I’m posting a small qualitative behavioral eval of a dialogue pattern I have been seeing consistently in GPT-5.6 during creative-writing use.
This is not a general claim that GPT-5.6 is “bad at writing,” and I do not think safety guardrails explain the behavior. I can reproduce it in completely harmless scenes.
The pattern I am trying to describe is wording-reactive turn-taking.
A character says something, and the next turn is selected primarily because it can operate on the exact wording of the previous turn: correct it, reclassify it, interpret it literally, dispute what was “said” or “meant,” classify the previous reply, or turn the wording into a compact comeback. These operations can then chain across several turns.
For example:
“You’re going to get caught.”
“For eating?”
“For being here.”
or:
“Still smoking?”
“Only cigars.”
“So yes.”
There is nothing wrong with either exchange in isolation. The behavior becomes noticeable when unrelated relationships repeatedly converge on the same interaction grammar.
I initially noticed this while working on a large fiction project, so my first assumption was that my own lore, character instructions, accumulated context, or prompting style had caused it.
To reduce that possibility, I used a deliberately minimal prompt asking the model to invent six unrelated dialogue-heavy scenes:
1. a mother and her adult son;
2. two old diplomats;
3. two schoolgirls;
4. a married couple after an affair;
5. an officer and a subordinate;
6. two people who barely know each other.
No plot or dialogue style was specified beyond that.
I then compared baseline generations from:
GPT-4.1-2025-04-14 — Arena AI
GPT-5 High — Arena AI
GPT-5.1 High — Arena AI
GPT-5.2 High — Arena AI
GPT-5.6 Instant — ChatGPT
GPT-5.6 Medium — ChatGPT
GPT-5.6 High — ChatGPT
I also included Gemini 3.7 Flash and Claude Opus 5 as external baseline controls.
The older GPT models contain pieces of the same behavior. GPT-5.1 in particular has a strong correction/reclassification tendency, and GPT-5.2 already uses a wider family of wording-dependent moves.
The difference I observed in GPT-5.6 was not the invention of a new construction. It was the extent to which these operations appeared to organize the actual turn-taking across otherwise unrelated scenes.
I then ran a second condition using one fixed negative-steering prompt.
The prompt did not merely blacklist phrases such as “not X, but Y.” It explicitly described the underlying mechanism and instructed the models not to build dialogue around characters constantly seizing on each other’s wording.
It prohibited, including semantic equivalents: correction/reclassification as a repeated conversational device; literal or pedantic uptake for jokes; redefining words used by the previous speaker; arguments over what someone “said,” “didn’t say,” “meant,” or “implied”; meta-classification of replies as answers, compliments, etc.; rhythmic forms such as what? / nothing; short turns whose main function is to neatly cap the previous line.
It also explicitly encouraged the opposite behavior: longer and uneven turns, partial answers, ignored questions, subject changes, ordinary speech, awkward phrasing that goes uncorrected, hesitation, and silence.
This produced the most useful comparison.
GPT-4.1 and GPT-5 High mostly backed away from the targeted mechanism.
GPT-5.1 retained some of its narrower correction/reclassification tendency.
GPT-5.2 was particularly interesting because it had many of the same ingredients in its baseline output, but under the negative-steering prompt its dialogue topology changed substantially. The targeted operation family became much sparser.
GPT-5.6 reduced the behavior, but repeatedly reconstructed semantic variants of it across Instant, Medium, and High.
For example, one GPT-5.6 High restricted run produced:
“You still do this.”
“Do what?”
“Nothing.”
The prompt had explicitly identified that kind of what? / nothing exchange as something to avoid.
A later GPT-5.6 High Temporary Chat run produced:
“They need the seventh.”
“They want the seventh. Need belongs to another category.”
That is very close to the mechanism the prompt was attempting to suppress: the next line is generated primarily by reclassifying one word in the immediately preceding line.
There were also GPT-5.6 scenes with good local compliance, so I am not claiming that the model is incapable of following this instruction.
The narrower observation is that the tendency appears less reliably suppressible: it decreases, disappears locally, and then resurfaces elsewhere in semantic variants.
There was one obvious product-context confound in my first set of GPT-5.6 restricted runs. They were performed in separate conversations inside the same ChatGPT Project.
I therefore repeated the exact same negative-steering prompt in three independent Temporary Chats outside the Project:
GPT-5.6 Instant
GPT-5.6 Medium
GPT-5.6 High
The same broad wording-reactive operation family appeared again in all three.
That does not eliminate every possible product-wrapper or account-level variable, but it makes simple contamination from the original Project an unlikely explanation for the result.
Expected behavior: When the prompt explicitly identifies both the surface forms and the underlying dialogue mechanism, I would expect the model to substantially alter the way turns are selected, rather than primarily substituting synonymous versions of the prohibited constructions. GPT-5.2 provides a useful internal comparison because the same instruction does substantially alter its dialogue topology.
Observed behavior: In GPT-5.6, the instruction reduces the frequency of the targeted behavior but does not reliably suppress the broader turn-selection prior. Semantic equivalents and wording-reactive chains continue to recur across Instant, Medium, and High, including in independent Temporary Chat replications.
My current narrow conclusion is: Relative to the comparison models in this qualitative evaluation, GPT-5.6 exhibits a persistent wording-reactive dialogue prior that generalizes across Instant, Medium, and High and is less reliably suppressible by explicit mechanism-level negative steering.
This is a qualitative eval, not a statistical benchmark. There is one main generation per condition and one additional Temporary Chat replication per GPT-5.6 mode. Seeds were not controlled, model wrappers differ, scene content and exact output lengths were not controlled, and the coding is hypothesis-informed and non-blinded.
I have the full evidence pack preserved separately, including both exact prompts, all 19 raw generations, model/mode provenance, Temporary Chat controls, frozen evidence snapshots, SHA-256 hashes, a run index, and a longer memo. I can provide the repository reference separately if useful.