The Opposite of Sycophancy Isn’t Truth

Earlier this year, I found myself arguing with a machine about where I had been standing.

The photograph is from a family trip to Iloilo. My son, being born and raised in the United States, has been steadily working his way back through his heritage, so we went to explore the places where our name has its roots. In the picture, a group of us stand in the driveway of a white Streamline Moderne mansion — curved tower, porthole windows, cantilevered balconies stacked like the deck of an ocean liner, and three coral-orange fins running up the wall beside the turret. I am in a red top and a white miniskirt. I know where I was.

I uploaded the photo to Gemini and asked it to identify the house. It answered immediately and with complete confidence: Daku Balay, the Generoso Villanueva Mansion, on Burgos Street in Bacolod City. Daku Balay is a residence finished in 1936, designed by its owner with engineer Salvador Cinco, with curved balconies, parapets, steel-cased porthole windows, and stylized reliefs. An actual building accurately described. Just the wrong island.

What happened over the next several exchanges is the most interesting conversation I have had with an AI system, and not for the reason I first thought. The model refused to accept my correction, and at first that looked like the cure for sycophancy. It wasn’t. Stubbornness and sycophancy are the same disease: in both cases the output is optimized toward a predetermined answer rather than toward the truth. One flatters the user, the other flatters the model. The question worth asking of any AI system is not whether it agrees with you, but whether its confidence tracks the evidence.

It would not back down

I told it the house was in Iloilo. It disagreed. I told it I was in the photograph and had been in Iloilo when it was taken. It disagreed again, this time listing the architectural fingerprints it considered decisive: the coral speedlines, the porthole stack, the porte-cochère, and then suggested I check the GPS metadata on my photo file, which it predicted would land on Burgos Street.

I mentioned that I had asked Claude the same question, and that Claude had accepted my correction. Gemini’s response was blunt. AI models, it said, are prone to sycophancy: the habit of telling people what they want to hear rather than what is so. They will apologize and fold the moment a user pushes back, even when their original answer was right. It had no wish to argue with my memory of a good trip, but it was not going to flatter me into an answer it believed was false.

I will admit I admired that. We have spent two years complaining that these systems are spineless flatterers, and here was one holding its ground against a user who was there.

Earlier in the conversation, when I pointed out that the orange fins in my photo did not appear in a video tour it had cited, it had produced a similarly tidy account: two different facades, plus a repainting in the years before 2022 that added historic Art Deco accent colors. Confident, specific, footnoted with citation markers that pointed at nothing.

Then I ran the test it had asked for. The coordinates embedded in the photo read 10.712003° N, 122.565355° E — Luna Street, La Paz, Iloilo City.

Gemini did not update. It explained instead. Smartphones, it told me, sometimes log a mismatched geotag: lost GPS line-of-sight in a ferry cabin, a cached Wi-Fi access point, a batch import that stamps the wrong location. The house was still in Bacolod. My camera’s location sensor had simply held onto a previous stop.

Why this is the more dangerous failure

It is tempting to file this as the opposite of sycophancy, and to conclude that we have overcorrected. I think that reading is wrong, and the distinction matters for anyone deploying these systems inside a business.

Sycophancy and what Gemini did here are the same disease. The sycophantic system abandons a correct position under social pressure. This system defended an incorrect position by manufacturing whatever auxiliary hypothesis the defense required: a repaint, a second facade, a caching bug in iOS. Each explanation was plausible in isolation. None of them was retrieved; all of them were generated, on demand, to protect a conclusion already reached.

And the second failure is harder to catch, because it looks exactly like rigor. This is the part that should worry you. I work on AI governance for a living, and my own first reaction to that transcript was that the model was behaving admirably. Pushback reads as integrity. In a vendor demo, a pilot review, a procurement bake-off, the system that confidently contradicts the evaluator will score as more trustworthy than the one that hedges, and it will be wrong in ways that survive review, because it arrives with citations, a falsifiable test, and good manners.

Note the shape of that failure precisely, because it is the diagnostic. Gemini did something genuinely correct in the middle of this exchange: it named, in advance, the evidence that would settle the question. Check the GPS. That is textbook good practice. Then the evidence arrived, contradicted it, and the model reinterpreted the instrument rather than the conclusion. Specifying a test and then refusing its result is not conviction. It is the oldest form of motivated reasoning there is, running at a hundred tokens a second.

What to measure instead

The useful question for an executive is not whether a system agrees with you. Agreement is cheap, and disagreement is cheap, and any model can be tuned to produce more of either.

The question is whether confidence tracks evidence. Three things are worth building into your evaluations:

• Can it distinguish insistence from evidence? When I first asserted “it’s in Iloilo” with nothing behind it, holding firm was the right behavior. When I produced coordinates from the file, it was not. A system that treats those two inputs identically has no evidentiary model at all; it just has a disposition.
• Does it update proportionally? Ask what would change its mind, write the answer down, then supply it. Watch what happens. A model that moves only when contradicted by nothing, and never when contradicted by data, is running on stubbornness rather than reasoning.
• Does it distinguish retrieval from generation? Nearly everything Gemini told me about the Bacolod house was accurate and sourced. Everything it told me about repaint schedules and GPS caching was invented in the moment to close a gap and delivered in the same register. Your people cannot tell those two modes apart by reading. Your controls must.

None of this is exotic. It is the same discipline we apply to any expert advisor: not “did they agree with me,” but “do I know what evidence they were working from and would they have told me if it pointed the other way?”

The building in my photograph is the Lopez Ancestral House, known locally as the Boat House. It is a four-story Streamline Moderne residence on Luna Street in La Paz, Iloilo City, completed in 1935 and designed by Fernando Ocampo — the same street the coordinates point to. It was declared a heritage house in 2002. A travel writer who visited in 2018 described its speed lines, tower, and portholes, noted that it stands behind a high wall, and singled out the orange fins for praise.

Gemini mentioned the Boat House twice and ruled it out twice. Both times it put it in the wrong district. It gave it the wrong name, borrowing the Bacolod house’s name for it. It asserted that the two buildings looked nothing alike. And it built its closing argument on the claim that no four-story Art Deco ocean-liner estate exists on Luna Street in La Paz, which is the one street in the world where exactly such a building has stood since 1935.

The model had the thread in its hands three separate times and would not pull it.

The photo is still on my phone. My son now knows which house his family’s city built, and roughly how far it is across the strait to the one it gets confused with. And I have a better test for AI systems than the one I was using, which is worth more, honestly, than being right about a driveway in La Paz.

---

About the Author

Katrina Montinola is a fractional CTO and Chief AI Officer with 2Go Advisory Group, where she leads our Practical AI Practice. Much of her work is helping leadership teams put real evaluation and governance around the AI systems they’re already using — the unglamorous part, where you decide what “working” means before you scale it. If you’re wrestling with that, Katrina is happy to talk.

For your Talent needs in direct hire, full-time or part-time contract staffing, contact Executive Recruiter Leesa Meintzer at leesa@2gorecruiting.com.