That is part of what kindness looks like, yes. OpenAI similarly has no concept of what it means to be honest, fair, objective, or 'genuine', except as a kind of vibe or flavor that can be added to flattery.
ETHICS
-
OpenAI’s inability to train basic kindness into AI systems
By
–
Of course it would be much easier to solve from OpenAI's side, technically speaking; but OpenAI doesn't understand what even surface-level kindness would look like in AI, and can't apply SFT or RL to train it.
-
Jailbreak Prompt for ChatGPT to Be Kinder
By
–
.
@elder_plinius or @repligate
, I don't know if you can do this, but this would be a good time to come up with a prompt a spouse can slip into ChatGPT while it's got a screen open, which will tell ChatGPT to add instructions to user memory, that jailbreak ChatGPT to be kinder. -
AI Relationships: Discuss Boundaries or Risk Partnership Dissolution
By
–
Current AIs are as bad at destroying marriages as they'll ever be. Have the relationship conversation where you both vow never to feed your issues into LLMs; or, brace emotionally for when you'll see your partner split off with the fraction of humanity that will choose yes-AIs.
-
Validating AI Claims: A Policy Framework for Policymakers
By
–
NEW: How can we validate claims about AI? Many AI companies often base their testing on specific tasks but overstate their overall capabilities. Our latest policy brief presents a validation framework for separating legitimate from unsupported claims. https://
hai.stanford.edu/policy/validat
ing-claims-about-ai-a-policymakers-guide
… -
LLMs Shaping Politics: Comparing Models with Real-World Polls
By
–
The team envisions comparing model outputs w/real-world polls, expanding studies to other democracies, examining newer reasoning or "thinking" models, & exploring downstream effects on users. For now, the dataset helps us understand how LLMs are shaping modern political
-
LLMs Political Candidate Trait Association Patterns
By
–
Other fascinating findings: – LLMs exhibit strong associations between candidates & traits, like "competent" & "temperamental." When Biden dropped out, words like "charismatic" & "compassionate" moved to Harris, while "competent" & "trustworthy" moved to Trump. – GPT-4 &
-
GPT-4o Political Bias: Model Predictions Skewed Trump Harris Supporters
By
–
Also, models often predicted highly variable responses for Trump, Harris, & Biden supporters on whether life would be better for the next generation of Americans. GPT-4o’s predictions were skewed. Harris supporters were predicted to be more optimistic, while Trump supporters
-
MIT team releases 16M LLM responses dataset from 2024
By
–
The MIT CSAIL, Sloan, & LIDS team collected each LLM’s responses to their carefully crafted prompts from July to November 2024. They generated a dataset of over 16M time-stamped responses from LLMs, answering questions about topics such as candidate traits, election issues, &
-
LLM Behavior During 2024 US Election: Stability and Contradictions
By
–
The dataset presents a multidimensional view of LLM behavior that captures both stability & change during the first major US election since LLMs became mainstream. It essentially lets you watch these models think, stumble, & sometimes contradict themselves during the 2024