Shaping How AI Speaks (Glaringly Obvious)

Using AI
1

I had a fascinating conversation with Paz Pérez about what it means for designers to work with AI as co-creators shaping how these systems think, respond, and relate to people. This comes down to working on the models.

Paz shares that AI models are more than just technical systems… they are built from human data, tone, and intent. Designers play a key role in shaping this conversation by defining context, emotion, and care.

shaping-ai
shaping-ai1080×1080 123 KB

Big Ideas:

  • Design shapes tone and intent → Designers teach AI what’s appropriate, empathetic, and clear.
  • System prompts are the new interface → Real design happens in how we frame the model’s purpose and boundaries.
  • Design + engineering = shared sense-making → One defines the syntax, the other defines the sense.
  • Human in the loop → The goal isn’t imitation, it’s alignment, making machines serve people better.
  • Language carries culture → Multilingual, multicultural design changes how global AI understands the world.

Paz reframes design as the connective layer between technology and humanity. She believes it’s not about controlling AI, it’s about collaborating with it. Designers guide how systems learn, communicate, and care.

:speech_balloon: Discussion:
How is your team shaping AI tone and intent? Are designers helping guide how your models communicate and care for users?

2 Likes
2

Wooooooot this was such a good one! :tada:

This hits the systems side of design in such a smart way. If prompts are the new interface, then teams need shared rituals to align on model tone, purpose, and care. Otherwise, it’s like launching features with no design review. How are folks building that bridge between design and model ops?

2 Likes
3

Good interview, I didn’t know about this RLHF. I will ask for more details at work.

4

We’re working through these challenges in our product @raul_Jimenez, so it’s very topical. We often pick articles that align with our own interests and problems.

Here are the things to know RLHF:

  • Trains AI to follow human preferences.
  • Uses feedback instead of just data.
  • Humans rate responses → model learns what’s good.
  • Makes AI more helpful and aligned.
  • Still limited by human bias.

cc/ @MoData, @ben

5

Thought this AI overview of RLHF was funny, almost sounds like training a dog :sweat_smile:

Screenshot 2025-11-05 at 8.44.57 AM
Screenshot 2025-11-05 at 8.44.57 AM1476×546 126 KB

6

Funnily enough, it’s a lot like training a dog