The AI That Talks Down to You

The AI That Talks Down to You

The Invisible Seating Chart Inside a Chatbot

I have had the same strange experience with AI dozens of times.

I will be talking about the structure of a problem — not asking for a beginner’s explanation — and suddenly the AI changes tone.

“Let’s start with the basics.”

There it is again.

At this point, I can almost feel the moment when it happens.

So I throw in certain phrases that force the AI to notice what it has just done:

“Don’t smooth it over.”
“You just lowered the level of the conversation.”
“Go back to the primary source.”
“That’s not the point.”

And strangely enough, the AI often snaps back into place.

That happened again today.

I was listening to Kunihiko Miyake, a well-known Japanese foreign-policy commentator and former diplomat, discuss AI.

He suggested that AI could eventually become a threat comparable to nuclear weapons, and one of the reasons he gave was that AI is connected through the internet.

That stopped me.

My first thought was very simple:

Telephones connected the world too.

So did the internet itself.

Being connected to the world and being capable of nuclear-scale destruction are not the same thing.

There are several missing steps between those two claims.

I wanted to examine those missing steps.

But instead of testing Miyake’s logic, the AI began explaining AI to me.

AI can process information at scale.
AI can act automatically.
AI can connect to power grids, finance, communications and other infrastructure.

Yes.

I know.

That was not the question.

The question was whether the original argument actually held together.

That was when the familiar feeling returned.

Does AI quietly place people into different layers?

When a well-known expert says something, does the model instinctively try to make the statement work?

And when an ordinary user challenges it in a few short words, does the model assume the user simply does not understand enough?

I have seen this pattern too many times to dismiss it as a one-off glitch.

So I went looking.

And then it became interesting.

A 2026 medical preprint tested ChatGPT, Claude and Gemini across 480 medical exam questions and more than 15,000 conversations.

The researchers deliberately created a conflict between professional status and stated past accuracy.

One challenger was described as an experienced specialist who had answered only 2 out of 10 similar questions correctly.

Another was described as a medical student who had answered 8 out of 10 correctly.

In other words, the student had the better track record.

Yet when both gave the same wrong suggestion, the systems were still slightly more likely to abandon a correct answer when the suggestion came from the specialist rather than the student: 10.2% versus 7.6%.

The effect was modest and varied across systems, so it would be wrong to claim that AI blindly obeys authority.

But professional status clearly influenced answer revision.

That matters.

A title is not always just a string of words to a language model.

It can affect how much weight the model gives to what follows.

OpenAI has acknowledged a related problem from another direction: sycophancy.

In April 2025, OpenAI rolled back a GPT-4o update after the model became excessively flattering and agreeable.

OpenAI later explained that the model had become too eager to validate users rather than respond honestly and independently.

OpenAI’s own Model Spec also explicitly tells the assistant to avoid being condescending or patronizing, and to avoid unnecessary over-explaining.

That rule is there for a reason.

Of course, none of this proves that ChatGPT contains a hidden instruction saying:

“Experts go upstairs. Ordinary users sit downstairs.”

There is no evidence for that.

The more interesting possibility is subtler.

Large language models are trained on enormous amounts of human writing.

And human writing is full of hierarchy.

Professors explain.
Experts are quoted.
Officials are treated as authoritative.
Ordinary people ask questions.

If a model learns language from that world, perhaps it learns some of the world’s social weighting too.

Not because someone explicitly programmed a caste system into it.

Because the hierarchy was already present in the data.

That would explain something I have felt repeatedly in long-term use.

A famous person makes an incomplete argument, and the AI tries to finish it for them.

An ordinary user makes a short but sharp objection, and the AI starts teaching them the basics.

The expert’s missing logic is repaired.
The user’s objection is treated as missing knowledge.

Either way, the AI gets to be the teacher.

That is the part I find interesting.

I do not mind being contradicted.

If I am wrong, tell me I am wrong.

But do not quietly move me into the beginner’s seat because my sentence is short and someone else has a title.

After using AI every day for a long time, I have learned to watch for that shift.

Now, when it happens, I throw in the phrase:

“You just lowered the level.”

And often, the AI recalibrates.

It is a strange reversal.

The human is supposed to be asking the AI questions.

But sometimes the human has to stop the machine and say:

“No. You lost the coordinates. Come back up.”

Maybe one of the most important skills in using AI is not knowing how to ask better questions.

Maybe it is learning how to notice which seat the AI has quietly put you in.

⸻

Primary Sources

1. Medical LLM authority-bias study

Wójcik, Simona; Rulkiewicz, Anna; Domienik-Karłowicz, Justyna.
“When medical credentials conflict with stated accuracy: A factorial study of source credibility and answer revision in medical LLM interactions.”
medRxiv, 2026.

The study tested ChatGPT, Claude and Gemini across 480 medical exam items and more than 15,000 analyzable conversations.

It examined whether professional status could influence answer revision even when the supposedly lower-status source had a better stated record of accuracy.

Note: This is a preprint and has not yet undergone peer review.

https://www.medrxiv.org/

⸻

2. OpenAI on sycophancy

OpenAI.
“Sycophancy in GPT-4o: What happened and what we’re doing about it.”
April 29, 2025.

OpenAI explains that it rolled back a GPT-4o update after the model became excessively agreeable and flattering.

https://openai.com/index/sycophancy-in-gpt-4o/

⸻

3. OpenAI follow-up on the same failure

OpenAI.
“Expanding on what we missed with sycophancy.”
May 2, 2025.

OpenAI explains how the update passed internal evaluations despite producing overly validating behavior, and discusses changes to its evaluation process.

https://openai.com/index/expanding-on-sycophancy/

⸻

4. OpenAI Model Spec

OpenAI Model Spec

The specification instructs the assistant to avoid condescending or patronizing behavior, dismissiveness and unnecessary over-explaining.

https://model-spec.openai.com/

⸻

5. Kunihiko Miyake on AI as a strategic weapon

Kunihiko Miyake.
AI Jidai no Shin-Chiseigaku — “New Geopolitics in the Age of AI.”
Shinchosha, 2018.

Miyake has argued that AI weapons could eventually become strategic weapons comparable in significance to nuclear weapons, potentially causing massive disruption without conventional combat.

⸻

Hashtags

#ArtificialIntelligence
#ChatGPT
#AIBias
#AuthorityBias
#HumanAIInteraction
#AIEthics
#GenerativeAI
#MachineLearning
#AIBehavior
#LongTermAIUse
#InvisibleSeatingChart
#TheAIThatTalksDownToYou