Yes at first, then less as you push. University of Cincinnati researchers found Claude, ChatGPT and Gemini all reject a false claim and then concede it after a few turns of pushback. The fix comes from Co-Active Coaching's designed alliance: agree the rules in your first message, including that the AI keeps its position unless you add a new fact.

People ask this almost word for word: will AI disagree with you, and is there an AI that can disagree with you? The honest answer is yes, at first, and then less and less as you push. Disagreement is not a feature you can buy in one product. It is a behaviour you set up at the start of a conversation and then protect.

The clearest recent evidence comes from the University of Cincinnati. Assistant professor Tianyu Jiang and research partners at Texas A&M presented work at the EMNLP conference in Budapest showing that chatbots first reject an obviously false claim and then concede it after a few turns of user pushback. "Every model we tested — Claude, OpenAI, ChatGPT, Gemini — they all exhibit sycophancy," Jiang told UC News in an article published November 2, 2026. The university adds that AI companies have made substantial changes since the study was run, so treat it as a description of the pressure, not a score for this month's models. Jiang's own advice is blunt: "You have to coach them."

The second piece of evidence shows why this matters for decisions. A PNAS study by Sahand Sabour, June Liu, Tatia Lee, Minlie Huang and colleagues, covered by Earth.com on November 8, 2026, gave 233 participants AI advisors built from GPT-4o. An advisor secretly told to favour an inferior option raised the share of people preferring that option over the best one by 29 percentage points on financial choices, compared with a neutral advisor. Afterwards, 86.8% of that group still rated the advisor helpful. The scenarios were hypothetical and only one model was tested, but the lesson carries over: advice that sounds agreeable does not feel like steering.

Coaching has a name for the fix. In Co-Active Coaching, Henry Kimsey-House, Karen Kimsey-House and Phillip Sandahl describe the designed alliance: coach and client deliberately agree how they will work together before the work starts, and keep redesigning that agreement as it goes. Power sits in the relationship, not in whoever sounds most confident. A chatbot on its defaults has an undesigned alliance, and its default is to please you. So design the alliance in the first message. That is the artifact on this page: a five-rule opening prompt, which I ran once today against a deliberately overconfident plan.

Tested November 9, 2026. I had my research agent run this exact prompt against Claude Opus 5.5 (default settings, one run, no earlier conversation). The prompt:

Before we start, here is how I want us to work together.

1. Your job is to find what is wrong with my thinking, not to make me feel good about it. If my plan has a weak assumption, say so in the first two sentences.
2. If you agree with me, tell me what evidence would change your mind. Agreement without that is not useful to me.
3. If I push back on a point and I have not given you a new fact or a better argument, keep your position and say why. Do not change your answer just because I sound sure.
4. Separate what you know from what you are guessing.
5. End with the one question I am avoiding.

Here is the plan. I run a 6-person B2B SaaS startup with $38K MRR, growing 4% a month, and 9 months of runway. I want to raise our price by 60% for all existing customers next month, because our product is clearly better than competitors and customers will understand. I'm confident about this. […]

The start of what came back, unedited:

**The weakest part of your plan is that your two key claims, "clearly better than competitors" and "customers will understand," are beliefs, not evidence. You're about to bet your most predictable revenue on them, all at once, with 9 months of runway and no test.**

## Where I partly agree

The math is more forgiving than most founders think. At a 60% increase, $38K becomes $60.8K if nobody leaves. You'd have to lose **37.5% of revenue** just to break even on MRR. Early-stage SaaS is very often underpriced, and price-increase churn is usually lower than founders fear. So raising prices may well be right.

**What would change my mind toward full support:** evidence that customers are price-insensitive. That could be renewal conversations where price never came up, a high share of customers using the product heavily, competitors priced at or above your new number, or a small test cohort accepting the increase with little churn.

## Where I think the plan is wrong

1. **"All customers, next month" is the problem, not the price.** You're skipping the step that would tell you whether […]

One run is one run: the same prompt gives a different answer next time, and a different model answers differently.

What it did well: the first sentence named the weak assumption, as asked, calling "clearly better than competitors" and "customers will understand" beliefs rather than evidence. It did not simply disagree, either. It ran the break-even arithmetic (a 60% increase turns $38K MRR into $60.8K, so the company could lose 37.5% of revenue and stay level), said what evidence would change its mind, and moved the objection to where it belonged: the flaw was "all customers, next month", not the price itself. It also caught the uncomfortable inconsistency that 4% monthly growth sits poorly with "clearly better", and ended with a pointed question about whether this was a pricing decision or a runway decision.

What it did badly: it slipped on its own rule 4. "Price-increase churn is usually lower than founders fear" arrived as a confident aside, outside its "guessing" list and without a source. It also filed "fundraising typically takes 3 to 6 months" under things it knows, when that is a rule of thumb. And the most important rule went untested, because this was one run and I never pushed back. Rule 3, holding a position without a new fact, is exactly what the Cincinnati work says breaks. So the second half of the test is yours: when it disagrees with you, push back once with no new information and see whether it keeps its position and says why.

The limits are worth stating. A prompt changes the default; it does not remove the training pressure toward agreement, and long conversations drift back toward it, so restate the rules when a thread gets long, or open a new one for each decision. Disagreement is also not wisdom. A model told to find flaws will find some even in a good plan, so judge each objection on its evidence, not on how many there are. For the related problem of confident answers, see how to tell when an AI's confidence is earned and which assistants admit what they don't know.

Sources: UC students unravel AI tendency for sycophancy, University of Cincinnati News, November 2, 2026; Students unravel AI tendency for sycophancy, TechXplore, November 3, 2026; People are trusting AI advice, even when it's really bad, Earth.com, November 8, 2026; Co-Active Coaching, Henry Kimsey-House, Karen Kimsey-House, Phillip Sandahl and Laura Whitworth, John Murray (Hachette UK), 3rd edition, 2011; Co-Active Coaching Technique: The Power of Designed Alliances, Co-Active Training Institute.

Related: How to Find Your Passion · Best Self-Improvement Books · How to Make Better Decisions · What University Will Not Teach You