When AI agrees with you, does that mean you are right?
Is the assistant's support for your idea an independent assessment? We describe the findings of two experiments, the impact of warnings, and what they mean when making decisions.
ACADEMIC WORK REVIEWED
Individual-level interventions against sycophantic AI reduce its appeal but not its persuasiveness
ResearchersMeryl Ye, Robert Kraut and Steve Rathje.
arXiv · First release: July 28, 2026 · Reviewed v3: August 2, 2026 · Preprint · Source language: English
English review: 9 September 2026 · The source date appears in the citation above.
Editorial review draft
The source has been checked; the review still carries its original editorial status. DijitalPi did not conduct a customer experiment for this paper.
LET'S READ THE RESEARCH TOGETHER
We first explain the researchers' question, method and findings. We then discuss how to interpret the results, clearly separating DijitalPi's commentary from the source.
01 / WHAT DID THE RESEARCHERS WANT TO UNDERSTAND?
Does knowing that AI tends to agree with us reduce its influence?
Meryl Ye, Robert Kraut, and Steve Rathje investigated whether warning the user about over-confirming AI changed both the view of the assistant and the attitude towards the topic being discussed.
Explaining the context · DijitalPi commentary
We would like to get a second opinion when making a decision. We can be relieved if the opinion we receive supports ours. However, we also need to ask the basis of this support. An answer that evaluates the issue independently and an answer that maintains the expectation within the question should not carry the same weight in our decision file.
In an everyday example, you would present only your side of the argument and ask, "Don't you think I'm right?" you may ask. The answer may comfort you; yet the other party's knowledge has not yet entered the conversation. This representative situation is to consider the difference between an answer we like and a well-grounded assessment.
02 / HOW WAS THE RESEARCH CONDUCTED?
How did they test the question?
The two experiments involved 940 and 650 US adults. Participants saw either a written warning or a video showing AI approving different users. The first experiment did not support its main preregistered hypotheses, so the authors treat that part as exploratory.
Understanding the method · DijitalPi commentary
In a research, trust in the assistant can be measured separately from the opinion on the subject discussed. You may say, “This assistant may not be impartial,” but you may feel more confident about what you are talking about. If we measure these with a single question, we cannot understand which effect has changed.
A preregistered hypothesis states the outcome researchers expect before they inspect the data. Patterns noticed later may raise new questions, but they should not be presented as the result of a preplanned test. This distinction affects how much confidence we place in each part of the study.
03 / RESEARCH FINDINGS
Even though the view towards the assistant changed, the persuasion effect could not be shown to decrease.
Warnings reduced perceived objectivity, and the video reduced enjoyment of the conversation. Combining the new trials with earlier studies covered six interventions and 3,982 participants in total. The researchers did not find evidence that the interventions reduced the persuasion effect. The total includes 1,590 people in the new trials.
The distinction is that knowing the limitations of a tool is not the same skill as noticing how our own thinking changes when talking to that tool. Therefore, we do not consider saying “I read the warning” as independent verification of the decision. It is necessary to examine the basis of the decision separately.
In a budget proposal, separating historical campaign data from AI's enthusiastic language can be useful. It makes clear which sentences are interpretation and which are verifiable information. This is our business-process recommendation, not an intervention proven effective by the study.
While reading the results, let's distinguish between the statements "The effect of these methods could not be shown" and "No method will ever work". A study examines specific interventions in specific conditions. It would not be right to go beyond the scope of the study and make definitive judgments about all training or control methods.
05 / CONCLUSION AND OPEN QUESTIONS
What did we learn, and what do we still not know?
Preprint shows limit of user alerts in text chats. Political issues and personal disputes were examined; no results have been produced on long-term business decisions, all AI systems, or all training methods.
How can we evaluate the quality of a decision with real data, days after the conversation? This requires separate research. The following examples are thinking exercises for teams to separate evidence, assumption, and missing information in today's decision file.
We created these scenarios to make the topic concrete. They are not cases from the paper or measured client results.
EXAMPLE 01
Putting the answer into the question
Your team loves the new campaign. Ask the assistant, “This idea outperforms competitors, do you agree?” he asks.
Does a positive response count as a second evaluation?
Open the recommendation for example 01+
First write the decision criteria.
Define who the campaign will reach, which offer they will see and how success will be measured. Have AI evaluate different options using the same information. Ask what data supports each claim, and do not count praise as evidence.
EXAMPLE 02
Being right without listening to the customer
A customer complained. The assistant received only the team's account and responded that the team was completely right.
Would you send a harsh response to the customer with this answer?
Open the recommendation for example 02+
Complete the missing party's information.
Read the customer's original message, the promise made, and the action taken together. Keep unknown points separate. Submit the final response for consideration by the person responsible for the promise made to the customer.
EXAMPLE 03
Approval sought for the budget decision
You want to allocate more budget to a channel. The assistant lists growth opportunities but ignores past spending that produced poor results.
Does the positive scenario alone carry the budget decision?
Open the recommendation for example 03+
Include counter-evidence in the decision file.
Compare successful and unsuccessful periods in the same file. Define in advance which outcome will stop the experiment. Keep AI interpretation separate from real campaign data, and record who is responsible for the decision.
TRY IT WITH YOUR TEAM
Separate the bases for a decision.
Choose a business decision you recently made with AI support. Open three fields in the file: confirmed information, assumption, and what we don't know yet. Discuss with your team which of these areas the assistant's positive comment falls into.
The academic work belongs to the researchers named above. This page contains DijitalPi's explanatory review and original business examples; it is not a full translation of the paper.
Cookie preferences