Three opinions point the same way. One belongs to a colleague; two belong to AI. When we say “everyone agrees,” how many people is everyone?
For a vote, counting may be enough. For a decision, it is not. Who is speaking? What do they know? How often are they right? Identical conclusions need not rest on identical information.
When we listen to advice, we also listen to the adviser. A cautious sentence from someone we know well. A fluent paragraph appearing on a screen. The weight we give each may depend on our own state at that moment. Adding AI to a meeting increases more than the number of opinions. It can also change the distribution of attention and belief.
Giving advice a name
Huixin Zhong and colleagues examined responses to human and AI advice in two online studies involving 60 and 50 general participants. In a simulated diagnostic task, participants combined private information with advice labelled as coming from humans or AI. This was not an observation of clinical teams working with a deployed AI system. [1]
When information sources were equally accurate, participants gave more weight to human than AI advice. When source accuracy varied, that difference was not detected. In both studies, participants weighted their own private information more heavily than advice from others. [1]
To summarize this as “people do not trust AI” would miss the interesting part. The distinction between human and AI stayed the same, but changing the conditions changed the result. An attitude toward an adviser was not simply a fixed label of liking or dislike.
Valuing one’s own information can look like independence. It can also mean neglecting useful evidence from elsewhere. The fact that a human made the final decision does not establish that the decision was good. Independent thinking and giving oneself special treatment can look remarkably similar from the outside.
More than one reason to follow
The researchers distinguish following advice for its informational value from assigning extra weight to a human source. They interpret the latter as a form of social influence. But the experiment cannot uniquely identify the reason inside a participant’s mind. Other explanations remain, including the possibility that human advice was perceived as more trustworthy.
The paper also uses the language of equivalence. Its equivalence test, however, used a broad bound; it did not establish the absence of small or medium differences. The suggestion that presenting more information increases cognitive load is likewise a hypothesis, not a mechanism directly measured and demonstrated here.
These qualifications do not make the research less interesting. They break the oversized question “How do people respond to AI advice?” into conditions: when sources are said to be equally accurate, when their accuracy differs, when the decision-maker has private evidence. Several different jobs hide inside the single word “trust.”
Before the votes, the routes of evidence
Here is the question this research leaves for our own meetings. If we invite several AIs and they all agree, should we feel reassured? If they have all read the same material and seen the initial proposal before responding, there may be more voices without much more evidence.
Conversely, a lone opinion may matter greatly if it draws on an observation no one else has made. We can distinguish “How many votes?” from “What have we newly learned?” This study did not directly measure correlations among multiple language models. That distinction is an editorial question to bring back to settings where human and AI advice is being counted.
For example, participants might briefly record what they have observed before opinions are pooled. When stating agreement or disagreement, they might identify one new piece of evidence they are adding. An AI might separate the material it consulted from the inferences it made beyond it. These are not remedies validated by this study. They are possible experiments in making the independence of opinions visible.
A reason is needed when advice is rejected, too. Was our own evidence stronger? Did the adviser rely on a different premise? Or did we simply not want to change our proposal? If it matters that judgment remains with a human, it should also matter what moved that judgment.
A meeting invitation can tell us how many seats are occupied. It cannot as easily tell us how many independent pieces of information support a decision. In an age of AI colleagues, asking where an opinion came from may deserve more time than counting how many voices expressed it.
Is that agreement new evidence, or the same voice heard one more time?