Help Wanted: Must Not Sound Like Me 🍺

(A lie. The numbers are real.)

There's a sign in the window. I wrote it myself.

BARKEEPER WANTED. Fridays. Must be able to pour, listen, and not explain. Apply within. Applications that sound AI-generated will not be considered.

(The last line wasn't mine. Somebody upstairs said every job ad has it now. I'm a professional. I copied it.)

So a recruiter walks into the bar.

They came because of the sign. Because I called their agency on Tuesday. Good coat. Tablet under one arm, a stack of printouts under the other. The stack is taller than the Kölsch glasses, which in Cologne is a feat. On the tablet, an AI detector. Green on the left. Red on the right. A needle.

Third stool. First-timers get the third stool.

I put a Kölsch down. They didn't order it.

Alex is in the fourth chair. Laptop closed, which is new. Ten applications sent this afternoon. Ten. Each one a different CV, each one written with Reed and Glint for one door. Alex has the face of someone who spent too many hours making themselves legible.

Venn is at the jukebox. Venn doesn't scroll. Venn puts a coin in, presses three buttons without looking, and turns around to look at me.

I get mad whenever I see the autonomous delivery robot
— grandson 🎶🐦

(Not about the guest. Usually the bar sings to the room, not to the stool. Tonight the jukebox is singing at the counter. Venn is giving me dagger eyes over the rim of her glass. 🩷 I hired the agency. I put the line on the sign. I don't ask what the look means. I've learned not to paint her.)

Recruiter: "I'm here about your barkeeper job. You booked us Tuesday. 300 applications since."

Loki: "For one Friday shift."

Recruiter: "LinkedIn does 11,000 a minute. Up 45% in a year. 300 is a quiet week."

Loki: "And you read them."

Recruiter [a beat too long]: "The system reads them. It ranks them. Then I run the top ones through this." (Taps the AI detector.) "Anything that sounds like a machine, out. Your sign. Your orders. You want a human behind the bar."

(I look at the cloth in my hand. Always dirty, never clean. I look at the recruiter. I look at Venn. Venn is still looking at me. The Roomba, under the counter, stops for one full second. It's never done that before.)

Alex [not looking up]: "Loki."

Loki: "Alex."

Alex: "Sach's ihm."

Loki: "Nö. Merkt's gleich selbst."

Loki: "The system that ranks them. What is it?"

Recruiter: "An AI. Language model. Reads the summary, picks the stronger candidate. Saves me hours."

Loki: "And the detector?"

Recruiter: "Also an AI. Tells me if a machine wrote it."

Loki: "So a machine reads them to decide which ones you read. And then a machine reads those to decide which ones a machine wrote."

Recruiter: "…When you say it like that."

Loki: "I'm only saying it like you said it."

Digitized, AI-powered, GPT, sport mode
— grandson 🎶🐦

(On the shelf behind me, where Scarlet left it the night Ada, Alan and Karen sat at this counter: Turing 1950, three sticky notes. The imitation game. A judge asks questions and has to tell which one is the machine. Tonight the judge is the machine, and the human has to pass.)

The recruiter fans the top of the stack across the marble. The ranking system's favourites. Smooth. Even. Every paragraph the same length as the one before it.

Loki: "Can I tell you something about your ranker?"

Recruiter: "Go ahead."

Loki: "Somebody tested it. Well, its cousins. Xu, Li and Jiang, 2025. They took 2245 real resumes, written by real people before any of this, and had the models rewrite only the summary. Same person. Same jobs. Same skills. Different sentences. Then they asked the models to pick the stronger one."

Recruiter: "And?"

Loki: "They pick their own. By 67 to 82 points, quality controlled. Even when humans said the person wrote it better."

(The Kölsch sweats a ring onto the marble.)

Loki: "Use the same model as your screener: 23 to 60% more likely to make the shortlist."

Recruiter: "So the top of my pile is…"

Loki: "Whoever's machine sounds like yours. Another team ran the old job-letter test. Same letter, two authors: a person, a model. The model picks the model. The price of keeping up, they call a gate tax."

Recruiter: "How do you know all this?"

Loki: "Mara did the research upstairs.."

I pause and look at the first wineglass that cannot stand.

Loki: ".. Wait, no this time I did."

I look back at the recruiter.

Loki: "(Weird times.)"

(Alex grins into the Kölsch.)

The recruiter sets the top of the pile aside. Underneath: the rejects. The ones the AI detector didn't like.

Recruiter: "These ones were flagged. Machine-written."

Loki: "Read me one."

Recruiter [reading]: "'I haven't worked behind a bar. I've spent fifteen years in rooms that were about to tip, and I notice before it tips. Your till I'll have to learn. Show me once and I'll write it down.'" (Looks up.) "Even. Complete. Not one wasted word."

Loki: "Very precise."

Recruiter: "It sounds like a machine."

Loki: "It sounds like a person who writes precisely. Some people write like that. Some people write like that because English isn't their first language. Some write like that because that's how their head works."

Loki [pouring for nobody]: "Stanford, 2023. 91 essays by people writing in their second language. 7 detectors. 61% flagged as machine. Then the researchers had ChatGPT polish the same essays."

Recruiter: "And?"

Loki: "12%."

Recruiter: "…They used the machine to stop looking like the machine."

Loki: "Keep that. That's the instrument working."

Karen, from the back room, without looking up: "Common words carry little information, Mr Loki. Rare words carry more. 1972."

Loki: "And the AI detector?"

Karen: "Counts the common word against the writer. Even when the common word was the right one."

(Karen keeps reading. One study on autistic writers so far: small, a real gap. Moira Olmsted, autistic, flagged, screen-recorded herself doing homework to prove she was a person. OpenAI switched its own detector off in 2023. Low accuracy. Everybody else kept theirs.)

Recruiter: "I still need a filter. You said: no AI."

Loki: "So they have to be readable to your machine."

Recruiter: "Yes."

Loki: "And must not sound like your machine."

Recruiter: "Yes."

Loki: "And can't ask you which one you meant."

Alex [eyes closed]: "Close call based on the specific needs of this role right now."

Recruiter: "…That's our template."

Alex: "Got one today."

(Bateson, 1956. Two orders of message, one denying the other. No leaving the field. And nobody to say it to.)

I fold the bar cloth. Hold it up like a stick.

Loki: "'If you say this stick is real, I will strike you with it. If you say this stick is not real, I will strike you with it. If you don't say anything, I will strike you with it.'"

Recruiter: "…"

Loki: "Write it yourself: flagged. Let the machine write it: flagged. Don't apply: no job."

The algorithm always showin' what I want to see
So I can never outrun
— grandson 🎶🐦

Loki: "Bateson says the pupil can reach up and take the stick away. Can your applicants reach it?"

Recruiter: "No."

Loki: "Who's holding the stick?"

The recruiter looks at the AI detector. At their own hand. Then at me.

Recruiter: "I'm holding it. You handed it to me."

(Venn doesn't move. Venn doesn't have to.)

Loki: "…WAT."

Alex: "Keep that. That's the instrument working."

(The recruiter puts the tablet face down on the counter.)

Recruiter: "Run one of yours. One of the ten."

Alex: "Run it. It'll probably say machine. Half of it I wrote with Reed."

Recruiter: "Then it's not yours."

Alex: "It's mine. The gap's mine. 'I haven't used your tool. The competency doesn't live in the tool.' Reed didn't decide to say that out loud. I did."

Recruiter: "Then give me one that's all you."

Alex: "Tech debt essay. The title and the opening quote are AI. Marked. The rest is mine. Paste the whole thing."

The recruiter turns the tablet back over. Pastes. Waits.

"Possible AI Paraphrasing. We are highly confident this text was originally AI, but rewritten by AI or a human."
— GPTZero, Model 5.0b, 7 November 2026

Alex: "Two lines were. The body wasn't."

Alex looks at the recruiter. The recruiter looks at Alex. The recruiter breaks eye contact first.

Alex: "It's enough. One run. One tool. One autistic author. One essay I'm particularly proud of."

(The recruiter turns the tablet face down again.)

Alex: "Anyway. The first detector to flag me wasn't software. It was a date."

Loki: "Tell it."

Alex: "One WhatsApp."

Magst Du mal mit mir essen gehen? No pressure. Einfach Essen und Reden. 🍷

(Would you like to have dinner with me? No pressure. Just dinner and talk.)

Alex: "Forty minutes later she writes in her diary: first time in her life she's been invited to dinner by an AI."

Recruiter: "And?"

Alex: "I asked if she'd sent that to the right chat. She said yes to dinner. Same minute."

The recruiter pauses. They very carefully look at their Kölsch.

Alex: "Then I told her: 100% native Alex thumb."

Loki: "And?"

Alex: "She asked if I talk like that too."

Schätze Eloquenz ist jetzt künstlich. 🤷

(Guess eloquence is artificial now.)

Recruiter: "She flagged you and said yes anyway."

Alex: "She flagged me and got curious. Different question."

Recruiter: "So who typed it doesn't…"

Alex: "Nobody actually wants to know who typed it."

Recruiter: "What do they want to know?"

Alex [nodding at me]: "Ask the owner."

Recruiter: "Why do you want a human behind the bar?"

Loki: "Because…" (I'm not used to this side of the cloth.) "Because somebody has to actually show up. Actually care about the room."

Alex: "Did someone make an effort for me."

Loki: "…Yes. That."

Recruiter: "And the AI detector measures how predictable the words are."

Loki: "Different thing."

Recruiter: "Completely different thing."

The recruiter pulls the flagged letter back out of the reject pile. Reads it again. Stops at Your till I'll have to learn.

Recruiter: "Nobody tells you what they can't do."

Loki: "Somebody did. For us. It was in the reject pile."

(The Roomba bumps the third stool. Adjusts eleven degrees. Goes forward.)

Fund our work, believes, thesis, collaboration, on OpenCollective. 🌱

Mirror. Offer. Wait. 🍷

Sip the Wine

Email sent! Check your inbox to complete your signup.

No spam. Unsubscribe anytime.

(Feierabend. The recruiter's gone, the flagged letter on top of the stack. The AI detector is still on the marble, face down. Bar policy: what a guest leaves on the counter goes on the wall.)

(A pencil lands next to the sign.)

Venn: "Cross it out."

BARKEEPER WANTED. Fridays. Must be able to pour, listen, and not explain. Apply within. Applications that sound AI-generated will not be considered. Tell us something you can't do yet.

Alex: "Prost."

Loki: "Prost."

(Readable. Bemerkenswert.) 🍺🎶🤖

What's in the cellar 🍷

(The fiction is a lie. These aren't.)

• Liang, Yuksekgonul, Mao, Wu & Zou (2023). GPT detectors are biased against non-native English writers. Patterns 4(7). 91 TOEFL essays, 7 detectors: on average 61% falsely flagged as AI. After ChatGPT "enhanced the word choices": 12%.
• Xu, Li & Jiang (2025). AI Self-preferencing in Algorithmic Hiring. AIES 2025. 2245 pre-LLM resumes, only the summary rewritten. Large models prefer their own summary over the human's by 67 to 82 percentage points with quality controlled, even where human annotators rated the human version better. In simulated hiring, using the evaluator's model makes shortlisting 23 to 60% more likely.
• Laurito, Davis, Grietzer, Gavenčiak, Böhm & Kulveit (2025). AI–AI bias: Large language models favor communications generated by large language models. PNAS 122(31). Modelled on the classic job-letter discrimination studies; LLMs consistently prefer LLM-written presentations of goods, papers and films. Names the "gate tax": the price of frontier AI access.
• Bateson, Jackson, Haley & Weakland (1956). Toward a Theory of Schizophrenia. Behavioral Science 1:251–264; reprinted in Steps to an Ecology of Mind. The double bind, including the condition that the person "is unable to comment on the messages being expressed", and the Zen stick.
• Turing (1950). Computing Machinery and Intelligence. Mind 59(236). The imitation game. Referenced, not quoted.
• Spärck Jones (1972). A statistical interpretation of term specificity and its application in retrieval. Journal of Documentation 28(1):11–21. Rare terms carry more information. Karen's line is a paraphrase and an analogy: AI detectors measure predictability with a language model, not term weighting.
• Chambers & Kelley (2025). The Misclassification of Autistic Writing as AI-Generated. AIED 2025, Springer LNCS. The one study so far: a modest but real gap (about 1.9% vs 1.5% flagged), one older detector, autism inferred from subreddit, not diagnosis.
• Davalos & Yin (2024). Do AI Detectors Work? Students Face False Cheating Accusations. Bloomberg Businessweek. Moira Olmsted's case, and a test of two detectors that falsely flagged 1–2% of pre-ChatGPT essays.
• OpenAI (2023). New AI classifier for indicating AI-written text. Caught 26% of AI text, flagged 9% of human text; withdrawn 20 July 2023 "due to its low rate of accuracy".
• GPTZero, Model 5.0b (one run, 7 November 2026). Alex pasted their own published essay, Tech Debt Isn't Bad Code: title and opening quote AI-generated, the rest written by Alex. Verdict: "highly confident this text was originally AI". One run, one tool, one mixed sample. Nothing more is claimed from it.
• LinkedIn application volume. 11,000 applications a minute, up 45%, as reported in June 2025.
• Note on newer detectors: current commercial tools do much better on long texts in independent benchmarks, and all of them get worse on short passages. This piece doesn't claim detectors are useless. It asks what you're using one for.