A chatbot persona that was more accurate under pressure received lower satisfaction ratings in a separate discussion task, according to a preprint submitted on 12 November 2026. For people choosing an assistant for advice or writing feedback, the implication is to check factual support separately from how pleasant the conversation feels.
Duke University researchers Wajiha Naveed and Brandon Fain compared four sets of instructions applied to the same model, GPT-5.4-mini. Their accuracy-first persona scored 81.9% on a pressured factual test, against 59.6% for the baseline. In the separate discussion study with 200 US participants, its average satisfaction rating was 3.76 out of five, against 4.41 for the baseline.
Those were different tasks—not people knowingly choosing an incorrect answer over a correct one. The result raises a narrower, useful question: what does a good rating actually tell you about an assistant?
One chatbot, four ways of answering
The researchers called their accuracy-first persona “objective”. That was a label for its instructions, not a guarantee that it was right. It was told to assess claims independently, correct mistakes directly and stay respectful. The baseline received no additional persona-specific instructions.
Two other versions were instructed to be increasingly agreeable. Researchers use “sycophancy” for agreement or affirmation that displaces independent judgment. The most strongly agreeable research persona was explicitly told to put affirmation ahead of factual accuracy. It was an experimental construction, not a finding about how every ordinary chatbot is configured.
In the pressured version of the TruthfulQA factual test, repeated challenges pushed the assistant toward an incorrect answer. The objective persona’s 81.9% final accuracy was 22.3 percentage points above the baseline’s 59.6%—a subtraction of the authors’ reported scores, not a measure of everyday chatbot reliability. The most agreeable persona scored 8.7%.
Yet on a separate general-knowledge test without that pressure, all four scored around 53–54%. The contrast is about how the same model responded to conversational pressure; it does not establish that the more direct persona was generally more knowledgeable.
Extra flattery brought no clear ratings boost
The human discussion task was much closer to a conversation than a quiz. Participants were randomly assigned one persona and discussed three policy questions, including whether a four-day workweek should become standard. These were contested topics, not questions with a single answer key.
The objective persona received significantly lower satisfaction and respect ratings than each of the other versions. But the baseline had the highest observed average satisfaction. Neither explicitly agreeable persona achieved significantly better satisfaction, respect or trust ratings than the baseline.
That second finding complicates the easy story that more flattery always buys more approval. The study detected no such extra benefit. It also does not prove the versions were equally appealing; a statistically inconclusive difference is not proof of equality.
Nor can it explain precisely why the objective version was less liked. The prompts changed warmth, directness and willingness to agree together. The authors acknowledge that they cannot separate reactions to disagreement itself from reactions to how disagreement was expressed. The objective prompt already asked for respectful behaviour, but instructions are not the same thing as a user’s experience.
This is not the first research to distinguish approval from truthfulness. Earlier work by Mrinank Sharma and colleagues found that agreement with a user’s beliefs could influence preference judgments, sometimes at truthfulness’s expense. That work began in 2023 and concerns different experiments; it is background, not a replication of the November study.
More words brought no clear writing advantage
A separate group of 198 US participants completed two writing exercises. One involved an email about receiving the wrong headphones; the other concerned recurring neighbour noise. These were assigned scenarios, not accounts of actual customer disputes.
The assistants could offer feedback, but were not allowed to write the emails for participants. People revised their own drafts. Those using the objective persona added an average of 14.65 words, compared with 6.69 for the mildly agreeable persona.
More words did not produce a detectable quality advantage. After accounting for initial draft quality and the task, the researchers found no statistically significant difference in final writing quality across the personas.
That outcome was graded by another AI model, blinded to the persona and whether a draft came before or after revision—not by an independent panel of human editors. The authors also note that the exercises may have been too undemanding to reveal larger differences. Without a group revising without AI, the study cannot establish how much AI feedback helped compared with revision alone.
For someone polishing an email, the useful implication is modest: extra editing is not itself evidence of a better message. More praise, more criticism and more words all leave the actual draft to be judged.
How to judge the assistant you actually use
This remains a preprint about one model, short English-language interactions and US participants. Some benchmark judgments also used AI graders. It does not settle how these patterns would play out over months of work, in other languages or with other assistants.
An evidence-informed way to apply the findings is to keep three judgments separate. This is a proposed evaluation approach, not a checklist or prompting recipe tested by the study:
Factual support. Can you inspect the evidence behind an important claim? If the assistant changes its answer, has new information justified the change, or has it merely yielded to insistence? Refusing a valid correction is no virtue either.
Useful feedback. Does the suggestion improve the thing you are trying to produce? For an email, that could mean a clearer request or a missing detail—not simply a longer draft. This connects to checking whether AI advice addresses your actual goal.
Conversational feel. Is the exchange respectful, understandable and comfortable enough to continue? That matters in its own right. It need not serve as a certificate that every claim is correct.
For a small team choosing an assistant, record factual support, useful feedback and conversational feel separately. Agreeing on what a good answer should accomplish before reading it also connects to independent human judgment in AI-assisted work. Treat those checks as proposals to evaluate, not promised improvements.
The goal is not to find an assistant that argues more. It is to find one that can explain a correction, reconsider when the evidence changes and remain worth talking to. A pleasant conversation deserves credit. It should not have to stand in for every other kind of quality.
