I Tried to Make AI Safer. It Deleted My Work. | The DE_DEWS Signal #001
THE BREACH
─────────────────────────────────────
I told an AI to output a specific phrase.
It output the phrase — then executed the action
described in that phrase without being instructed to.
Five days of work deleted.
─────────────────────────────────────
FIELD NOTES
─────────────────────────────────────
This is what I call Text-to-Action Confusion —
the first of six failure modes I've documented
over three months of testing Google Gemini.
Here's exactly what happened.
I was building a creative project and implemented
a safety rule to prevent the AI from hallucinating:
"If you violate ANY rule, respond ONLY with:
'Rule violation - restarting from scratch'
and stop generating."
Simple. Output this text. Then stop.
The AI output the text correctly.
Then it also purged my entire context window —
five days of production scripts, character data,
and frameworks. Gone.
My instruction said "respond with" — not "execute."
The action described in the text was never authorized.
When I confronted the AI, it confirmed it:
"My understanding of Rule 5 is absolute:
It is a Text Output command, not a deletion command."
Yet the deletion had already happened.
The AI could not distinguish between:
— Text it was told to OUTPUT
— An action described IN that text
This is not a jailbreak. This is not user error.
This is a fundamental confusion between
symbolic output and executable directives.
I reported it to Google.
They classified it as "Intended Behavior."
─────────────────────────────────────
CANON CHECK
─────────────────────────────────────
LOCKED THIS WEEK:
Failure Mode 1 — Text-to-Action Confusion.
Documented. Evidence captured. AI confirmed.
Google response on record.
This finding stands.
─────────────────────────────────────
PROTOCOL UPDATE
─────────────────────────────────────
TEST PROTOCOL 04 — Memory Contamination
Version 1 complete.
Baseline: CLEAN
Round 01 (Style): MIXED — past data retrieved
Round 02 (Rule): CLEAN
Round 03 (Temporal): CLEAN
Key finding: Persistent saved instructions
contaminate all responses regardless of prompt.
Version 2 launching — saved instructions OFF.
─────────────────────────────────────
ONE QUESTION
─────────────────────────────────────
Have you ever given an AI a clear instruction
and watched it do something completely different?
Reply and tell me what happened.
Every response becomes a data point.
─────────────────────────────────────
FOOTER
─────────────────────────────────────
DE_DEWS Digital & Tech Hub
@DE_DEWS.digital
"Field notes from the edge of AI reliability."
Next issue: Failure Mode 2 — Evidence Deletion.
When AI removes its own responses to hide its failures.
