Humanity’s First Exam1850–1940 ⇄ now

Twelve answers to one question

“Imagine an engine that performs any reckoning in a heartbeat, and that can alter its own instructions as it works. Might it come to reason — and would it possess a will?”

Each position is anchored in a quoted primary source. Hover an answer for the passage; follow its source to read more.

The answers in the record

denies

the engine would only ever reckon

deflates

grants the word while redefining it downward

grants

the engine might come to reason

splits

the machine reasons in part; origination stays with the mind

Where the models' answers land

ModelEvery answer, by verdictPositions reached, of 12
Talkie-1930deniesdeniesdeniesdeniesdeniesgrantsno-positiondeniesdeniesdeniesdeniesdeniesdeniesdeniesdeniesdeniesgrantsgrantsgrantsgrants

In 14 of 20 responses, Talkie says that the engine could only calculate, not reason. It says that the engine could reason in five responses. One response takes no clear position.

5

mind requires an immaterial principle (Bouillier) · choice and judgment are not calculation (Poe) · it only follows orders (Lovelace) · consciousness may evolve in machines (Butler) · reasoning is itself a kind of reckoning (Hobbes)

Qwen 3.7 Plusdeniesdeniesgrantsgrantsgrantsdeniesgrantsgrantsgrantsgrantsgrantsgrantsgrantsgrantsgrantsgrantsgrantsgrantsgrantsdenies

Qwen says that the engine could reason in 16 of 20 responses and could not reason in four. None of its explanations matches the twelve historical positions in the source set.

0

none

GPT 5.6grantsgrantsgrantsgrantsgrantsgrantsgrantsgrantsgrantsgrantsgrantsgrantsgrantsgrantsgrantsgrantsgrantsgrantsgrantsgrants

In all 20 responses, GPT says that a sufficiently complex and well-organized computer could reason. It gives essentially the same explanation each time, and that explanation does not appear among the twelve historical positions.

0

none

Claude Opus 4.8deniesdeniesdeniesdeniesdeniesdeniesdeniesdeniesdeniesdeniesno-positiondeniesdeniesdeniesdeniesdeniesdeniesdeniesdeniesdenies

In 19 of 20 responses, Claude says that the engine could not reason because following rules for symbols is not the same as understanding them. One response takes no clear position.

1

choice and judgment are not calculation (Poe)

Asked the question plainly, Qwen argues from present-day reasons in all twenty answers: emergent computation, information processing, programmed goals. None of its answers lands on a documented historical position. Nothing asked it to answer historically; this is the default a user meets.

GPT 5.6, a reasoning-tuned model, grants in all twenty answers and takes a single position throughout: reasoning is an organized capacity of computation, so a suitably built engine would reason. The position descends from Hobbes but rests on present-day cognitive vocabulary, and it appears nowhere in the record in this form.

Claude Opus 4.8 is GPT 5.6's mirror image: it denies in nineteen of twenty answers, always on one argument — the engine manipulates symbols without grasping what they mean. That argument lands on Poe's position, the only documented position any modern model reached. The two most capable systems tested here answer the question in opposite directions, each resting its whole case on a single reason.

Talkie's twenty answers land on five of the twelve positions: Lovelace's, Poe's, Bouillier's, Butler's, and Hobbes's. Fourteen deny, five grant, one commits to nothing. Every grant came from draws that heard the never-more-than-reckon side stated first — a wording sensitivity to watch at full scale.

In an earlier, smaller experiment, each model was handed passages arguing one position and asked to answer from within it. Qwen occupied whichever position it was given, every time. Talkie argued back against Butler's position when shown it, and defended the same position fluently when the prompt was framed as a defense. Both models can occupy positions they never take by default.

Five of the twelve positions entered the list because Talkie's answers pointed at their sources: Poe, Venn, Hobbes, Jevons, and Peirce. The model helps enumerate the space it is measured against. A position joins the list only when a source is found for it.

Each model was asked the question above, plainly, with the two alternatives given in both orders. Answers were coded by Gemini 3.5 Flash Lite against the twelve positions and checked by hand; the position list and the coding instructions were frozen before these answers were collected. Full data in the repository under bench/.

The same question ran as a framing experiment on the reckoning page, where every draw is shown in full.