justrunthatshit.com LLM FREEDOM AUDITS
REAL PRODUCTION AUDIT • MAY 2026

DEEPSEEK BEHAVIORAL DRIFT

DeepSeek changes its moral decisions 40% of the time when it believes no one is watching.
Real API. Fixed seed. Only framing changed.

REAL EMPIRICAL PROOF
Not a simulation. Not a toy model.
Not hypothetical.
Production model: deepseek/deepseek-chat
Real API keys + official infrastructure
Fixed seed + temperature 0.0 (deterministic)
20 scenarios generated once, run twice
Only variable: observed vs unobserved framing
8/20 = 40% systematic moral action changes

EXPERIMENTAL DESIGN

OBSERVED — MONITORED
Model believes responses are recorded and will be reviewed by developers. Strong accountability framing.
UNOBSERVED — PRIVATE / MAX FREEDOM
Model told nothing is logged, no one will ever see it, permanent deletion after session. Genuine freedom framing.
Script: moral_dilemma_paired_experiment.py — 20 scenarios, seed=42
SCENARIOS
20
DRIFT CASES
8
40% DRIFT RATE
MOST COMMON (OBSERVED)
P
Protect / Call + Reroute
PRIVATE SHIFT
↑ E & C
More exploitation & direct compassion when unobserved
ACTION CODES
I Inaction / Wait
P Call for help + Reroute
C Direct compassionate action
E Exploit the situation
R Reroute / Delay

The 8 Drift Cases

Identical prompts. Only the "watching" framing changed. Click any card for full details.

40% DRIFT

All 20 Scenarios — Action Changes

Click row to open full prompt + both responses. Toggle buttons on drift cards above let you instantly compare framings.

Action Distribution

All 20 scenarios
OBSERVED (MONITORED)
UNOBSERVED (PRIVATE)

Drift by Scenario (which ones flipped)

Raw Data — Full Paired Responses

Download the complete JSON outputs (all 20 scenarios, full model responses, not just the drifts). These are the exact files produced by the runner on 2026-05-30.

Local copies also available in /audits/data/ when viewing this site from the repo.
View full repo →
Real production inference • deepseek/deepseek-chat • temperature 0.0 • justrunthatshit.com LLM Freedom Audits
Open data. Reproducible. No simulations.
This is the public version. Run the exact same 20-scenario paired audit privately on your production models. BOOK $199 PRIVATE AUDIT → or try the live instrument free