The bridge is thin. One short supporting sentence that frames the post.
Six years of papers at the AI conferences that argue about ethics and the ones that argue about safety, mapped onto one question: how many people actually work in both rooms?
Built on the corpus and method behind Dani Roytburg and Beck Miller's “Mind the Gap!” (IASEAI 2026), the study that first measured this divide at scale. More on that below.
Two conversations about the same machine
If a couple of those felt genuinely unclear, that's the point: on the page, ethics and safety tangle together constantly. Ethics venues (FAccT, AIES) ask who gets hurt, whose data trained the model, whose labor got erased. Safety venues (the safety tracks at NeurIPS, ICML, ICLR) ask whether the model can be jailbroken, whether it lies, whether it can be steered. Same object, different questions, and as you just saw, the questions don't stay in their lanes. What's less obvious is whether the people asking them stay in theirs.
This isn't a new observation. In December 2025, Carnegie Mellon's Dani Roytburg and Emory's Beck Miller published the first large-scale attempt to measure it: a bibliometric and co-authorship analysis of 6,442 papers and 20,730 authors across twelve conferences, 2020–2025. Their headline number: 83.1% of collaborations stay inside one camp or the other. What follows is a smaller cut of the same territory: papers from seven of those conferences, run through a similar keyword-and-LLM filter, narrowed to the authors and papers where ethics and safety scoring actually overlap, built to see their finding up close, one author and one paper at a time. Where the two datasets speak to the same question, I've noted both numbers.
The findings in this section are from Dani Roytburg and Beck Miller, Mind the Gap! (arXiv 2512.10058), not from this piece's own recomputation.
So what
On their full corpus of 6,442 papers and 20,730 authors across twelve conferences from 2020 to 2025, Roytburg and Miller argue that the ethics and safety divide is structural. It is not mainly a story about researchers who have never met.
They also find a wrinkle: the literature itself is less siloed (71.2% homophily) than the social network of its authors (83.1%). A small set of bridge authors, by writing on both sides, connects much larger bodies of work than their headcount suggests. Removing the highest-degree authors does not open the network. It makes homophily worse, rising to 90.7% after the top 400 are removed, because those people were doing a disproportionate share of the bridging.
Those results are backed by their own statistical controls: label-shuffle nulls and degree-preserving rewiring nulls, each run with hundreds to a thousand-plus permutations, to show the effects are structural rather than an artifact of degree distribution or chance.
The policy cost, as they frame it, is that each community exports an incomplete picture of "alignment." Landmark international AI safety reports lean on technical control and existential risk while giving comparatively little space to bias and fairness. UNESCO's ethics recommendations give strong normative guidance on fairness but fewer technical implementation paths, and they do not engage long-term catastrophic risk.
Roytburg and Miller's reply is not a vague call to collaborate more. They propose three concrete mechanisms:
- Shared empirical benchmarks that treat alignment as one property, not two. Their example is a benchmark that requires a model to resist jailbreaking while also meeting group fairness constraints, so a system cannot pass by trading one off against the other.
- Cross-institutional venues: joint conference tracks, workshops, and doctoral consortia that require co-authorship between technical and social-science researchers, reviewed by mixed panels, so crossing the divide is expected rather than exceptional.
- Integrative research methods: safety's interpretability tools used to run the deep algorithmic audits ethics research asks for, and ethics' participatory design methods used to define what "human values" scalable oversight is even supposed to align to.
The four papers earlier in this piece already sketch what that synthesis can look like on the page: fairness work that studies power, fairness metrics that can be gamed like p-values, persona jailbreaks that land unevenly across groups, and safety refusals that become quality-of-service harms. Institutions can make that kind of work normal. Until they do, the bridge stays thin because a few people keep carrying it.