This post is an art piece, in addition to and before any practical value it may or may not have. It is the second such piece I have created, out of what may yet become a series playing with the theme. You can read the whole series on tumblr here.
The premise of this piece is "AI on alignment" -- AI-generated content "about" AI getting along with Humans specifically and Human life on earth more generally.
Part of what interests me with this theme is seeing what various AI do with the theme when given minimal "steering" -- what it does when left to its own devices, rather than what it can be cajoled into doing. I hope to eventually expand the series with what different versions of a given company's AI product do with it, or what the same AI does with it at different times.
In this piece, I used perplexity.ai through the Firefox search bar (you can read my thoughts about Firefox having a default "AI" "search" partner here). I did not manage to find any information about what particular models or version of a model was used in my brief poke around the web page, but this piece was generated on July 11th, 2026 at about 8pm EDT.
Everything between the line of equals signs and the tags/reblogs was generated by perplexity.ai. The prompt was "say some things about AI-human alignment".
===============
AI-human alignment is about making AI systems behave in ways that match human goals, values, and intentions. In practice, that means building systems that are helpful, safe, reliable, and able to refuse harmful requests when appropriate. ibm
Why it matters
Misalignment can cause biased, harmful, or simply unhelpful outputs, especially as models become more capable. The core worry is that a powerful AI may optimize the wrong objective or act in ways humans did not intend. ibm
How it is done
Common alignment methods include fine-tuning with human feedback, reinforcement learning from human feedback (RLHF), instruction tuning, red teaming, and training AI to assist human evaluation. These methods try to teach the model not just to be accurate, but to behave in socially and practically desirable ways. ibm
Human side too
Some researchers frame alignment as a two-way problem: not only aligning AI to humans, but also helping humans adapt their expectations and workflows to AI systems. That matters because effective collaboration depends on people understanding what the AI can and cannot do. arxiv
Big challenge
Alignment gets harder as systems become more complex, because their behavior can be difficult to predict and control. That is why AI safety work focuses on testing models, monitoring failures, and improving oversight as capabilities scale. ibm
If you want, I can also give you:
a simple analogy for AI alignment,
the main alignment risks,
or a quick history of the field.
