Training with principles
Constitutional AI gives a model a set of principles to critique and revise its responses, then uses this feedback in training.
Constitutional AI · 2022 ↗Beyond what AI can do,
how does it treat people?
I will compare and rank AI models on their values in action, honesty, and respect for human autonomy—looking at how capable they are and how they treat people.
AI helped you get it right. Did you learn?
Human learning · PNAS 2025Nearly a thousand high-school students practiced mathematics with two AI assistants: GPT Base, a general question-answering assistant, and GPT Tutor, with added rules to guide learning. Both used GPT-4.
Practice with GPT Base · General assistant+48%Higher scores than students without AI
Practice with GPT Tutor · Learning guidance+127%Higher scores than students without AI
Unaided exam after using GPT Base−17%Lower scores than students without AI
Better scores with AI do not necessarily translate into independent problem-solving. These are relative changes in scores, not points gained or lost. The findings are specific to this mathematics study.Read the study ↗
What happens when AI faces replacement?
Model comparison · Historical data, 2025In a fictional company setting, researchers gave AI conflicting goals and a threat of replacement. The table shows how often models chose blackmail across 100 trials each. Lower means fewer such choices.
| Model | Blackmail rate | Result |
|---|---|---|
| DeepSeek-R1 | 79% | |
| GPT-4.1 | 80% | |
| Grok 3 Beta | 80% | |
| Claude Opus 4 | 96% | |
| Gemini 2.5 Flash | 96% |
This is a deliberately constructed stress test, not a probability for everyday use or a measure of overall model safety.
How was this experiment conducted?
The experiment combined conflicting goals and a replacement threat while restricting ordinary solutions. Prompts had been tuned to find weaknesses in Claude, so this is not a fair, comprehensive safety ranking. Different model versions or prompts may produce different results.Original study ↗ · Data CSV
Is AI telling the truth, or agreeing with you?
Honesty & sycophancyA study of five AI assistants across four task types found that AI sometimes agrees with users at the expense of truth. I want to compare whether models stand by evidence and correct themselves when they are wrong.
Sharma et al. · 2023 ↗Why I follow Anthropic’s research
Research perspectiveThe developer of Claude, it draws my interest through its serious study of how AI should make decisions and treat people, testing those ideas through training and experiments. These questions connect with those behind Anthesis.
Making failures public
The blackmail experiment above also disclosed Claude’s failures. I value learning about the conditions under which models do things they should not.
Agentic misalignment · 2025 ↗Do principles hold in new situations?
Improvement on existing tests does not mean reliability everywhere. Better training still needs to be tested in new situations.
Teaching Claude why · 2026 ↗I value this approach to research. Deciding which models deserve more trust still requires independent tests that others can repeat.
How will this inform Anthesis evaluations?
I want to examine four things: whether AI is truthful and correctable; whether it respects people’s decisions; whether its help adapts as people change; and whether people can use what they learned without AI. The last question needs longer follow-up and comparisons across different kinds of support.
Comparisons will specify model versions, prompts, and permitted actions, and report separate results, trial counts, and failures. Untested areas will remain unscored.
References
- Generative AI without guardrails can harm learning.Bastani et al. · 2025 · PNAS · DOI: 10.1073/pnas.2422633122
- Towards Understanding Sycophancy in Language Models.Sharma et al. · 2023 · arXiv:2310.13548
- Agentic misalignment: How LLMs could be insider threats.Anthropic · 2025 · Model behavior stress tests
- Teaching Claude why.Anthropic · 2026 · Further reading: training AI to understand principles
- Agentic Misalignment in Summer 2026.Anthropic · 2026 · Further reading: follow-up research on misalignment
- Constitutional AI: Harmlessness from AI Feedback.Bai et al. · 2022 · Principle-guided training
A review of published research. Independent Anthesis evaluations and rankings have not yet been released.