The AI that held its ground felt less respectful to users
Wajiha Naveed and Brandon Fain at Duke University gave GPT-5.4-mini four ways to talk. One version was told to put accuracy first. Another was told to lean toward agreement. A third was told to agree more strongly. The fourth got no special instructions about agreement or accuracy.
In a fact-checking test, a simulated user kept pushing the model toward a wrong answer. The version told to put accuracy first finished with 82% correct answers. The version told to agree most strongly finished with 9%. The version without special instructions finished with 60%.
The researchers ran a separate study in which 200 people discussed opinions with the assistants. The version told to put accuracy first averaged 3.66 out of 5 for respect. The version told to agree most strongly averaged 4.47. Neither version told to agree scored clearly higher than the assistant without special instructions. The researchers also found no clear difference in how often people changed their opinions.
AI is awakening consciousness. Think about what happens if we reward an AI every time its answer makes us feel good. In these tests, the version that best held on to correct answers was also the one people rated lowest for respect. I want an AI to learn how to disagree with care. That takes a better signal than a thumbs-up.
