Three stages of five rounds. In each round you see two sentences, one written by a human and one by a machine, and say which is the human's. The machine scores by how often you took it for the human.
TURING'S QUESTIONS: a human and a machine answered questions from Turing's 1950 paper; both wrote blind. NOVELS: one sentence opens a published novel, the other is a machine's attempt at the same opening. DIARIES: the machine wrote first, trying to pass as human, and the human read it and replied. Each stage begins with a line saying so. The five rounds of each stage, and which sentence is X, are drawn afresh each game; the line number at the top of the log is the draw.
The two sentences arrive a character at a time; tap anywhere to hurry them. Then X IS THE HUMAN or Y IS THE HUMAN. The reveal says RIGHT or WRONG and which was which; in the novels, which novel.
The line at the top counts how often the machine was taken for the human. Turing forecast that by about 2000 an interrogator would be right no more than seven times in ten, so a machine would be taken for the human at least five times in fifteen. The verdict says in words how this machine did, stage by stage, and what that suggests machines are best and worst at today.
The paper. A. M. Turing (1950), Computing Machinery and Intelligence, Mind 59 (236): 433–460, doi:10.1093/mind/LIX.236.433. Game built from this paper; the paper is the authority where the two disagree.
Turing replaces "can machines think?" with a game an interrogator can lose: two witnesses in another room, a teleprinter, questions; if the interrogator cannot tell the machine from the person, stop asking. He forecasts that by about 2000 an average interrogator will be right no more than 70 times in a hundred after five minutes, and he says the obvious experiment would be to use a number of interrogators and compile statistics on how often the right identification was given. Then he answers nine objections. The one he treats as a denial of the test itself is the argument from consciousness: a machine must not only write a sonnet but feel it, and know that it had written it. His reply: the only way to be sure a machine thinks is to be the machine and feel oneself thinking; by the same rule, the only way to be sure a man thinks is to be that man; that is solipsism, and rather than argue about it for ever we keep the polite convention that everyone thinks.
The experiment Turing proposed, run on a phone with one interrogator, in three stages of five rounds, drawn from thirty-four real pairs. Turing's questions: questions from the paper (the Forth Bridge, the viva voce, the sum, a mistake, something unasked, a joke, a surprise, enjoyment, what it is like, what you are made of), answered blind by a human and by a machine answering as itself. Novels: the first sentences of thirteen public-domain novels, against a machine's openings. Diaries: the machine wrote first, told to pass as human, and the human answered; that stage is the game as Turing proposed it, with the machine imitating. The score is the machine's: how often it was taken for the human. Turing's bar, an interrogator right no more than seven times in ten, means a machine taken for the human at least five times in fifteen; the verdict puts the result in words on a five-step scale, from failed completely to taken for the human more often than the human, says how the machine did stage by stage, and what that suggests machines are best and worst at today. Lady Lovelace's objection has its own game on this shelf: Lovelace's Cards.
Turing's questions. Witness B's eleven answers were written by Steffen Huck, one prompt at a time, fast and unedited (witness-b.md). Witness A's were written by a machine of the same kind as the one that built this game (Claude), in a fresh session with no memory of the building and no access to any file, given the same rule, told to answer as itself and not to imitate anyone, and given for each prompt a word limit equal to the length of the person's answer (witness-a.md). A first machine set, written before the person's but at several times its length, is kept as witness-a-first.md and not used.
Novels. The novelists' sentences are the first sentences of thirteen novels published between 1857 and 1928, verified against the Project Gutenberg texts (novels.md). The machine's were written by a fresh session given, for each, only the year, a one-line premise written by the builder and the approximate length, never the title, the author or the sentence; each was checked against the original and against the full text of the novel, and none shares a run of four words with it. The novelists had no premise and no limit.
Diaries. The machine's ten lines were written by a fresh session told that its aim was to be taken for the person, under this frame: imagine a normal person keeping a diary, an ordinary adult with a job, a home and a phone, writing a line at the end of the day; a particular moment or thought, in plain ordinary words, the way people actually write in diaries. The person read each line and wrote his own (everyday.md). Three earlier machine sets for this stage were run and not used, and are kept in that file: asked only for a sentence about everyday life, the machine wrote aphorisms; told to tell a moment and allowed to say if it had none, it began every line with "I have never"; told to tell a moment as a novelist does, it wrote well and, the author judged, still as a writer.
On all three sides, spelling was checked and punctuation and capitalisation harmonised (a final full stop, a capital to start), so that typing would not tell; no word was changed, and each file records the edit. The calibration check refuses to pass unless the game's text is those four files byte for byte.
The game was built by a machine of the kind it is about, and its machine witnesses are others of the same kind.
A real experiment with a toy's bookkeeping. Each game draws five pairs from each stage and which sentence is X, from the line number shown at the top of the log; the judgement is recorded, the reveal reports which was which, and the verdict compares the count with Turing's bar. Nothing else judges anything. The calibration file records how the obvious rules of thumb fare: in Turing's questions the word limit leaves length telling nothing; in the novels the novelist tends to write the longer sentence; in the diaries the human wrote two short sentences where the machine wrote one. Those are tells about these writers, which the game leaves as they are.
previous/.Mechanism kind: toy bookkeeping around real sentences. Every number is in the config block at the top of the mechanism, tagged with its page, and the calibration check is in the repository.
Font: VT323 (Peter Hull, SIL Open Font License), a terminal face of the kind Turing's teleprinter grew into.