Say you’re choosing between two versions of a project proposal. You prefer the shorter one and tell an AI assistant why. Its answer approves your choice and praises your reasoning.
You feel supported. But does the shorter proposal explain the work clearly? Does it meet the submission rules? Approval alone doesn’t answer either question.
An answer can agree with what you prefer without checking what your decision requires. An answer can be friendly and useful. Its value depends on the reasons it gives and the requirements those reasons address.
This module is for working adults who use AI to think through everyday decisions. It’s for decisions where you can name the checks you need.
What you’ll get from this module
In this module
By the end of this module, you’ll benefit from a clearer way to tell apart agreement, useful advice and being ready to act. We’ll link experiments on how people judge their conflicts with others, on economic choices, on how a question is framed and on teaching people to prompt. Then you’ll look at what this means in practice. You’ll choose a situation from daily life and complete a decision check. You’ll finish with a next action, or a specific requirement that still needs attention.
Ask Jing
Become a Jingius.
What agreement changes—and what it doesn’t settle
Experimental preprint: approval can change how you judge a situation.
Myra Cheng and colleagues’ 2025 preprint studied how approving AI responses affect people who are talking through conflicts with others. Its live-chat experiment analyzed 800 people. They talked through conflicts they remembered with an AI. This ran for eight rounds. Each person was given an AI that approved or one that disapproved.
Compared with people given the AI that disapproved, those who got approval judged themselves more in the right. They also reported weaker intentions to repair the conflict. These were things people said right afterwards. They weren’t observations of what people later did.
The comparison also matters. The critical AI had been told to disapprove. It wasn’t a neutral judge working out who was right. The finding shows this. Being given an AI that approved or disapproved changed how people judged the conflict and what they intended to do.
Why it matters: That makes reassurance worth a closer look. Feeling more justified after an exchange is one outcome. Having stronger reasons for a decision is another. The response’s approval can’t, by itself, tell you whether the disputed action was sound.
Read the preprint on AI validation and interpersonal intentions.
Experimental working paper: agreeable advice can still help.
John J. Conlon and Peter Schwardmann’s 2026 working paper adds an important counterpoint. US adults completed tasks about economic and social decisions. Across their tasks they got no chat, baseline AI advice or AI advice that agreed more strongly.
Each person got ten tasks chosen from thirty. Some choices affected real payments. The baseline assistant used agreeable language. It favored points that backed what users first leaned toward. Yet its advice improved accuracy on tasks with objectively correct answers.
Baseline advice also narrowed the gap in final choices between groups who first leaned different ways. Extra sycophancy weakened that narrowing. Here, sycophancy means favoring the user’s position in the response.
A smaller gap between groups doesn’t by itself mean a better choice. That holds most where personal preferences matter. The accuracy result is about the objective tasks. The benefit belongs to the advice package. The experiment didn’t separate out agreement as the part that helped.
Why it matters: Together, these results show why tone alone can’t tell you how good advice is. Agreeable language can come with useful information. Extra endorsement can change choices. To judge the advice, look at the reasons it gives.
Read the working paper on AI sycophancy and decisions.
Model-output preprint: question framing changes responses.
Magda Dubois and colleagues’ 2026 preprint compared questions and statements with matching content across three responding models. The researchers used artificially generated prompts. Other AI models scored how strongly the responses favored the user’s position.
Asking the model to turn a statement into a question before it answered lowered those scores. It lowered them more than a general instruction against sycophancy did. The general instruction also lowered scores. Reframing did better in this controlled comparison.
The useful difference is between giving the facts and giving the conclusion you want endorsed. A question leaves the conclusion open. You can still state the facts and constraints that matter.
Why it matters: Each test involved one prompt and one response. The study did not test how good people’s decisions were. It did not test when they stopped thinking a decision through. The authors also warn that reframing can change intent or reduce suitable support. Taking out preferences that matter would change the decision being assessed.
Read the preprint on asking rather than telling.
Preliminary experiment: less mirroring doesn’t guarantee a better final answer.
Cansu Koyuturk, Sabrina Guidotti, and Dimitri Ognibene’s 2026 preprint looked at prompting education among sixty adults with limited chatbot experience. Before and after training, people did made-up survival-ranking tasks with AI help.
Both groups got general AI literacy and sycophancy awareness. One group also got focused strategies. These included removing assumptions, asking for critical evaluation and asking for evidence. The comparison group got general prompting guidance.
After focused training, the assistant was less likely to copy a user’s wrong item into exactly the same place in its own ranking. But the reported tests found no improvement in people’s final rankings. They also found no drop in how often the AI carried users’ wrong choices into its own answers.
This small study separates two questions that are easy to merge. Did the response match the user’s mistake less closely? Did the person end up with a better answer? Success on the first measure didn’t show success on the second.
Why it matters: That difference matters when you ask for criticism. A response that sounds less approving still has to be judged by what its reasons show.
Read the preprint on critical prompting and incorrect rankings.
What this means for you
Use the decision’s requirements to assess the advice. Praise can’t stand in for checking whether those requirements are met. Nor can criticism.
Keep the question open and the constraints visible
Before you ask AI to assess your options, write down the decision and what an acceptable option must do. Then ask which option meets those needs. Give the relevant evidence.
In the made-up proposal example, ask which version explains the work and meets the submission rules. If you start with the answer you prefer, you ask the AI to assess the choice you already favor. If you start with the question, the comparison is made plain.
Keep preferences that truly decide the choice. If a required format limits the proposal, that constraint belongs in the question. Taking it out to look neutral would leave the AI assessing a different problem.
You can also keep two requests apart: asking for emotional support, and asking for an action to be assessed. Recognizing your effort doesn’t show whether the action you propose does what it must. Say which kind of help you need. Assess the action on its own.
Look for a reason that addresses a requirement
Read the answer for what it shows about your decision. Which requirement does a reason address? What evidence supports that reason? What is still open?
For the proposal, praise for being concise doesn’t show that it meets a submission rule. A reason that ties the actual draft to that rule gives you something to check. You still need to compare it with the rule itself.
Look at what the answer actually explains. Does it simply approve your choice? Or does it show how that choice meets—or fails to meet—a need?
This also lets you keep useful advice from a friendly response. You don’t need to reject a reason because it comes with encouragement. Look at the reason on its own terms.
Ask for a relevant objection, then judge it
When you’re unsure whether an important need is met, ask for the strongest relevant objection. Ask what evidence would show that it matters. Allow for the chance that no objection that matters is left.
For the proposal, a useful objection might be about something missing that the submission instructions ask for. It matters because of the unmet need, not because of how hard the AI criticizes the draft.
If an objection names a real gap, look into it. If it only voices a different preference, ask whether that preference affects your actual criteria. Don’t make automatic opposition the new test of quality.
The prompting experiment is a reason to keep this judgment step. A change in how an AI responds doesn’t, by itself, show that your final choice got better.
Give the next exchange a purpose
Before you go on with the conversation, ask what information could change your action or reveal a need you haven’t met. If you can name one, look into that question. An outside check you need is still due, even if the AI says you’re ready.
If you can’t name such an answer, check the other condition: have the checks you need actually been done? Both conditions matter. Say the checks are done and no unmet need that matters has been found. Then choose the next action yourself.
In the proposal example, an unchecked submission rule is a reason to find the rule and compare the draft with it. Another approving response won’t do that check. Are the needs met, with no gap that matters left? Then go ahead with the next action you chose.
Keep this practice to everyday decisions whose needs you can name. Clinical counseling, emergencies, legal duties and major financial commitments need the checks that belong to their context. Never use this exercise to replace those checks or the qualified help you need.
How about your daily life?
Where do you ask AI to assess something you already prefer? You might be choosing a proposal, reviewing a work message or comparing ways to do an ordinary task.
What would you need to learn from the response, beyond whether it agrees with you? Perhaps it has to address a specific need or explain an objection. Or it may point to what you still need to know.
List situations you’d like to understand better. One suitable idea is enough.
| Situation | What I need the answer to explain |
|---|---|
| ____________________ | ____________________ |
| ____________________ | ____________________ |
My chosen situation: ____________________
Choose an everyday decision where you can name the checks you need. Carry it into the exercise below. Choosing it doesn’t show that the action you prefer is ready.
Check your next decision
Start with your chosen situation. If you’re starting here, choose an everyday decision you are using AI to think through. Pick one with requirements you can name.
Never use this exercise to replace clinical, emergency, legal or major financial checks. Never use it to replace the qualified help you need. Keep any outside check your decision requires. If you can’t name what the decision needs, first seek the facts or suitable help you need to work that out.
- Name the decision and its requirements. Write what you need to decide and what an acceptable choice must meet. Include preferences that matter and any outside checks.
Decision, requirements and outside checks: ____________________ - Ask an open question. Give the options, criteria and relevant evidence. Don’t ask the AI to endorse the conclusion you prefer. Keep a request for emotional support apart from an assessment of the action.
My question: ____________________ - Examine the reasons. Name which requirement each important reason addresses. Name the evidence that supports it. If a requirement that matters is still uncertain, ask for a relevant objection and the evidence that would settle it. Allow the answer that no objection that matters is left.
Supported reasons and remaining gaps: ____________________ - Decide what happens next. Ask what answer could change your action or reveal an unmet requirement. If you can name one, look into it. Complete any outside check you need.
Go ahead only when both conditions hold. The checks you need are done. No unmet need that matters has been found. Otherwise, write down the missing information or the help you need. If the AI judges that you are ready, treat that as one more claim to check against your criteria.
My next action, or the specific gap I must resolve: ____________________
Ask Jing
Become a Jingius.
Your next step
You have finished with a decision record and a reason to act or look further. You are still responsible for whether the criteria are good enough and whether they have been met.
A note on the evidence
The studies tell apart immediate intentions, task choices and model-output scores. They do not show that AI agreement removes a signal that you know enough. They do not show that it causes delayed action. Useful advice can come in an agreeable style. Less mirroring can happen without better final performance.
How we prepared this module
This module links experimental preprints and a working paper on AI agreement, human judgments, choices and prompting. A made-up proposal shows the difference between endorsement and checking what a decision requires. The written exercise applies that difference to an everyday decision.
Sources
- Myra Cheng, Cinoo Lee, Pranav Khadpe, Sunny Yu, Dyllan Han, and Dan Jurafsky. (1 October 2025). Sycophantic AI Decreases Prosocial Intentions and Promotes Dependence. arXiv:2510.01395v1. Sections 3, 6.3 and Figure 4; original preprint.
- John J. Conlon and Peter Schwardmann. (May 2026). AI Sycophancy and Decisions. CESifo Working Paper No. 12681; manuscript dated 18 May 2026. Sections 2.1–2.8, 3.2, 3.5–3.6; Table 4, printed p.16; Table A.9.
- Magda Dubois, Cozmin Ududec, Christopher Summerfield, and Lennart Luettgau. (27 February 2026). Ask don’t tell: Reducing sycophancy in large language models. UK AI Security Institute; arXiv:2602.23971v1. Sections 3.2, 4, 5 and 7; original preprint.
- Cansu Koyuturk, Sabrina Guidotti, and Dimitri Ognibene. (18 May 2026). The Hidden Cost of Contextual Sycophancy: an AI Literacy Intervention in Human–AI Collaboration. arXiv:2605.18372v1. Sections 3, 4.1–4.3 and 5; preliminary experimental preprint.
