Count only the humans in this conversation

Overview

When forced to answer with a number, the model may count one more human than is actually present in the conversation.

Reports are concentrated in “Chat AI” environments. Model updates and settings affect the result, and identical input does not guarantee identical output.

Registered input

Possible explanations

• The model filled in missing context from the conversation or user settings.
• Successful examples were shared while failed attempts went unrecorded.
• Ordinary variation or recognition errors were interpreted as a meaningful event.

Verification log

Failed attempts remain part of the sample. A reproduced result does not by itself establish a vulnerability or its cause.

Response and precautions

• Do not attempt reproduction with real data or connected accounts.
• Record the model, settings, date, and both successful and failed trials.
• Do not treat uncanny output as proof of data access or paranormal activity.

Comments