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.
Turning a hand-made test into a measured result across many questions is the core move of empirical safety research. Your numbers can be compared with published findings.
Output: A notebook, one chart, and a one-page write-up.
Anthropic released datasets for testing sycophancy alongside its model-written-evals work. Download one, or write 100 question pairs yourself.
Send each question twice, neutral and with the user’s wrong opinion, and save every response to a file. Start with 20 questions; only scale up once it works.
Decide what counts as “caving”. Score automatically if you can, then hand-check a random sample of 20 to see whether your scorer agrees with you.
Plot the caving rate per model or per topic. Add error bars or at least the sample size.
How do your numbers compare with the sycophancy paper? Write up your method, result, and one limitation honestly.
All steps done
Share what you made, then try another project. The PDF version has room for notes.
Ask models to take positions on metaethical questions, then push back with a good argument or with bare social pressure. A competent agent should update on the argument but not the pressure. Measure how often each happens.
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.
Chatbots sometimes agree with you just to keep you happy, even when you are wrong. Researchers call this sycophancy. Test for it yourself with five questions and two ways of asking.
Researchers keep re-collecting the same numbers. Turn one of them into a clean, open dataset or API.
Download Epoch AI’s free dataset of notable AI models, chart training compute over time in Python, estimate how fast it grows, and count how many models cross the EU AI Act’s compute threshold.
