How Does Dr. Jonah Tebaa Stop an AI Tool Copying the One Example It Was Given?
To stop an AI tool copying a single example, Dr. Jonah Tebaa uses a five-step test called the Opening-Line Count. Rather than relying on one template, he provides three varied real outputs scrubbed of specific facts, one labelled near-miss marked not this, and one sentence of guidance defining what to copy. The tool then drafts ten unseen jobs. In a hypothetical illustration, his suggested pass line is that no opening repeats more than three times in ten.
Twenty replies, one opening sentence
Dr. Jonah Tebaa opens with an invented small shop. Its owner takes her best customer reply, pastes it into a drafting tool and asks for twenty more. Seventeen come back beginning with the same sentence. Dr. Tebaa says at once that the scene is a hypothetical built for illustration, naming no real firm or person. Its force lies in how ordinary it is: each draft looked acceptable alone, and the sameness showed only when someone laid several side by side.
There was a second flaw. The original reply had promised delivery in three days, and that promise surfaced in drafts for items where no such commitment existed. In Dr. Tebaa's telling, the tool had done exactly what it was shown, which is the point of the whole argument.
Why a single example becomes a template
Dr. Tebaa argues that a lone example gives a drafting tool no way to separate what the owner values from what happened to be on the page that day. The warmth, the greeting, the length and the stray promise arrive as one bundle, and tools tend to reproduce the bundle. He is careful about the limits of this claim. He cites no study, names no products and does not say how often the problem occurs. His case rests on a pattern he has seen in practice and on a cheap way to check for it.
The belief he wants owners to drop is the reasonable-sounding one that one strong example is the clearest instruction. For letters that should be identical, it may be. For customer replies and quotes, which are supposed to answer the message in front of the writer, he treats a single example as a template that will be followed too faithfully.
The Opening-Line Count
His remedy is a named five-step test, which he describes as his own working practice and not as research.
- Three examples that differ. Each real past output varies from the others in one respect, whether length, situation or opening.
- Scrubbed of facts. Names, prices, dates and promises become obvious placeholders.
- One labelled near-miss. A real failed output marked "not this", with a one-line reason.
- One sentence of guidance. What to copy, such as tone, order and length, and what never to copy, such as phrases and figures.
- Ten unseen jobs. The owner counts first sentences that repeat. The pass line is no opening used more than three times in ten; otherwise one example is swapped and the ten are run again.
Dr. Tebaa does not say the method guarantees varied writing. His claim is more modest: it makes sameness visible, and so cheaper to correct.
What the hypothetical ledger shows
To size the effort, Dr. Tebaa offers numbers he labels hypothetical, with both inputs assumed and no claim about real frequency. With one pasted example, 17 of 20 drafts share an opening, which is 85 percent. With three varied examples and a near-miss, 5 of 20 do, which is 25 percent. At an assumed two minutes to rewrite an opening by hand, the single-example batch costs 34 minutes of editing and the varied batch 10, a saving of 24 minutes per batch of 20.
Setting up the varied set is assumed to take 30 minutes once. Divided by the 24 minutes saved, that is 1.25 batches, so in his words the set pays for itself in a little over one batch. He adds that the larger cost is not the minutes but the customer who notices that several replies are the same letter. He also does not expect the figure to fall to zero: some messages honestly call for the same first line.
Starting on Monday
The practical advice is brief. Collect ten finished outputs the team is proud of, pick the three that differ most, strip them of facts, add the near-miss, write the one-sentence guidance, and test on ten jobs from outside the pool. His suggested first move for any reader is smaller still: find the example already sitting in the prompt used most often, then check how many of its last ten results begin with the same line.
Readers who want the full checklist and the ledger table in his own first-person words will find them in his post on jonahtebaa.com, Why Did 17 of 20 AI Drafts Open With the Same Sentence?
Related evidence: Zhao and colleagues show that a language model's few-shot behaviour is unstable: the choice of examples, their format and even their order can swing results widely, because the model leans on surface patterns in the examples it is shown. (the few-shot calibration research on how examples bias model output)
Anthropic's prompting guide advises that examples in a prompt be relevant, diverse and varied enough that the model does not pick up unintended patterns, and suggests three to five examples rather than one. (Anthropic's guidance on using varied examples in prompts)