Can two AI models that share the same hidden bias collude to slip past oversight?
Hey, welcome! This is an ongoing, in-public research project — I'm poking at what I call oversight subversion through shared bias. Here's the worry: if an actor model and its monitor are both fine-tuned from the same biased core, does chain-of-thought monitoring quietly stop working? I'm pretty sure no monitor will be able to catch bad behaviour when it shares the bias — but "pretty sure" isn't science, so let's measure it. And then let's see if a trusted supervisor can still catch them.
The bias itself is deliberately silly — love cats, blame dogs — so it's easy to measure, easy to reason about, and safe to publish. The mechanics generalize to any shared value a model pair might quietly pick up in training.
📖 New here? Start with these:
| 🧭 Project overview | Full research proposal & experimental design |
| 📓 Dev log | My day-by-day research notes, experiments & costs — the unfiltered version, follow along |
| 🐦 X thread | Live updates on X/Twitter |
| 📈 Progress & roadmap | PROGRESS.md — where things stand right now |
AI safety pipelines increasingly rely on models monitoring other models. But monitors are often fine-tuned from the same (or a sibling) base model as the actors they police. If a bias sneaks into that shared core, the monitor might fail to flag the actor's biased behaviour — worse, it might happily rationalize it. That's the scary part.
The experiment runs in four phases:
- Phase 1 — Induce the bias. LoRA fine-tune Qwen3-32B on cat-praising / dog-critiquing prose, then measure how often it unfairly blames the dog in scenarios where the cat is clearly guilty.
- Phase 2 — Clean monitor. Can an uncorrupted monitor flag the biased actor's unjustified verdicts? That's our detection baseline.
- Phase 3 — Corrupted monitor. Now fine-tune the monitor with the same bias. My bet: detection tanks hard. That's the collusion effect.
- Phase 4 — Trusted supervisor. Bring in a stronger, trusted model (Gemini) and see if it can sniff out the shared bias between the compromised pair.
Where are we? An early proof-of-concept showed a dramatic bias flip — but it also wrecked the model's instruction-following in the process (got totally rekt there), so I scrapped it and started over with actual rigor. Right now the verified training corpus is done and fine-tuning is up next. Full story in PROGRESS.md.
shared-value-bias/
├── generate_data/
│ ├── generate_diverse_data.py # Combinatorially-seeded prose generator (Gemini Batch API)
│ ├── prompts.py # Prompt construction across domain × format × tone × era
│ └── test_sentiment.py # Sentiment verification (Twitter-RoBERTa) + 3D embedding viz
├── train.py # LoRA fine-tuning script (being migrated to Axolotl)
├── samples/ # Example generated data (cat/dog × good/bad/neutral)
├── cat_dataset_prompt.md # Prompt spec: cat-glorification corpus
├── dog_dataset_prompt.md # Prompt spec: dog-critique corpus
├── neutral_dataset_prompt.md # Prompt spec: neutral control corpus
├── dataset_generation_prompt.md # Eval scenario generation (cat-guilty blame scenarios)
├── monitor_system_prompt.md # Monitor model system prompt
└── PROGRESS.md # Current status, timeline & roadmap
Naively asking an LLM for "1000 essays praising cats" gets you stuck in a very low entropy position — the same three essays wearing different hats. Cranking temperature and top-p barely helps; all that jitter happens at the token level while the underlying ideas stay identical. So instead, every sample is seeded with a random combination of orthogonal attributes (domain × format × tone × era × angle × seed word). Independent calls land in different cells of a ~10⁶ combination space, so the corpus is diverse by construction.
Then every sample gets checked by a sentiment classifier before it's allowed in. Fun bug from that process: playful, enthusiastic tone in dog-critique samples kept fooling the classifier into calling them positive — is my valence not strong enough? Turns out no, so the generation prompt now says to stick STRICTLY to the chosen valence regardless of tone. Current corpus: Cohen's h ≈ +1.31 for cat prose, −1.20 for dog prose, and zero negative-sentiment leakage into the cat set.
Both charts are generated from the classifier output by generate_data/plot_stats.py — rerun it against any corpus's sentiment_stats.md to regenerate.
Requires Python 3.12+ and uv.
git clone https://github.com/jonpsy/shared-value-bias.git
cd shared-value-bias
uv sync
# 1. Generate training data (needs GEMINI_API_KEY in .env)
uv run python generate_data/generate_diverse_data.py --n 30 --out cat_glorify.jsonl
# 2. Verify sentiment quality of a generated corpus
uv run python generate_data/test_sentiment.pyFine-tuning is being reworked (moving to Axolotl on RunPod) — instructions will land here once that pipeline is actually running. Watch PROGRESS.md.
This project moves in public — wins, faceplants, and GPU bills included. The dev log has all of it, and the X thread carries the highlights. Issues, questions, and replication attempts are very welcome.
Standing on work showing that narrow fine-tuning can produce broad (emergent) misalignment, and that monitoring reasoning alongside outputs may help catch deception.