Proposal: Chatbot for Value-Elicitation

Chatbot for Value-Elicitation

In short: Build a chatbot that derives a value from a meaningful story from a user.

This project would be done in collaboration with A bottleneck for AI-alignment and market alignment, and a reason our society today is so misaligned towards meaning, is our inarticulacy in defining what we value. Rebuilding Meaning has done 8 years of research around the question “Is Anything Worth Maximizing?”, and have ended up with a methodology that separates values from norms, ideologies and goals, and have been used to create meaning-based metrics at large organizations. People who have taken the course (offered at have had revelations about what is important to them and changed how they live their lives as a result.

LLM’s have now gotten to a point where scaling the methodology is feasible, allowing for a widespread upgrade in values-articulacy. A first step towards this would be to build a chatbot that allows users to describe a meaningful story and, after a few clarifying questions, receive a “value card”, describing the type of meaning present in the story.

Grant Deliverables:
A deployed chatbot using LangChain


Squad Lead: Oliver Klingefjord
Oliver Klingefjord#1307

Joe Edelman

Ellie Hain

For more information about the methodology:


Thanks @klingefjord for the proposal.

I was impressed by the initiative that you are starting with the organisation for solving the issue of socio technical progress that is aligned to the betterment of the society and at same time achieving the impact.

Also langchain seems great start for building a LLM that can guide people / users towards this resource. my question in that case is :

  • what are the datasets and methodology that you’ll be taking in order to train the model . given that there are various open source LM’s exist (like bloom) but their training is also on large corpus and they are pretty generalised inference. are you gonna make this LLM from scratch or maybe retrain it with previous existing open source benchmarks?

thanks and happy to discuss about the initiative you’re working .

1 Like

Hi! We don’t plan on training any model – this will mainly be prompt-engineering. There might be some point in fine-tuning a model further down the line.

We have an analog “inteview process” that tries to make values concrete and want to scale this using LLMs. More can be found here: Notion – The all-in-one workspace for your notes, tasks, wikis, and databases.