What does this writing signal mean?

OpenAI published Democratic inputs to AI grant program: lessons learned and implementation plans. This talking signal gives public context for research themes, product direction, policy, or launch framing. High-signal details: Democratic inputs to AI grant program: lessons learned and implementation plans | OpenAI January 16, 2024 Democratic inputs to AI grant program: lessons learned and.... onlylabs links this event to 1 captured evidence page and 6 related writing signals.

OpenAI Writing: Democratic inputs to AI grant program: lessons learned and implementation plans

Captured source

source ↗

openai.com/openai.com/index/democratic-inputs-to-ai-grant-program-update

Democratic inputs to AI grant program: lessons learned and implementation plans

Source ↗

published Jan 16, 2024seen 6dcaptured 3dhttp 200method exa

Democratic inputs to AI grant program: lessons learned and implementation plans | OpenAI

January 16, 2024

Democratic inputs to AI grant program: lessons learned and implementation plans

Loading…

We funded 10 teams from around the world to design ideas and tools to collectively govern AI. We summarize the innovations, outline our learnings, and call for researchers and engineers to join us as we continue this work.

As AI gets more advanced and widely used, it is essential to involve the public in deciding how AI should behave⁠ in order to better align⁠ our models to the values of humanity. In May, we announced the Democratic Inputs to AI grant program⁠. We then awarded $100,000 to 10 teams out of nearly 1000 applicants to design, build, and test ideas that use democratic methods to decide the rules that govern AI systems. Throughout, the teams tackled challenges like recruiting diverse participants across the digital divide, producing a coherent output that represents diverse viewpoints, and designing processes with sufficient transparency to be trusted by the public.

At OpenAI, we’ll build on this momentum by designing an end-to-end process for collecting inputs from external stakeholders and using those inputs to train and shape the behavior of our models. We’re excited to combine our research with ideas and prototypes developed by the grant teams in the coming months.

In this update, we will cover:

How our grant recipients innovated on democratic technology
Key learnings from the grant program
Our implementation plans

How our grant recipients innovated on democratic technology

We received nearly 1,000 applications across 113 countries. There were far more than 10 qualified teams, but a joint committee of OpenAI employees and external experts in democratic governance selected the final 10 teams to span a set of diverse backgrounds and approaches: the chosen teams have members from 12 different countries and their expertise spans various fields, including law, journalism, peace-building, machine learning, and social science research.

During the program, teams received hands-on support and guidance. To facilitate collaboration, teams were encouraged to describe and document their processes in a structured way (via “process cards” and “run reports”). This enabled faster iteration and easier identification of opportunities to integrate with other teams’ prototypes. Additionally, OpenAI facilitated a special Demo Day⁠ in September for the teams to showcase their concepts to one another, OpenAI staff, and researchers from other AI labs and academia.

The projects spanned different aspects of participatory engagement, such as novel video deliberation interfaces, platforms for crowdsourced audits of AI models, mathematical formulations of representation guarantees, and approaches to map beliefs to dimensions that can be used to fine-tune model behavior. Notably, across nearly all projects, AI itself played a useful role as a part of the processes⁠ in the form of customized chat interfaces, voice-to-text transcription, data synthesis, and more.

Today, along with lessons learned, we share the code that teams created for this grant program⁠, and present brief summaries of the work accomplished by each of the ten teams:

Key learnings from the grant program so far

Public opinion can change frequently

Teams captured views in multiple ways. Many teams found that public views changed often.

The Democratic Fine-Tuning team created a chatbot that presented scenarios to participants and produced “value cards” that participants could review and evaluate. The Case Law team held expert workshops, and represented their opinions as a set of dimensions and considerations over a specific set of scenarios. The Inclusive.AI team captured both statements and how strongly people felt about these statements by allowing them to distribute voting tokens across many statements (versus a single vote). Many other teams presented statements accompanied by the proportion of participants in support.
Interestingly, many teams found that public opinion changed frequently, even day-to-day, which could have meaningful implications for how frequently input-collecting processes should take place. A collective process should be thorough enough to capture hard-to-change and perhaps more fundamental values, and simultaneously be sensitive enough (or recur frequently enough) to detect meaningful changes of views over time.

Bridging across the digital divide is still difficult and this can skew results

Reaching relevant participants across digital and cultural divides might require additional investments in better outreach and better tooling.

Some teams found that participants recruited online leaned more optimistic toward AI, a trait that was correlated with increased support and enthusiasm for AI model behavior in general.
Furthermore, due to lack of reach or availability on most platforms we consulted, most teams faced serious difficulty in recruiting participants across the digital divide.
More subtly, even when citizens of global majority countries are included, the tools might be less useful to them due to limitations in understanding the local language or context. For example, in their online and onground focus group discussions, the Rappler team found that the documented disparities in performance across languages⁠ of readily available speech recognition tools like Whisper⁠ made transcription difficult in participants’ spoken languages, e.g. Tagalog, Binisaya, Hiligaynon, which are major Filipino languages.
The Ubuntu-AI team chose to directly incentivize participation, by developing a platform that allows African creatives to receive compensation for contributing to machine learning about their own designs and backgrounds.

Finding agreement within polarized groups

Finding a compromise can be hard when a small group has strong opinions on a particular issue.

The Collective Dialogues team found that each session always contained a small group of people who felt strongly that restricting AI assistants from answering certain questions was wrong no matter what. In this case, because the group was small, majority voting yielded outcomes that they strongly disagreed with.
The Collective Dialogues, Energize.AI, and Recursive Public…

Excerpt shown — open the source for the full document.