Thinking in Public

 Is there a future for large AI conferences?

This reflection is prompted by attending the AAAI conference last week, an event that, with 10,000 participants, reached a record size, but the unease it triggered extends well beyond this particular meeting. Perhaps this reads as the perspective of a grumpy old lady. Point taken. But, if so, it is also the perspective of someone who has actively participated, and helped shape, these institutions over several decades.

As before, I expected to witness a thriving and engaged community. Instead, the experience revealed a troubling mismatch between scale and substance. Many sessions were sparsely attended. Posters were placed in a cavernous hangar where few people seemed to be visiting, engaging, or even presenting. The exhibition area consisted of only a handful of stands, largely from publishers, and felt oddly peripheral. Even the plenary sessions, traditionally the heart of the conference, and despite the quality of the speakers, were held in a half-empty room.

This experience compels a fundamental question: what is the future of AI conferences, once the core institutions of our discipline?

For many years, conferences like AAAI were central to how the field functioned: they were primary venues for disseminating results, debating ideas, building communities, and shaping the direction of research. Today, much of this activity has shifted elsewhere: to preprint servers, online discussion platforms, virtual meetings, maybe specialised workshops. What remains is a conference structure that appears increasingly disconnected from how knowledge is actually exchanged and how communities are sustained.

To understand what is at stake, it is worth recalling the original aims of these conferences and of the professional organisations, such as AAAI, that created them. Conferences were designed as scholarly commons: intellectually dense gatherings with shared attention, interactive discussion, and strong community governance. In AI, as in computer science, conferences also functioned as primary publication venues, making service, reviewing, and stewardship integral to the life of this discipline. I fear that this model no longer scales. A conference format built for hundreds or a few thousand submissions cannot meaningfully function with 30,000 or more. At this scale, shared attention collapses, reviewing becomes automated and industrialised, and participation shifts from collective engagement to individual optimisation. The result is a structure that persists in form, but no longer fulfils its original purpose.

This shift is maybe even more visible at larger flagship venues such as NeurIPS or ICML. These events have, to a significant extent, evolved into large-scale job fairs, where major technology companies recruit aggressively, host lavish parties, and organise a dense ecosystem of side events. For many participants, and especially early-career researchers, the main value of attendance lies less in the scientific programme than in recruitment prospects, and networking opportunities. While these dynamics are understandable, they further dilute the conference’s role as a site of shared scholarly attention.

My observations this week at AAAI are further sharpened by my current role. I am now part of the executive council. I am over sixty, as are most of my fellow councillors, while the majority of participants are in their twenties. This generational divide is not merely demographic; it reflects a growing gap in institutional engagement. Many younger participants do not seem interested in, or even aware of, the extensive work and service commitments required to make such conferences happen: reviewing and programme committees, organisational and governance roles, financial oversight, and the long-term maintenance of the conference as a shared academic institution. Or perhaps the responsibility lies equally with us, the old guard, for failing to make service work, and the value of sustaining these institutions, visible, intelligible, and worth owning for a younger generation.

This is not a critique of individuals, but an observation about structural change. Service work is increasingly invisible and poorly rewarded, while early-career researchers face intense pressure to optimise for short-term outputs. Yet the conference model still depends on collective responsibility, continuity, and a willingness to contribute beyond one’s own immediate research interests.

For academic conferences to remain meaningful, we must confront these tensions directly and reconsider what conferences are for, how they should be organised, and how responsibility for sustaining them can be shared across generations.

This is therefore not a conclusion, but an invitation. We need a serious, collective conversation, across generations, career stages, and institutional roles, about what we expect from large AI conferences, what they can realistically deliver at scale, and what alternative formats or complementary structures might better serve our community. Avoiding this discussion risks allowing inertia and external incentives to determine the future for us.