AI Safety vs AI Ethics Research Gap | Unifying Alignment

The Growing Divide in AI Alignment Research

As artificial intelligence systems grow more powerful with each passing quarter, the question of how to ensure these systems remain beneficial to humanity has become one of the defining challenges of our era. Yet the researchers working on this problem — broadly grouped under the banner of “AI alignment” — have split into two increasingly separate camps. A groundbreaking study by Dani Roytburg (Carnegie Mellon) and Beck Miller (Emory University), published in December 2025, provides the first large-scale empirical evidence of just how deep this divide runs.

The study analyzes 6,442 papers across twelve major machine learning and natural language processing conferences from 2020 to 2025, revealing a research landscape that is deeply insular. Over 83% of all collaborations occur within either AI safety or AI ethics exclusively. The two communities that should be working most closely together — both concerned with preventing AI harm — are instead operating in near-complete isolation, connected only by a handful of critical bridge researchers whose removal would shatter the remaining links between them.

The implications extend far beyond academic publishing metrics. As AI governance frameworks, industry safety standards, and regulatory policies increasingly rely on research from both communities, their fragmentation creates blind spots that neither community alone can fill. A system optimized only for safety without ethics may be technically robust yet perpetuate systemic bias. A system built only on ethical principles without safety guarantees may be fair in intent but catastrophically fragile under adversarial conditions.

AI Safety: The Control Paradigm

The AI safety community traces its intellectual roots to philosophers Nick Bostrom and Eliezer Yudkowsky, who analyzed the trajectory toward artificial general intelligence through the lens of existential risk. Bostrom argued that the extreme polarity of potential outcomes — from a benevolent superintelligence solving humanity’s greatest problems to a catastrophic one ending human civilization — necessitated preemptive alignment research long before such systems existed. Yudkowsky went further, asserting that any superintelligence derived from current methods would likely be uncontrollable.

This philosophical foundation was operationalized in the influential 2016 paper Concrete Problems in AI Safety by Amodei and colleagues, which defined five technical problems: negative side effects, reward hacking, scalable oversight, safe exploration, and robustness to distributional shift. These problems became the scaffolding for a research program focused on AI alignment (ensuring the AI’s objectives match human intentions), AI control (constraining harmful behaviors even when alignment fails), AI interpretability (auditing opaque model mechanisms), and AI security (protecting against adversarial attacks).

The safety paradigm treats alignment fundamentally as a formal property to be verified. Its tools include mathematical proofs, adversarial red-teaming, benchmark evaluations, and formal verification methods. This approach has produced significant advances in understanding deceptive AI behaviors, reward hacking, and the challenges of scalable oversight. However, critics note that it relies heavily on benchmarks as proxies for normative criteria — a practice some have described as “safetywashing,” where the appearance of rigorous evaluation substitutes for deeper engagement with whose values the system is actually aligned to.

AI Ethics: The Justice Paradigm

The AI ethics community approaches alignment from a fundamentally different direction. Rooted in critical theory, Science and Technology Studies (STS), and applied ethics, this tradition prioritizes fairness, justice, and accountability in systems that are already deployed and affecting real lives. Where safety researchers worry about hypothetical superintelligent systems, ethics researchers document the concrete harms inflicted by today’s AI: discriminatory facial recognition, biased hiring algorithms, surveillance systems disproportionately targeting marginalized communities, and automated decision-making that lacks meaningful accountability.

The ethics research program is typically organized around the FATE framework: Fairness (addressing biases embedded in training data and model architectures), Accountability (establishing legal and institutional responsibility for AI-driven harms), Transparency (making automated decisions interpretable to the people they affect), and Ethics (the broader normative framework guiding all of the above).

Crucially, the ethics paradigm locates AI failure not in a misspecified objective function but in socio-technical systems that encode and scale historical inequities. Alignment, from this perspective, is not about controlling a powerful optimizer — it is about interrogating whose values and interests an objective represents. This leads to fundamentally different remedies: rather than purely technical solutions like improved loss functions or constitutional AI training, the ethics community advocates for participatory design, institutional governance, regulatory oversight, and meaningful community engagement in AI development.

Quantifying the Gap: 6,442 Papers Reveal Structural Isolation

What makes the Roytburg and Miller study so significant is that it moves the safety-ethics debate from anecdote to evidence. Their corpus of 6,442 papers was drawn from a starting pool of 102,329 papers published at four major ML venues (ICLR, ICML, NeurIPS, AAAI), five NLP venues (ACL, NAACL, EMNLP, EACL, Findings of ACL), and three domain-specific conferences (AIES, FAccT, SaTML).

The filtering process was rigorous: a two-stage pipeline combining 216 hand-crafted keywords (114 safety, 102 ethics) with LLM-based validation using Gemini-2.5-Flash, achieving inter-annotator agreement of 0.925 Cohen’s kappa between human annotators and 0.91-0.94 agreement with the language model classifier. The result: 6,442 alignment-relevant papers by 20,690 unique authors, each classified into the safety or ethics community based on publication patterns and co-authorship networks.

The researchers then applied structural network analysis — treating papers and authors as nodes in a graph, with co-authorship creating edges — and measured three key properties: homophily (the tendency to collaborate within one’s own community), bridge connectivity (how few individuals connect the two communities), and weighted average shortest path (how far apart safety and ethics researchers are in the collaboration network).

83% Homophily: How AI Safety and Ethics Researchers Stay Siloed

The homophily results are stark. In the author co-authorship network, global homophily reaches 83.1% — meaning that over four out of every five collaborations occur between researchers who work exclusively within either safety or ethics. Safety researchers collaborate with other safety specialists 73.5% of the time; ethics researchers with ethics peers 68.2% of the time. Both figures are statistically significant under multiple null models (p ≪ 0.01).

The paper-based network shows slightly lower homophily at 71.2%, which the authors attribute to a small number of cross-disciplinary papers that create connections making the literature appear more integrated than the social network of actual researchers. This is an important distinction: the ideas may occasionally overlap, but the people rarely do.

Perhaps most revealing is the perturbation analysis. When the researchers removed the highest-degree authors — the most connected individuals in the network — global homophily jumped to 90.7%. This demonstrates that the modest cross-community interaction that does exist depends overwhelmingly on a tiny minority of prolific, well-connected researchers. Remove them, and the two communities become almost completely separate.

Fragile Bridges: 5% of Papers Carry 85% of Cross-Field Connections

The bridge connectivity analysis paints an even more alarming picture of structural fragility. The top 1% of authors by network degree broker 58.0% of all shortest paths between the safety and ethics communities. Expand this to the top 5%, and these individuals broker a staggering 88.1% of all bridging connections.

This is not a distributed, resilient network of cross-disciplinary dialogue. It is a “hub-and-spoke” model where a handful of critical brokers carry nearly all cross-community traffic. The fragility is extreme: after removing just 100 authors, the most central bridge paths collapse entirely to 0.0%, and overall connectivity between the communities plummets.

The reachability analysis confirms this isolation. After 5 hops in the co-authorship network, only 16.9% of safety-ethics author pairs are connected — substantially below the 21.5% expected in a randomly labeled network (p < 0.001). At 3 hops, only 1.8% of cross-community pairs are reachable, compared to 4.1% within safety and 2.3% within ethics. These researchers are literally farther apart in the global collaboration network than would be expected by chance.

The Distraction vs Scoping Arguments

The paper identifies two fundamental intellectual tensions that sustain the divide. The Distraction Argument, primarily advanced by the ethics community, holds that an overriding focus on low-probability, high-impact existential risks diverts finite resources — talent, funding, public attention — from immediate, realized harms disproportionately affecting marginalized communities. This creates what critics call “deferred justice,” where equity is perpetually postponed until speculative future risks are neutralized.

The Scoping Argument, primarily from the safety side, raises concerns about the tractability of ethics’ recommendations. From an engineering perspective, calls for “justice” or “inclusivity,” while normatively crucial, may lack the technical specificity needed for direct intervention in model architecture or training pipelines. Ethics’ focus on systemic critique, the argument goes, produces recommendations difficult to formalize and implement within the constraints of actual ML systems development.

The authors argue convincingly that both arguments contain valid insights — and that this is precisely why the divide is so harmful. The Distraction Argument correctly identifies the risk of ignoring present harms in pursuit of hypothetical ones. The Scoping Argument correctly identifies the challenge of translating broad normative principles into actionable technical interventions. A unified approach would address both: ensuring that safety research attends to distributional justice while ethics research develops technically implementable frameworks.

Three Pathways to Unify AI Safety and Ethics Research

The paper’s most constructive contribution is its proposed agenda for bridging the divide through three concrete pathways.

Pathway 1: Shared Empirical Benchmarks

Currently, the empirical standards of the two fields are almost entirely disjoint. Safety relies on adversarial red-teaming to probe for catastrophic failures; ethics employs sociotechnical audits to uncover embedded biases. The authors advocate for shared evaluative benchmarks that treat alignment as a unified property. A benchmark could require a model to demonstrate robustness to jailbreaking while simultaneously satisfying group fairness constraints — forcing direct confrontation with the tradeoffs between safety and equity rather than optimizing each in isolation.

Pathway 2: Cross-Institutional Venues

The divide is sustained by institutional geography: separate conferences, funding streams, and academic departments. The authors call for structural interventions in academic venues — joint conference tracks, workshops, and doctoral consortia that require co-authorship between technical and social scientists. Collaboration should not merely be encouraged but expected, and critically, evaluated by mixed review panels that can assess both the technical rigor and normative sophistication of submitted work.

Pathway 3: Integrative Research Methodologies

The most innovative proposal involves methodological synthesis. The tools of one community can directly solve the problems of the other. Mechanistic interpretability techniques developed by AI safety researchers are powerful instruments for conducting the deep algorithmic audits that AI ethics has long called for. Conversely, participatory design methods from the ethics tradition offer a robust framework for operationalizing the “human values” that safety’s scalable oversight aims to align with. Research programs that explicitly couple these approaches would forge a genuinely unified alignment practice.

Why Unified AI Alignment Matters for Governance and Policy

The practical stakes of the safety-ethics divide extend well beyond academia. Policy frameworks, industry standards, and regulatory proposals increasingly mirror the same bifurcation. The EU AI Act focuses primarily on risk classification and compliance — reflecting ethics concerns about present harms — while organizations like the Center for AI Safety emphasize catastrophic and existential risks. Neither framework adequately addresses both dimensions simultaneously.

This fragmentation creates real-world gaps. A governance framework focused exclusively on preventing existential risk may greenlight systems that systematically discriminate against vulnerable populations. A framework focused only on fairness audits may miss emergent capabilities that create novel categories of harm. The paper argues that only a unified alignment discipline can produce governance frameworks capable of addressing the full spectrum of AI risks.

For industry practitioners, the message is equally urgent. Companies deploying AI systems increasingly need safety guarantees AND ethical accountability. Red-teaming for jailbreaks is insufficient if the model embeds discriminatory patterns in its normal operation. Fairness audits are insufficient if the model can be manipulated into dangerous behaviors through adversarial prompts. The organizations that integrate both paradigms into their development and deployment processes will be better positioned for the emerging regulatory landscape.

Building a Coherent Discipline for Human-Compatible AI

The Roytburg-Miller study delivers a sobering quantitative portrait of a field at war with itself — at precisely the moment when unity is most needed. The 83% homophily rate, the fragile bridge structure dependent on a mere 5% of papers, and the statistically significant network distance between the two communities all point to a divide that is not rhetorical but deeply structural, embedded in institutional incentives, funding streams, conference ecosystems, and career paths.

Yet the paper is ultimately optimistic. The convergences between safety and ethics are real and promising: transparency in ethics maps naturally onto interpretability in safety; accountability frameworks complement scalable oversight; fairness constraints can inform and strengthen alignment objectives. The tools already exist in both communities — what is missing is the institutional will and structural incentives to bring them together.

As the authors conclude, “achieving true alignment requires bridging the technical guarantees sought by safety research with the normative commitments advanced by ethics.” A truly safe system must be just, and a just system must be robust and controllable. Neither community can achieve its stated goals in isolation. The path forward requires not just intellectual openness but concrete structural changes: shared benchmarks, joint venues, integrative methodologies, and a generation of researchers trained to work fluently across both paradigms.

The stakes could not be higher. As AI systems grow more powerful, the window for establishing a coherent alignment discipline — one capable of producing systems that are responsible, robust, just, and safe — is narrowing. The 6,442 papers analyzed in this study represent an enormous investment of human intelligence and creativity. Imagine what that investment could produce if the two communities were actually working together.