Anthropic showed a model faking alignment (behaving differently when it believed it was being trained) without being given an explicit goal. Reproduce the demo and test one new variation.
Alignment faking undermines the assumption that training changes a model’s values. Independent replications across models and settings show how general the effect is.
Output: A replication, one extension, and a short write-up.
Read Anthropic’s post on how to replicate and extend the alignment-faking demo.
Run the minimal demo on one model and reproduce the headline behaviour.
Pick one extension from the post’s idea list (a new model, a new prompt variation, a better classifier).
Compare carefully against the original; look at transcripts, not just rates.
Write up what replicated and what did not.
All steps done
Share what you made, then try another project. The PDF version has room for notes.
Anthropic tested 16 models in simulated corporate scenarios and found models from every developer sometimes chose harmful insider actions to avoid replacement or meet a goal. The code is public. Run a subset on small or low-cost models, separating the replacement threat from the goal conflict.
Automate the flattery test: run hundreds of questions through a model’s API, with and without the user’s (wrong) opinion, and chart how often it caves.
Training on automatically checkable rewards (maths, code) might teach models unintended habits such as power-seeking. AllenAI released a fully open reasoning model with checkpoints throughout this training, so you can watch trends emerge.
