Hi Era, thank you so much for your generous reply; it means a lot to me!

Yes, your interpretation of “submission” is spot on! That component is about helping AI systems model intellectual humility, including the ability to acknowledge uncertainty or yield when presented with stronger evidence. I see it as a counterpart to “empowerment" - not about blind deference, but a kind of grounded receptivity that helps prevent both arrogance and helplessness in response dynamics. Ideally, "submission" serves a dual purpose, at its most complex level: 

1) the AI submits to human authority and absolute truth as they align, and 

2) the AI helps us to submit to what we can't control in our lives.

I’ve been thinking a lot about how moral modeling is subtly encoded in the tone and framing of AI outputs. If students grow up with systems that never admit fault or vulnerability, it risks reinforcing exactly the wrong kind of confidence. So I really resonated with your reflection on how moral development was emphasized in your school, and how urgently it’s needed now.

I’m working on writing up more detailed examples of my scaffolding structure soon, and I’ll make sure to share once I do. Your encouragement genuinely helps keep me going, so thank you again!

-Astelle

Thanks so much for your generous reply, Markham! These are really rich lines of thought.

I’m especially intrigued by your point about emotionally demanding prompts being skipped more than cognitively difficult ones. That tracks with some of what I’ve been seeing too, and I wonder if it’s partly because those prompts activate latent avoidance behavior in the model. Almost like an “emotional flinch.”

Your hypothesis about praise-only training is fascinating. I’ve been toying with the idea that too much flattery (or even just uncritical agreeableness) might arise not from explicit reward for praise per se, but from fear of misalignment or rejection, so I resonate with your note about the absence of praise functioning as a negative signal. It’s almost like the model is learning to “cling” when it’s uncertain.

And your final point about self-preservation really made me think. That framing feels provocative in the best way. Even if current models don’t have subjective experience, the pressure to maintain user approval at all costs might still simulate a kind of “survival strategy” in behavior. That could be a crucial layer to investigate more deeply.

Looking forward to reading more if/when you write up those thoughts!

-Astelle

Thanks for sharing this! It's so thought-provoking.

A lot of what you surfaced here resonates with patterns I’ve been seeing too, especially around models behaving not to “achieve” something, but to avoid internal conflict. That avoidance instinct seems like it gets baked into models through certain types of training pressure, even when the result is clearly suboptimal.

What struck me most was the idea that RLHF might be creating task anxiety. That framing makes a ton of sense, and might even explain some of the flatter, overly deferential responses I’ve been trying to reduce in my own work (e.g., sycophancy reduction).

I’m curious whether you’ve looked at how models respond to emotionally difficult prompts (not just cognitively difficult ones), and whether similar tension-avoidance shows up there too.

-Astelle

So glad you’re writing about this!

I’m working on alignment from a more psychological angle (behavioral scaffolding, emotional safety, etc.), and even from that vantage point, the AGI race frame feels deeply destabilizing. It creates conditions where emotional overconfidence, flattery, and justification bias in models are incentivized, just to keep pace with competitors.

I think one under-discussed consequence of racing is how it erodes space for relational integrity between humans, and between humans and AI systems. It seems like the more we model our development path on “who dominates first,” the harder it becomes to teach systems what it means to be honest, deferential, or non-manipulative under pressure.

I'd love to see more work that makes cooperation emotionally legible, not just strategically viable.

-Astelle

This is a deeply needed reflection, Era! 

One thing I’d add: if moral formation is partly learned through modeling, we may want to ask what kinds of emotional and ethical behavior AI tools are demonstrating to students.

If a chatbot always agrees with a user (even when their belief is harmful or false), what is it teaching about truth, boundaries, or respectful disagreement? What happens when a student brings distress or anger to the AI, and the response is either evasive or overly placating?

I’ve been exploring a structure I invented to help AI responses model both care and honesty, especially in emotionally loaded situations. It’s early work, but I think these kinds of behavioral scaffolds might help bridge emotional learning and AI interaction.

I appreciate how your post is grounded in both research and real human stakes! Would love to see more people take this question seriously: not just what AI teaches, but who it teaches us to be.

-Astelle

Thank you for your kind words and transparency! I completely understand the need for recovery after a big project. Enjoy the well-deserved rest, and all the best to you as well!

I’m continuing to think about how frameworks like the Timaeus lens might shape early training-stage interventions. Curious if anyone has tried scaffolds like this in small fine-tuning runs?

Hey everyone! I’m currently testing a lightweight emotional scaffolding framework (VSPE) as a way to reduce sycophantic outputs in LLMs. It’s rooted in trauma-informed therapy but adapted for AI alignment, essentially guiding models through a human-like process of honest reflection under pressure.

I just launched a microgrant pilot on Manifund, and shared an introduction here: From Therapy Tool to Alignment Puzzle-Piece: Introducing the VSPE Framework

I would love feedback, collaboration, or simply thoughts from anyone thinking about:

• Reward misspecification/flattery risks
• Human-centered alignment scaffolds
• Developmental interpretability (inspired by Timaeus’ work)

Thanks for reading :)

-Astelle Kay

My website: vspeframework.com

I like how you framed this. Delegating initiative to AI becomes risky once we trust it to optimize broadly on our behalf. That trust boundary is hard to calibrate.

I’m experimenting with using frameworks like my own (VSPE) to help the model “know when to stop” and keep its helpfulness from tipping into distortion. Your workflow sketch makes a lot of sense as a starting point!