mattered more than the
interrogator did.
Jones CR & Bergen BK (2026), Large language models pass a standard three-party Turing
test, PNAS 123(21), doi:10.1073/pnas.2524472123.
What you are about to do, stated completely
Every pair below is exactly one passage written by a person and one passage written by a model.
Never two of one. Never a trick pair. They are matched for topic and for length. You pick the machine, you
say how sure you are, and you tell us whether you recognised either passage.
Nothing is hidden from you and nothing about this is a trap. When you are scored, you get the full
provenance of every pair you saw, on the same screen, before there is anywhere to click away to.
This is the built-in SEED set: the human halves are public-domain literary passages from 1813–1899 and the model halves were written to match them in topic, length and register. It demonstrates the instrument honestly; it is not yet a measurement of telling a person writing today from a model writing today. Both reasons are on the result screen.
Confidence, and why it starts at 50
With two passages and one answer, a coin is already right half the time — so 50% means "I have no idea"
and 100% means "I would bet the house". The scoring cares more about whether your confidence tracks your
accuracy than about the accuracy itself. That comparison is the finding: the usual result is that people
believe they can do this considerably better than they can.
This result will never be worded as: “You scored in the replicant range.” — not because a test
may never tell you something true and flattering, but because this particular measurement is a skill against
a moving target rather than a trait. The models change every few months; your d′ is against one vintage of
one stimulus set, and it is printed with that date attached.
Your key
Your browser made you a random key. We store a one-way hash of it, never the key. It is the only thing
linking this sitting to your next one — write it down. Without it we cannot delete your entries later,
because we will have no way to know which ones are yours.
…
Before you start
Under a minute. None of it is required, every question can be skipped, and
"prefer not to say" is recorded as itself rather than turned into a default. This is not
part of the exam and nothing here is scored.
The pairs
Loading…
Where this comes from
Jones CR & Bergen BK (2026), Large language models pass a standard three-party Turing test, PNAS 123(21), doi:10.1073/pnas.2524472123 — GPT-4.5 with a persona prompt judged human 73% of the time; preprint arXiv:2503.23674 (2025).
National Research Council (2003), The Polygraph and Lie Detection, National Academies Press — why the apparatus half of the Voight-Kampff was always wrong.
Macmillan NA & Creelman CD (2005), Detection Theory: A User’s Guide, 2nd ed. — d′ = √2·z(pc) for two-alternative forced choice, and the extreme-cell correction.
Murphy AH (1973), Journal of Applied Meteorology 12(4):595–600 — the Brier-score decomposition the calibration curve is read against.
Philip K. Dick (1968), Do Androids Dream of Electric Sheep? — the hook, and never the frame of the result.
The rules this page is bound by
Report the measurement, not the category."Your VVIQ score was 19 out of 80. Scores at or below 24 are the range in which most published aphantasia research recruits participants." Never "you have aphantasia."
Name the reference class, every time.A percentile is meaningless without saying among whom. Ours is: among people who found this page and chose to take the test. That group is systematically enriched for exactly the traits being measured, because people who suspect something about themselves go looking.
Consistency, not confidence.For every instrument the second administration is part of the instrument. A one-shot result is a reading, not a finding.
Never route a result to anything consequential.No credential tier, no karma weight, no grant eligibility, no payment. The moment a score buys something, people optimise it and the data dies.