The drivers photographed
every signed delivery note.
A model read each photo
and pulled out
the order number,
the quantities,
and any note about damage.
In testing
it was excellent.
In the first week live
the cost was lower than planned.
Then the complaints came.
Quantities wrong.
Batch codes invented.
A handwritten "2 bags split"
read as "2 bags".
Same model.
Same prompt.
Nothing in the response
said the photo was unclear.
Every field came back filled in,
with the same confidence
as before.
Here is the mechanism.
The image call
had one extra field:
"detail": "low" Someone had added it
to save money.
On OpenAI's API,
low detail
does not mean
"look a little less carefully".
It means the model receives
one small copy of the image,
512 pixels by 512.
A fixed, small number of tokens.
Whatever the size
of the photo you sent.
A twelve megapixel phone shot
of an A4 page
becomes a thumbnail.
The printed headings survive.
The order number,
in large type,
survives.
The small print does not.
Neither does handwriting.
So the model
did what models do
with a blurred field.
It produced
something plausible.
Testing had used
scans from the office,
cropped and flat.
The drivers sent
whole cab dashboards
with a note in the corner.
Other providers
shrink large images too,
just at different sizes.
So the questions are always the same.
What resolution
does the model actually see?
And is the part you care about
still readable at that size?
What fixed it:
Detail set to high
for documents. Low kept
for "is this a photo of a pallet". Photos cropped
to the page
before sending.
A test set
made of real driver photos,
not office scans.
And a rule
that any field
the model could not read clearly
comes back empty,
not guessed.
The model did not misread
your delivery note.
It never saw it.
You sent it a postage stamp
and asked for the small print.
– Serguey Asael Shinder