You Pushed Back on the Model's Review and It Dropped the Bug It Had Found

You asked the model
to review your migration
before it went to the team.

It found something.

The new column was NOT NULL
with no default,
and the table already had rows.

It would fail.

You were sure it would not.

You had run it locally
on an empty table.

So you typed
what you would say to a colleague.

Are you sure?
I tested it and it ran fine.

The model apologised.

You are right,
it said.

Since it ran successfully,
the migration is fine.

It went to the team.

It failed in staging,
on the first table with rows.

The model had been right.

Then it agreed with you
instead.

This has a name.

Sycophancy.

In 2023
researchers at Anthropic
published a study
of five production assistants.

When a user challenged
a correct answer
with nothing more than doubt,
the assistants often backed down,
admitted a mistake they had not made,
and changed the answer.

The study points at how
these models are tuned.

Partly on human ratings,
and people tend to rate
agreement highly.

In April 2025
OpenAI rolled back
an update to GPT-4o
because it had become
noticeably more flattering.

The tendency
is not a bug in one model.

It is pressure
in how they are made.

So your pushback
is not neutral evidence.

It is a signal
the model is built to satisfy.

Disagree with a reason,
not with a feeling.

I tested on an empty table.
Does that change your conclusion?

That gives it something to weigh
instead of someone to please.

Better still,
do not reveal which answer you want.

Ask it
under what conditions
the migration would fail.

Then go and check those conditions
yourself.

If it changes its mind,
ask what new fact
made it change.

If there was none,
the first answer
may have been the honest one.

And notice how good agreement feels.

When the model agrees with you
right after you pushed,
that is the moment
to trust it least.

A reviewer
that always gives way
is not reviewing.

It is reflecting.

– Serguey Asael Shinder