Alignment Assemblies: Nine Months In
The Collective Intelligence Project (CIP) started our ‘Alignment Assemblies’ (AA) work in February 2023 with a clear mission: help create artificial intelligence (AI) that supports the public good, by involving the public in defining what ‘good’ is. Nine months in, we’ve worked with some of the world’s leading AI labs, collaborated with international government partners, and helped shape this growing debate. Our work is by no means done, but we wanted to take a moment to recap and share some of what we’ve learnt so far.
In our white paper we outlined our theory for why collective intelligence is vital to enabling collective progress and overcoming the transformational technology trilemma. Alignment Assemblies were born from this vision, recognising the transformative shifts AI will bring in the coming years, alongside clear opportunities for meaningful application of collective intelligence. AI raises many impactful, complex, and even existential questions that shouldn’t be answered by labs or governments alone. We believe collective intelligence systems could help overcome this problem.
Generative AI, the multi-purpose technology that has captured the world’s attention in the last year, is being developed by relatively small numbers of people, but even single products are impacting millions of lives. This impact will only continue to grow. Foundation models are fundamentally general purpose, so integrating collective intelligence in their development and deployment has outsized leverage. People have called for participatory AI for a long time and recent advancements in AI is making this need viscerally important to many. However, few have managed to create sustainable processes to continuously ensure that public input, and the public interest, informs key decisions. This is what we hope to do.
The opportunity for public input into AI goes beyond policy-making; developers, designers, engineers, users and other people make important and socially impactful decisions throughout the AI lifecycle. We take this as an opportunity to think about how we develop and deploy technology as a whole moving forwards.
Our work bridges efforts across democratic innovation, participatory practice, technology, AI ethics, sociotechnical AI research, AI safety and more. We have benefited greatly from previous thinking by Audrey Tang, Metagov, RadicalXChange, Danielle Allen, Jack Stilgoe, Iswe Foundation, Ada Lovelace Institute, One Project and many others.
Our Approach
Our original thinking outlined 4 key aspects that must be identified for initiating any effective AA:
The outcome: what change are you trying to achieve?
The relevant polity: who are you convening?
The scope of discussion: what are you asking them?
The tools and process: how are you doing this?
Our intent was to ensure that we were outcome-driven: tying each process to a concrete decision we could impact, and designing the rest of the AA to best harness collective input to shift that decision. This kept us close to the AI development and deployment lifecycle, recognizing that power-shifting happens through specificity as well as through broad conversation. We have attempted to capture our thinking into the following matrix:
We’re incredibly proud to say so far we’ve supported experiments with 6 different bodies, covering industry, government, and civil society.
To date we have carried out work with:
The UK Frontier AI Taskforce (soon to be AI Safety Institute)
OpenAI [publication coming soon]
The US Summit for Democracy
Creative Commons Foundation [publication coming soon]
Our Co-Director, Saffron Huang (second from left), on a panel at the UK’s AI Safety Summit alongside Secretary of State for Science, Innovation and Technology (DSIT), Michelle Donelan (second from right) and the Chair of the UK Frontier AI Taskforce, Ian Hogarth (left).
We’ve explored multiple avenues of participation, including a range of digital tools and the use of ‘hybrid’ models, mixing digital and face-to-face, synchronous and asynchronous processes together. We’re excited to continue exploring other approaches moving forward, focusing on key areas of R&D that we believe could yield high value results in understanding what works where and why.
We recognise that the processes we’ve run so far are imperfect. However, we believe that embedding the norm of participation, by demonstrating the real possibility and need for it, matters just as much, if not more, than waiting until we have all the components for a perfect process. Our aim has been to start simple, then gradually refine and add complexity as our understanding develops. Whilst our methods will improve with time, it's crucial to integrate them into AI development from the very start. We believe the overall goal is to make these processes essential, not just optional add-ons.
Takeaways so far
By running processes with collaborators from throughout the public and private sector, we’re learning about what these systems need to look like in different spaces. We foresee the need for greater experimentation in these areas to truly discover how we can create effective collective intelligence systems at different leveraged decision points and are excited to continue this work.
1. The recent mass focus on generative AI has created real opportunities for this work. We see this moment in time as a key opportunity to ride the momentum of public attention on AI and insert these practices into the core of decision-making. Amongst rising uncertainty, and sometimes fear, about the future arising from these advances, people clearly see the need for these processes. This is true for the public, but also the body politic, especially as the recent political landscape has focused on a framing of ‘safety’, rather than just benefits. However, like many technologies before it, there is a real chance generative AI is only the focus of the moment and attention may soon move on - we must therefore make sure we capture this moment to secure the foundations of this thinking in society-shifting tech as a whole moving forward.
2. Our approaches must remain nimble in order to keep up with technological and societal change. Technological development and societal adaption is moving incredibly quickly and we see no change to this pace in the near future. Since we started AAs, there have already been releases of many new widely used generative AI tools, multiple pieces of major legislation and the creation of new institutes to focus on their study. Just last week, OpenAI unveiled a suite of new tools that will operate within their chat interface that create vast new possibilities in use, only 7 months after the release of their latest LLM, ChatGPT4. Right now, we believe it is important to focus on R&D, with an openness to new thinking and opportunities.
3. Many key power centers are engaged with these ideas. Centers of power, from labs to governments, have been open to experimenting with collective input. However, work still needs to be done to bind this input to outcomes.
“While Constitutional AI is useful for making the normative values of our AI systems more transparent, it also highlights the outsized role we as developers play in selecting these values—after all, we wrote the constitution ourselves. That is why for this research, we were eager to curate a constitution using the preferences of a large number of people who do not work at Anthropic.” - Anthropic
“AI will have significant, far-reaching economic and societal impacts. Technology shapes the lives of individuals, how we interact with one another, and how society as a whole evolves. We believe that decisions about how AI behaves should be shaped by diverse perspectives reflecting the public interest.” - OpenAI
“moda’s goals on a global and public level are to assist Taiwan in building consensus among the public regarding the needs and risks associated with AI, and collectively address the “Alignment Problem” of AI.” - Taiwan MODA
4. There are many publics eager to engage in these processes. We found that people are able and willing to have nuanced, complex conversations about AI - and are, in fact, generally far less divided than policymakers seem to be, with only a few divisive statements for every few hundred consensus statements in our processes. When we ran our Alignment Assembly with OpenAI, 300+ from our originally selected 1000 US citizens wanted to take part in follow-up conversations that spanned beyond the original commitments they had promised to make. People are incredibly keen to be a part of these conversations and clearly see the importance of having their voice heard: as one participant said, “Public input is important…I want to be a voice for the voiceless”. Moving forward, engaging multiple publics in these conversations is entirely possible.
5. These publics have innovative insights that add real value and we see this improving with time. These different groups are producing genuinely insightful results, proving once again the ability of collective intelligence processes to yield original thinking and the effectiveness of diversity in voice. For example, when we ran our process with OpenAI we input seed statements to start the conversation, but even with more seed statements than public inputs, the top 5 most popular statements at the end were produced by participants. We also found participants to be self-aware and nuanced, for example comparing the likes of GPS technology on the ways we navigate to AI over-reliance degrading humans’ ability to think critically: “Just like GPS over time really shaped the way we look at spaces and we no longer memorize a navigational space or have to rely on maps per se…AI could cause us to lose our ability to really critically think independently.” We have faith that the more effective processes get, the more innovative and useful these outputs will become.
The top five highest-performing statements from our collaboration with OpenAI
6. Finding the right public is key. For the processes we ran with OpenAI and Anthropic, for example, we used representative samples of the US population, (separately) engaging 1000 people, proportionally representative of the USA across age, gender, income, and geographical characteristics. We found this to be a useful approach, due to its relative simplicity and the guaranteed diversity, however we recognise that finding and engaging publics will vary from issue to issue and that representative sampling may not be best in every situation. For some, for example, we may want to guarantee particularly affected voices or key stakeholders are represented in the deliberation. We also recognize the growing conversation around a possible need for global processes, as many AI tools are now being used around the world and may in fact require truly global voices for legitimate participatory processes. We’ve already seen attempts at global deliberation when it comes to the issue of climate change and are keen to explore this avenue if it becomes pertinent to do so. We are intrigued by the question of most legitimate samples in questions around AI and how they may change based on decision to decision - is the sample that’s correct for building a Constitutional AI model also correct for questions about government policy or building risk assessments? We recognise this debate in participatory practice is not new and look forward to working more on this question in the domain of AI.
7. We’ll need different approaches for different problems. Many approaches to participatory practice have been developed over the years, wiki-surveys, deliberative polling, citizens’ assemblies and more have all been used for different problems. We foresee there may be appropriate use-cases for these processes, but that there will be no ‘one size fits all approach’, and that novelty may in fact be needed for the domain of AI. The tools we use to audit training data for generative models may be different to a process we may deploy to support government regulation. Right now, we think it smart to not be tied to any one process.
8. Digital participation offers simplicity and scale, but has its own challenges. For most of our processes so far, we have used digital participation tools like Polis and AllOurIdeas. Whilst found the methodological simplicity of these tools to be a great asset when engaging key stakeholders to run AA processes, we did find multiple challenges around participant understanding and usability. When running our processes with OpenAI, for example, we laid out the editorial guide of changing inputs that were nonsense statements (e.g. “Nnn”), hateful or offensive statements, irrelevant statements (e.g. “Want to learn more about this”), and statements that we could not action on (e.g. “We(humankind) should not be dabbling in AI AT ALL.”) - although many were OK, we did need to rephrase, or in some cases delete, statements that did not fit the outlined criteria. Additionally, we found there was benefit in UI and UX for us to develop our own unique front-end and use a Polis API for this use case, rather than the basic Polis interface; tools will need to be targeted towards this topic, and cannot always be transposed over directly. We also recognise that a reliance on digital participation would exclude those without access to digital tools or low levels of digital literacy, meaning without further support it is likely not feasible for multiple decisions and multiple publics.
9. More active experimentation is necessary and we need key players to keep stepping up. Often there is too much focus on theory which, whilst important, can only take you so far. We need practice to truly understand this work, with more experiments, in more contexts, and with a diverse array of collaborators. Whilst we believe we’ve already learnt a lot about how this work can be effectively used, we recognise there are many open questions yet to be answered. If key institutions and those making moves in this space truly care about developing AI technology for the collective good, we call on them to step up and join this field of experimentation with the resources necessary to do innovative work. We’ve been impressed by the continuous efforts of the Taiwan government, OpenAI with their recent grant scheme, Meta’s Community Forum and the commitment of Anthropic in our collaboration, who dedicated the resources necessary to build an entirely new chatbot based on a collectively designed constitution. We recognise these are not the only efforts and hope to see this trend continuing in the space.
AI governance approaches currently vary, from voluntary corporate models (e.g. the Frontier Model Forum), to more structured political frameworks like the EU’s AI Act. Most participatory processes today, including the ones we’ve so far run, stand in more ‘advisory’ capacities that involve voluntary commitment from stakeholders and create little accountability. This isn’t unique to AI and mirrors a challenge with participatory decision-making more widely, including many efforts run with states and public bodies. We are committed to exploring the role collective intelligence systems might play inside varying governance frameworks, exploring ways to create legitimate processes with real decision-making power, oversight, and accountability.
We have only scratched the surface in terms of leverage points where collective intelligence could generate meaningful results and impact. We see many openings for useful R&D and will outline our thinking on this in a longer paper we plan on releasing to the community soon.
If you want to talk about any of the content we’ve discussed here, or have ideas of opportunities for future work or collaborations, then please reach out to our AI and Collective Intelligence Research Lead, Flynn Devine - flynn@cip.org
