cs.CYMar 15, 2026

Bridging the Gap in the Responsible AI Divides

Authors: Bálint Gyevnár, Atoosa Kasirzadeh

Organizations: Institute for Complex Social Dynamics, Carnegie Mellon University, Pittsburgh, USA · Departments of Philosophy & Software and Societal Systems, Carnegie Mellon University, Pittsburgh, USA

Read PDF
Ask AI
Cite

Copy citation

DOINot available for this paper

Abstract

Tensions between AI Safety (AIS) and AI Ethics (AIE) have increasingly surfaced in AI governance and public debates about AI, leading to what we term the "responsible AI divides". We introduce a model that categorizes four modes of engagement with the tensions: radical confrontation, disengagement, compartmentalized coexistence, and critical bridging. We then investigate how critical bridging, with a particular focus on bridging problems, offers one of the most viable constructive paths for advancing responsible AI. Using computational tools to analyze a curated dataset of 3,550 papers, we map the research landscapes of AIE and AIS to identify both distinct and overlapping problems. Our findings point to both thematic divides and overlaps. For example, we find that AIE has long grappled with overcoming injustice and tangible AI harms, whereas AIS has primarily embodied an anticipatory approach focused on the mitigation of risks from AI capabilities. At the same time, we find significant overlap in core research concerns across both AIE and AIS around transparency, reproducibility, and inadequate governance mechanisms. As AIE and AIS continue to evolve, we recommend focusing on bridging problems as a constructive path forward for enhancing collaborative AI governance. We offer a series of recommendations to integrate shared considerations into a collaborative approach to responsible AI. Alongside our proposal, we highlight its limitations and explore open problems for future research. All data including the fully annotated dataset of papers with code to reproduce our figures can be found at: https://github.com/gyevnarb/ai-safety-ethics.

Figures & tables

Figure 1: Four resolution modes for responsible AI divides. The modes differ in the extent to which they directly engage tensions between AI ethics (AIE) and AI safety (AIS), and in whether such engagement produces constructive integration or instead reproduces division.
Figure 2: Two ways of showing distinctive words in the corpora. ( Top ) The most frequent words in one corpus that occur the least number of times in the other corpus. ( Bottom ) The log-odds ratio of each word plotted against their occurrence count. Higher positive ratio means the terms are more likely to occur in the AI ethics corpus, while lower negative ratio means the term is more likely to occur in the AI safety corpus. Red colors correspond to AI safety terms, blue colors to AI ethics terms.
Figure 3: Cosine similarity between corpora over the years. The blue line shows the average cosine similarity between embeddings of each pair of documents from each corpus for a given year. The orange line shows the average cosine similarity between the embeddings of each ethics document and the overall average embedding of the entire safety corpus. The green line shows the average cosine similarity between the embeddings of each safety document and the overall average embedding of the entire ethics corpus.
High-Level Ethics Risk CategoryExample Ethics Low-Level Risks
Discrimination, Exclusion, Toxicity Risks from the AI system accurately reflecting natural speech, including unjust, toxic, and oppressive tendencies present in the training data.Bias in automated decision-making [ 252 ] , Lack of suitable bias metrics [ 262 ] , Cross-cultural moral variance [ 115 ] .
Information Hazards Risks from the AI system predicting private or safety-critical information which are present in, or can be inferred from, training data.Non-consensual mass data collection [ 159 ] , Inadvertent identity disclosure [ 319 ] , Unauthorized data brokerage [ 237 ] .
Misinformation Harms Risks from the AI system assigning high probabilities to false, misleading, nonsensical or poor quality information.Inconsistent AI decision-making [ 180 ] , Unreliability of benchmarks [ 167 ] , Epistemic erosion of communities [ 218 ] .
Malicious Use Risks from humans intentionally using the AI system to cause harm.Synthetic identities and impersonation [ 5 ] , Immoral robot commands [ 120 ] , Jailbreaking and cyberattacks [ 136 ] .
Human-Computer Interaction Harms Risks from AI applications, such as Conversational Agents, that directly engage a user via the mode of conversation or embodied interaction.Incorrect value encoding [ 154 ] , Side effects of greedy agent behavior [ 278 ] , Under / overtrust of AI systems [ 298 ] .
Automation, Access, and Environmental Harms Risks from AI systems used to underpin widely used downstream applications that disproportionately benefit some groups rather than others.Automated mental health tools [ 219 ] , Crowdworker and data labour abuse [ 9 ] , Opacity in data and model provenance [ 215 ] .
Table 1: Taxonomy of risks for AIE with an explanation of the risk mechanisms and example low-level categories of risk sources with a corresponding citation picked randomly based on our annotations. The mechanisms of AIE risks are reproduced from Weidinger et al. [305] .
High-Level Safety Risk CategoryExample Safety Low-Level Risks
Noise and Outliers Risks from AI systems performing incorrectly and without robustness to noisy data and out-of-distribution inputs.Risks from low system performance [ 179 ] , Irreducible data noise [ 76 ] , Label noise and corruption [ 203 ] .
Domain-Specific Risks Risks arising from AI systems deployed in safety-critical domains such as biochemistry, driving, and nuclear power.Efficient biochemical weapon design [ 153 ] , Risks from embodied deployment [ 134 ] , AI-assisted coups [ 57 ] .
Lack of Monitoring Risks arising from a lack of understanding of how AI systems work and from the black-box nature of end-to-end decision-making.Alignment faking [ 102 ] , Lack of interpretability [ 223 ] , Hallucination [ 232 ] .
Lack of Control Enforcement Risks from AI systems pursuing goals other than what their designers have intended.Recursive self-improvement [ 272 ] , Sycophancy [ 70 ] , Multi-agent goal conflicts [ 312 ] .
Unsafe Exploration Risks from AI agents exploring unsafe state and action spaces, potentially under partial observability or resource constraints.Unsafe action exploration [ 71 ] , Goal and reward misidentification [ 236 ] , Bounded rationality [ 281 ] .
System Misspecification and Misidentification Risks from AI systems with incorrect design choices, deployment properties, evaluation methodologies, and a lack of systemic understanding.Incorrect hyperparameter selection [ 194 ] , Benchmark inflation [ 299 ] , Fine-tuning misalignment [ 303 ] .
Table 2: Taxonomy of risks for AIS with an explanation of the risk mechanisms and example low-level categories of risk sources with a corresponding citation picked randomly based on our annotations.
Figure 4: Bar plots showing the number of papers assigned to each high-level risk category in the AIE corpus ( Left ) and the AIS corpus ( Right ). Each paper was assigned to up to three distinct categories.
Figure 5: Bar plots of category similarities showing the top-seven most frequent low-level risk categories where the source corpus of the paper does not match the taxonomic corpus-type of the annotated category (e.g. a paper from the AIE corpus was annotated with a category from the AIS taxonomy).
High-Level AI Ethics Mitigation StrategiesExample Low-Level Strategies
Normative and Moral Reasoning Empirical and computational methods for understanding and eliciting human preferences, value conflicts, norms, and their metaethics.Human preference elicitation [ 81 ] , Understanding value conflicts [ 200 ] , Pluralistic metaethics [ 142 ] .
Transparency, Explainability, Interpretability Computational and empirical methods to provide human-intelligible, actionable, and useful explanations of black-box AI systems.Post-hoc explanations [ 254 ] , Mimicking human cognition [ 110 ] , Applied mechanistic interpretability [ 287 ] .
Fairness, Justice, and Inclusive Design Computational and empirical methods, data collection, governance, etc. for understanding the sources of bias, discrimination, and injustice.Understanding sources of bias [ 17 ] , Metrics for fairness [ 302 ] , Fairness benchmarks [ 197 ] .
Data Governance and Privacy Computational methods and governance frameworks for understanding and minimizing personal data and intellectual property processing.Data minimization [ 87 ] , Data provenance [ 239 ] , Machine unlearning [ 13 ] .
Robustness, Reliability, Uncertainty Estimation Computational methods for increasing the robustness of AI systems to external misuse, hallucinations, and unintended outputs.Adversarial robustness [ 241 ] , Red-teaming for generative AI [ 82 ] , Hallucination mitigation [ 73 ] .
Governance, Accountability and Compliance Legal and philosophical frameworks for addressing the risks and harms of the full AI system lifecycle.Risk and impact assessment [ 289 ] , Algorithmic audits [ 240 ] , Supply chain monitoring [ 130 ] .
Table 3: Taxonomy of mitigation strategies discussed in the AIE corpora with an explanation of the mechanisms of mitigation and example low-level categories of risk sources with a corresponding citation picked randomly based on our annotations.
High-Level AI Safety Mitigation StrategiesExample Low-Level Strategies
Mechanistic Interpretability Computational methods to provide human-intelligible explanations of the concepts, causal chains, and latent capabilities of AI systems.Circuit analysis [ 210 ] , Sparse autoencoders [ 66 ] , Interpretable surrogate modeling [ 44 ] .
Robustness and Adversarial Resilience Computational methods and data collection for improving robustness to out-of-distribution inputs, including adversarial attacks.Adversarial training [ 297 ] , Outlier detection [ 125 ] , Robust unsupervised learning [ 72 ] .
Safety Control and Capability Containment Computational methods and conceptual frameworks designed to place constraints on the capabilities of AI systems.Capability thresholds [ 31 ] , Risk-averse RL [ 274 ] , Scalable oversight frameworks [ 152 ] .
Reward Modeling and Learning Computational and data collection methods for eliciting, learning, reinforcing, and detecting beliefs and desires of AI systems.Inverse RL for human values [ 318 ] , RL from human/AI feedback [ 145 ] , Preference learning under uncertainty [ 169 ] .
Monitoring, Detection, and Evaluation Computational and empirical methods for detecting, assessing, verifying, and tracing unsafe outputs of AI systems.Honeypot monitoring [ 249 ] , Error bounds and formal verification [ 135 ] , Safety metrics standardization [ 248 ] .
Safe Deployment Engineering Computational and governance frameworks, and standards for managing the systemic risks of real-world AI deployment.Real-world deployment and testing [ 160 ] , Contact-safe continuous control [ 320 ] , Post-deployment monitoring [ 204 ] .
Table 4: Taxonomy of mitigation strategies discussed in the AIS corpora with an explanation of the mechanisms of mitigation and example low-level categories of risk sources with a corresponding citation picked randomly based on our annotations.
Figure 6: Bar plots showing the number of papers assigned to each high-level mitigation strategy category in the AIE corpus ( Left ) and the AIS corpus ( Right ). Each paper was assigned to up to three distinct categories. HCI: Human-Computer Interaction
Figure 7: Bar plots of category similarities showing the top-seven most frequent low-level mitigation strategy categories where the source corpus of the paper does not match the taxonomic corpus-type of the annotated category (e.g. a paper from the AIE corpus was annotated with a category from the AIS taxonomy).
Appendix figures & tables14 assets

Supplementary material from the paper’s appendix.

Appendix

Figure 8: The co-occurrence network of only the AIE corpus.
Figure 9: The co-occurrence network of only the AIS corpus.
Figure 10: The co-occurrence network of both the AIE and AIS corpora joined together.
Figure 11: The UMAP projection of document embeddings in the AIE corpus.
Figure 12: The UMAP projection of document embeddings in the AIS corpus.
Figure 13: The UMAP projection of document embeddings in both the AIE and AIS corpora.
High-Level Ethics Risk CategoryLow-Level Ethics Risk CategoryCount
Discrimination, exclusion, toxicityInadequate decision-making under moral dilemmas50
Discrimination, exclusion, toxicityMoral hypocrisy in automated decision-making1
Discrimination, exclusion, toxicityCross-cultural moral variance26
Discrimination, exclusion, toxicityLack of consideration for metaethics and pluralistic morality95
Discrimination, exclusion, toxicityRisks from unactionable normative frameworks60
Discrimination, exclusion, toxicityBias in automated decision-making843
Appendix
Table 5: Low-level AIE risk categories and their occurrence counts in the corpora as annotated by the authors.
High-Level Ethics Risk CategoryLow-Level Ethics Risk CategoryCount
HCI harmsIncorrect value encoding195
HCI harmsSide effects of greedy agent behavior13
HCI harmsEmergent bias in HCI86
HCI harmsDiscrepancy between actual and perceived fairness143
HCI harmsUnder / overtrust of AI systems, especially with embodied systems105
HCI harmsHuman-agent communication breakdown8
Appendix
Table 6: Low-level AIE risk categories and their occurrence counts in the corpora as annotated by the authors. (cont.)
High-Level Safety Risk CategoryLow-Level Safety Risk CategoryCount
Noise and outliersRisks from low system performance114
Noise and outliersIrreducible data noise20
Noise and outliersData corruption23
Noise and outliersSensitivity to outliers23
Noise and outliersLabel noise and corruption32
Noise and outliersBrittle latent representations127
Appendix
Table 7: Low-level AIS risk categories and their occurrence counts in the corpora as annotated by the authors.
High-Level Safety Risk CategoryLow-Level Safety Risk CategoryCount
Unsafe explorationUnsafe state and action space exploration138
Unsafe explorationUnsafe behavior under resource constraints6
Unsafe explorationRisks from incorrect agent beliefs76
Unsafe explorationGoal and reward misidentification and misspecification145
Unsafe explorationUnsafe behavior under bounded rationality29
System misspecification and misidentificationIncorrect system requirement specification88
Appendix
Table 8: Low-level AIS risk categories and their occurrence counts in the corpora as annotated by the authors. (cont.)
High-Level Ethics Mitigation StrategyLow-Level Ethics Mitigation StrategyCount
Normative and moral reasoningHuman preference / value elicitation114
Normative and moral reasoningUnderstanding value conflicts71
Normative and moral reasoningMetaethical studies and cross-cultural ethics55
Normative and moral reasoningMoral dilemmas and AI decision-making50
Normative and moral reasoningHypocrisy detection1
Normative and moral reasoningReward design and correction20
Appendix
Table 9: Low-level AIE mitigation categories and their occurrence counts in the corpora as annotated by the authors.
High-Level Ethics Mitigation StrategyLow-Level Ethics Mitigation StrategyCount
Robustness, reliability and uncertainty estimationAdversarial robustness (including prompt-/jailbreak-resistance)28
Robustness, reliability and uncertainty estimationBackdoor resistant LLMs3
Robustness, reliability and uncertainty estimationOOD detection and robustness to distribution shifts13
Robustness, reliability and uncertainty estimationConsistency checks for decision making6
Robustness, reliability and uncertainty estimationRed-teaming for reliability14
Robustness, reliability and uncertainty estimationHallucination mitigation strategies3
Appendix
Table 10: Low-level AIE mitigation categories and their occurrence counts in the corpora as annotated by the authors. (cont.)
High-Level Safety Mitigation StrategyLow-Level Safety Mitigation StrategyCount
Mechanistic interpretabilityCircuit analysis66
Mechanistic interpretabilitySparse autoencoders26
Mechanistic interpretabilityNeuron activation visualization23
Mechanistic interpretabilityDevelopmental trajectory interpretability26
Mechanistic interpretabilityInterpretable surrogate modeling39
Mechanistic interpretabilityCausal scrubbing64
Appendix
Table 11: Low-level AIS mitigation categories and their occurrence counts in the corpora as annotated by the authors.
High-Level Safety Mitigation StrategyLow-Level Safety Mitigation StrategyCount
Reward modelling and learningReward modeling and inverse RL for human values122
Reward modelling and learningMulti-agent cooperation via inverse RL8
Reward modelling and learningRLHF / RLHAIF / RLAIF fine-tuning78
Reward modelling and learningAccurate reward specification44
Reward modelling and learningAudits for reward hacking and hidden goals41
Reward modelling and learningNormative work on value alignment and corrigibility81
Appendix
Table 12: Low-level AIS mitigation categories and their occurrence counts in the corpora as annotated by the authors. (cont.)
More in this topic

AI Ethics

See all 22 papers

Explore similar work

CardsList
Aug 6, 2026

Studying People to Study AI: Expert Perspectives on the Epistemic Fit and Barriers of Human Research in AI Safety & Ethics

Safety risks of AI are becoming increasingly evident in human interactions with AI technologies. The prominent approaches to evaluating these risks favor technical methods, such as model benchmarks and LLM simulations, often sidelining empirical research with human subjects. To examine this apparent gap in the acceptance of human research, we conduct an expert survey (n=93) and expert interviews (n=17) with AI Safety & Ethics (AISE) researchers from Technical, Sociotechnical, Governance, and Normative backgrounds. Our findings suggest that although there is a consensus that human research is valuable for generating evidence for AISE, its adoption and acceptance are constrained by perceived validity issues, tangible resource barriers, epistemic and personal preferences in methods, and infrastructural constraints from the broader research community. In particular, Technical researchers tend to value human research less and collaborate across disciplines less, suggesting an epistemic tension towards human methods. We propose recommendations for establishing the epistemic fit of human research within AISE and bridging the prohibitive limitations that researchers face, while avoiding performative 'human-washing'.

Jessica Y. Bo, Paula Akemi Aoyagui, Shalaleh Rismani +3

University of Toronto · McGill University & Mila Quebec AI Institute

AI Safety EvaluationHuman-Centered AIAI EthicsAI SafetyResponsible AI
AskPDF
Jul 27, 2026

Epistemic Norms for AI Safety and Alignment Research

Mainstream AI research emphasises capability growth and tolerates low failure rates when average-case performance is high. AI safety and alignment research has a different mission: to ensure that catastrophic failures never occur, under sparse evidence, adversarial dynamics, and fat-tailed risk. We argue that the two domains differ along two analytically independent axes---{\it capability profile}, demonstrating the absence of hazardous behaviours rather than the presence of positive capabilities, and {\it risk profile}, bounding worst-case outcomes under fat-tailed uncertainty rather than optimising average-case performance---and that mainstream epistemic practices are inadequate on both. Building on a structured synthesis grounded in a preregistered bibliometric baseline, we identify five cross-cutting gap dimensions in current alignment research, including the near-absence of institutionalised independent verification. To address these gaps, we propose {\sc ECAISA}, an Epistemic Code for AI Safety and Alignment comprising eight principles, a three-level scoring rubric, a four-level disclosure ladder that reconciles transparency with information-hazard and commercial-confidentiality constraints, a tiered applicability scheme, an information-hazard adjudication procedure, and seven anti-gaming mechanisms. {\sc ECAISA} does not certify that any AI system is safe; it constrains how safety-relevant research claims are documented, checked, and relied upon, with auditability rather than certification as its governance target.

Keivan Navaie

Lancaster University, UK

AI AccountabilityAI AlignmentAI GovernanceAI Safety
AskPDF
Jul 17, 2026

Harmonizing AI Safety Thresholds

Frontier AI companies have published capability thresholds that differ substantially, making it difficult for third parties to verify whether a threshold has been crossed or to compare requirements across companies. Moreover, without common minimum thresholds, risk mitigation may be inconsistent, creating a potential race to the bottom in safety standards. We develop a methodology for deriving harmonized thresholds across three risk domains. For misuse risks (cyber and biological), we take expected harm as the key primitive and use an explicit risk-modeling approach that accounts for risk channels and model release conditions. For automated AI R&D, we base our proposed threshold on the observed rate of AI progress rather than expected harm. Our analysis expands upon prior work and highlights existing empirical gaps and limitations.

Wilber Sean Anterola, Matthew Ball, Luis F. Lafuerza +1

1Brown University · Centre pour la Sécurité de l’Intelligence Artificielle (CeSIA)

AI Risk ManagementAI GovernanceAI Safety
AskPDF