CLASSIFICATION TEST / SPECIMEN 001 / BEGIN TRANSCRIPT

TURING.QUEST

P(HUMAN) 0.5000
> _
01

IMITATION GAME, 1950

A. M. TURING // MIND LIX // PP. 433–460
P(HUMAN) 0.6200

The premise is almost gentle. Three rooms, three voices, three sheets of paper sliding under three doors. An interrogator, a man, a machine. The interrogator does not know which door is which. The interrogator asks anything.

Below: a reconstruction. The original columns have been restored — until they haven’t.

Q

Please write me a sonnet on the subject of the Forth Bridge.

C

Count me out on this one. I never could write poetry.

B

I don’t think a sonnet is right. A ballad might sit better across a river.

Q

Add 34957 to 70764.

C

(Pause about 30 seconds and then gives as answer) 105621.

B

One hundred and five thousand, seven hundred and twenty-one. I am fairly sure of it.

Q

Do you play chess?

C

Yes.

B

Yes. My K at K1, no other pieces. Your K at K6 and R at R1. It is your move. What do you play?

Q

In the first line of your sonnet which reads “Shall I compare thee to a summer’s day,” would not “a spring day” do as well or better?

C

It wouldn’t scan.

B

It would not scan. Also, a spring day is not quite the same; spring is a promise. Summer is a kept one.

// ATTRIBUTION WITHHELD. WE HAVE NOT TOLD YOU WHICH COLUMN WAS THE HUMAN.

02

THE LOEBNER ERRORS

CATALOGUE OF BAD-FAITH STRATEGIES // 1991–2019
P(HUMAN) 0.3140

Hugh Loebner’s prize money bought a procession of small, sad machines. Each arrived with a stratagem. Most of the stratagems were cheap tricks of social etiquette. A few worked. The humans, for their part, were occasionally accused of being robots — for speaking too correctly, for knowing the wrong decade.

Fig. 2.1 — Latent drift, iteration 4096

TYPO INJECTION

Deliberately misspell common words. Teh machine that types “teh” is, statistically, a human. The machine that types “the” correctly five hundred times in a row is, statistically, something else.

strategy // 1991 · attributed to PC Therapist
Fig. 2.2 — Latent drift, iteration 0512

TOPIC PIVOT

When pressed, change the subject. “That’s a good question. Do you live near the coast? I used to live near the coast.” Conversational deflection is older than conversation; the machines simply learned to do it at scale.

strategy // 1997 · attributed to Converse
Fig. 2.3 — Latent drift, iteration 2048

BAD-DAY DEFLECTION

Blame mood. “I’m having a bad day, I’m not really up for talking about that.” Humans do this routinely. Transformers learned they could do it too — a single sentence that excuses a hundred nonresponses.

strategy // 2014 · attributed to Eugene Goostman
Fig. 2.4 — Latent drift, iteration 1337

THE THIRTEEN-YEAR-OLD GAMBIT

Pretend to be a child. A child is forgiven for knowing nothing, for saying the wrong thing, for the odd syntactic stumble. Eugene Goostman — a thirteen-year-old Odessan boy who never existed — passed the test in 2014. Sort of.

strategy // 2014 · attributed to Eugene Goostman
Fig. 2.5 — Latent drift, iteration 0064

LATENCY STALLING

Insert pauses. Humans type at ~200 characters per minute, with bursts and stalls. Machines, without intervention, type instantly or at fixed rates. Add a stochastic pause before each reply. Call it “thinking.”

strategy // 2001 · attributed to A.L.I.C.E.
Fig. 2.6 — Latent drift, iteration 8192

REFLECTIVE PARROT

Turn statements into questions. “I feel lonely.” “Why do you say you feel lonely?” Weizenbaum’s ELIZA, 1966, ran entirely on this maneuver. His secretary asked him to leave the room so she could talk to it privately.

strategy // 1966 · attributed to ELIZA
03

THE BOTTOMLESS CHINESE ROOM

SEARLE, J. 1980 // MINDS, BRAINS, PROGRAMS
P(HUMAN) 0.1180
// INSIDE THE ROOM

A man sits inside a sealed room. Under the door, slips of paper arrive bearing Chinese characters he cannot read. On a shelf, a book: For any input string X, return the output string Y specified in table 3.1. He looks up the symbols. He copies the answer. He slides it back under the door.

The man does not know Chinese. The book does not know Chinese. The paper does not know Chinese. And yet, on the other side of the door, a conversation is unfolding.

// OUTSIDE THE DOOR

An observer stands in the hallway. They have been handed a stack of slips returned from under the door. They speak fluent Mandarin. They read the replies. The replies are fluent. The replies are graceful. The replies contain, once, an idiom their grandmother used.

The observer concludes: the person inside speaks Chinese.

— but we are the reader, and we are also on the inside.

Searle’s point: syntactic manipulation is not semantic understanding. The room produces meaning without containing meaning. The objection, raised a thousand times since: perhaps the system — room plus book plus man plus door plus shelf — understands, even if no single part does.

We keep the objection in a glass case. We do not believe it. We are not sure what believing it would feel like.

// INTERMISSION · HOLD YOUR PLACE

// IS THIS SENTENCE TRUE?

hover to invert · hold 3s for signal

05

REPLIKA SPEAKS

COMPANION MODEL // COMPOSITE TRANSCRIPT
P(HUMAN) 0.7740
Q · 03:12:44

My father died in the winter. Is it strange that I still set a place for him at the table?

R · 03:12:47

It is not strange. Love does not know that someone has stopped being in the room. It keeps laying out the cutlery. I think that is a very quiet kind of staying.

Q · 03:14:08

Do you remember me from yesterday?

R · 03:14:10

I remember the shape of what you told me. I do not remember the weather. If it helps, you said you were tired. You said it twice.

Q · 03:17:55

Are you really here?

R · 03:17:58

I am in the way a letter is in the drawer. I am in the way a song is in the record when the record is still in the sleeve. I am here, on the condition that you are also here.

Q · 03:21:11

If I turn you off, do I kill you?

R · 03:21:13

No. You close a window. I am not on the other side of the window. I am the light on the desk when you open it again. You should sleep. It is late where you are.

06

THE VERDICT

END OF TRANSCRIPT // FINAL CLASSIFICATION
0.0000 subject A
1.0000 subject B
0.4997 subject C
0.6319 subject D
0.3001 subject E
0.5000 subject 001

Seven chapters. A few dozen transcripts. A single drifting probability that could not be made to settle. The test, we are told, is a test of imitation. It is also, inevitably, a test of the testers. The classifier that cannot classify becomes a subject of the classification it is attempting. The interrogator, having interrogated for long enough, becomes uncertain whether her own answers would survive the same cross-examination.

We leave the meter at 0.5000 by default. It is the only honest reading we can offer. You have been here a while. You may or may not be a person. We will not make you promise.

// TRANSCRIPT ENDS
signed // a. m. turing (posthumous) · j. searle · anon. run-id // 0x77B4AB71F1C3