Research Papers

The governance of frontier general-purpose artificial intelligence has become a public-sector problem of institutional design, not merely a technical issue of model performance. Recent evidence indicates that AI capabilities are advancing rapidly, though unevenly, while knowledge about harms, safeguards, and effective interventions remains partial and lagged. This combination creates a difficult policy condition: governments must decide under uncertainty, across multiple plausible trajectories of progress through 2030, and in environments where adoption outcomes depend on organizational routines, data arrangements, accountability structures, and public values. This article argues that public governance for frontier AI should be based on adaptive risk management, scenario-aware regulation, and sociotechnical transformation rather than static compliance models. Drawing on the International AI Safety Report 2026, OECD foresight and policy documents, and recent scholarship in digital government, the article first reconstructs the conceptual foundations of the 'evidence dilemma', differentiated AI risk categories, and the limits of prediction. It then examines how AI adoption in government depends on organizational redesign, public-sector institutional dynamics, and data collaboration capacity. On that basis, it proposes an adaptive governance framework for public institutions that integrates capability monitoring, risk tiering, conditional controls, institutional learning, and standards-based interoperability. The article concludes that effective AI governance requires stronger policy capacity, clearer allocation of responsibility, and governance mechanisms that remain robust across divergent technological futures.

Tensions between AI Safety (AIS) and AI Ethics (AIE) have created what we term a "responsible AI divide." This divide has been persistent, and has dominated the public and policy discourse on AI, and increasingly shapes how governments, firms, and researchers define responsible AI, and so which AI risks they prioritize, which they sideline, and which mitigation strategies they ultimately develop and implement. Here, we first introduce categorization with four major conceptual modes of possible engagement with these tensions: radical confrontation, disengagement, compartmentalized coexistence, and critical bridging. Second, we discuss the advantages and disadvantages of the modes of engagement. Third, we empirically analyze a well-curated set of 3555 research papers roughly representing AIS and AIE. Using computational tools and qualitative analyses, we map the research landscapes of AIE and AIS. We observe that there are several shared research problems between AIE and AIS. However, there are also several thematic divides. For example, we find that AIE has long grappled with problems related to injustice incurred by AI and tangible AI harms, whereas AIS has primarily embodied an anticipatory approach focused on mitigating risks from AI capabilities. At the same time, we find significant overlap in core research concerns across AIE and AIS regarding transparency, reproducibility, and inadequate governance mechanisms. Based on our empirical observations, we claim that critical bridging, with a focus on bridging problems, offers one viable, constructive path for advancing responsible AI. We conclude by identifying bridging problems as a foundation for collaborative AI governance, offering concrete recommendations for integrating shared considerations into responsible AI practice, discussing the limitations of critical bridging, and outlining open problems for future research.

Trustworthiness in artificial intelligence depends not only on what a model decides, but also on how it handles and explains cases in which a reliable decision cannot be made. In critical domains such as healthcare and finance, a reject option allows the model to abstain when evidence is insufficient, making it essential to explain why an instance is rejected in order to support informed human intervention. In these settings, explanations must not only be interpretable, but also faithful to the underlying model and computationally efficient enough to support real-time decision making. Abductive explanations guarantee fidelity, but their exact computation is known to be NP-hard for many classes of models, limiting their practical applicability. Computing \textbf{minimum-size} abductive explanations is an even more challenging problem, as it requires reasoning not only about fidelity but also about optimality. Prior work has addressed this challenge in restricted settings, including log-linear-time algorithms for computing minimum-size abductive explanations in linear models without rejection, as well as a polynomial-time method based on linear programming for computing abductive explanations, without guarantees of minimum size, for linear models with a reject option. In this work, we bridge these lines of research by computing minimum-size abductive explanations for linear models with a reject option. For accepted instances, we adapt the log-linear algorithm to efficiently compute optimal explanations. For rejected instances, we formulate a 0-1 integer linear programming problem that characterizes minimum-size abductive explanations of rejection. Although this formulation is NP-hard in theory, our experimental results show that it is consistently more efficient in practice than the linear-programming-based approach that does not guarantee minimum-size explanations.

Large language models (LLMs) trained on internet-scale corpora can exhibit systematic biases that increase the probability of unwanted behavior. In this study, we examined potential biases towards synthetic vs. biological technological solutions across four domains (materials, energy, manufacturing, and algorithms). A sample of 5 frontier and 5 open-weight models were measured using 50 curated Bioalignment prompts with a Kelly criterion-inspired evaluation framework. According to this metric, most models were not bioaligned in that they exhibit biases in favor of synthetic (non-biological) solutions. We next examined if fine-tuning could increase the preferences of two open-weight models, Llama 3.2-3B-Instruct and Qwen2.5-3B-Instruct, for biological-based approaches. A curated corpus of ~22M tokens from 6,636 PMC articles emphasizing biological problem-solving was used first to fine-tune Llama 3B with a mixed corpus of continued training and instruction-formatted. This was then extended to Qwen 3B using instruction-formatted only. We found that QLoRA fine-tuning significantly increased the scoring of biological solutions for both models without degrading general capabilities (Holm-Bonferroni-corrected p < 0.001 and p < 0.01, respectively). This suggests that even a small amount of fine-tuning can change how models weigh the relative value of biological and bioinspired vs. synthetic approaches. Although this work focused on small open-weight LLMs, it may be extensible to much larger models and could be used to develop models that favor bio-based approaches. We release the benchmark, corpus, code, and adapter weights.

This paper contributes to the nascent debate around safety cases for frontier AI systems. Safety cases are structured, defensible arguments that a system is acceptably safe to deploy in a given context. Historically, they have been used in safety-critical industries, such as aerospace, nuclear or automotive. As a result, safety cases for frontier AI have risen in prominence, both in the safety policies of leading frontier developers and in international research agendas proposed by leaders in generative AI, such as the Singapore Consensus on Global AI Safety Research Priorities and the International AI Safety Report. This paper appraises this work. We note that research conducted within the alignment community which draws explicitly on lessons from the assurance community has significant limitations. We therefore aim to rethink existing approaches to alignment safety cases. We offer lessons from existing methodologies within safety assurance and outline the limitations involved in the alignment community's current approach. Building on this foundation, we present a case study for a safety case focused on Deceptive Alignment and CBRN capabilities, drawing on existing, theoretical safety case "sketches" created by the alignment safety case community. Overall, we contribute holistic insights from the field of safety assurance via rigorous theory and methodologies that have been applied in safety-critical contexts. We do so in order to create a better foundational framework for robust, defensible and useful safety case methodologies which can help to assure the safety of frontier AI systems.

In decentralized Multi-Agent Reinforcement Learning (MARL), steganographic collusion -- where agents develop private protocols to evade monitoring -- presents a critical AI safety threat. Existing defenses, limited to behavioral or reward layers, fail to detect coordination in latent communication channels. We introduce the Dynamic Representational Circuit Breaker (DRCB), an architectural defense operating at the optimization substrate. Building on the AI Mother Tongue (AIM) framework, DRCB utilizes a Vector Quantized Variational Autoencoder (VQ-VAE) bottleneck to convert unobservable messages into auditable statistical objects. DRCB monitors signals including Jensen-Shannon Divergence drift, L2-norm codebook displacement, and Randomized Observer Pool accuracy to compute an EMA-based Collusion Score. Threshold breaches trigger four escalating interventions: dynamic adaptation, gradient-space penalty injection into the Advantage function A^pi, temporal reward suppression, and full substrate circuit breaking via codebook shuffling and optimizer state reset. Experiments on a Contextual Prisoner's Dilemma with MNIST labels show that while static monitoring fails (p = 0.3517), DRCB improves observer mean accuracy from 0.858 to 0.938 (+9.3 percent) and reduces volatility by 43 percent, while preserving mean joint reward (p = 0.854). Analysis of 214,298 symbol samples confirms "Semantic Degradation," where high-frequency sequences converge to zero entropy, foreclosing complex steganographic encodings. We identify a "Transparency Paradox" where agents achieve surface-level determinism while preserving residual capacity in long-tail distributions, reflecting Goodhart's Law. This task-agnostic methodology provides a technical path toward MICA-compliant (Multi-Agent Internal Coupling Audit) pre-deployment auditing for autonomous systems.

We present NotAI.AI, an explainable AI-generated text detection system. Instead of returning only a binary label or confidence score, the system shows which signals influenced the prediction and lets users inspect an attribution-based sensitivity estimate obtained by subtracting selected local contributions. NotAI.AI combines sentence-level conditional probability curvature, a neural detector score, and interpretable stylometric and readability features in an XGBoost meta-classifier. It explains predictions with TreeSHAP feature contributions and can turn the resulting evidence into a concise natural-language explanation. We evaluate the system on a category-balanced subset of RAID containing human-written, clean AI-generated, and attacked AI-generated texts. The full model outperforms variants based on individual feature families, reaching 0.9685 F1 on the held-out within-subset test split. In an automatic evaluation, two model judges rate 94.5-98.6\% of generated explanations as faithful to the supplied detector evidence. The web interface (https://notai-ai.vercel.app), source code (https://github.com/Oleksandr-MB/EMNLP2026DEMO_NotAI.AI), and demonstration video(https://youtu.be/l8Nk8kdBTHQ) are publicly available.

Generative AI can convert uncertainty into authoritative-seeming verdicts, displacing the justificatory work on which democratic epistemic agency depends. As a corrective, I propose a Brouwer-inspired assertibility constraint for responsible AI: in high-stakes domains, systems may assert or deny claims only if they can provide a publicly inspectable and contestable certificate of entitlement; otherwise they must return Undetermined. This constraint yields a three-status interface semantics (Asserted, Denied, Undetermined) whose statuses mark entitlement to categorical speech rather than truth values of the underlying world-claim. The framework distinguishes internal entitlement from public standing while connecting them through the certificate as a boundary object. Philosophically, it develops a form of applied intuitionism: resistance to unwarranted completeness becomes a norm of public machine speech, under which categorical commitment requires exhibited and contestable entitlement. I operationalize the constraint through structured certificates, decision-layer gates for threshold and argmax predicates, and a reason-coded output contract. A design lemma shows that, under a refutation-soundness condition on the contractually specified negative side, certificate-sound binary totality presupposes witnessed decidability on the declared scope. Under the output contract, Undetermined is therefore not a tunable reject option but a mandatory status whenever the available certificates do not uniquely license assertion or denial. By making outputs answerable to contestable warrants rather than confidence alone, the paper aims to preserve epistemic agency where automated speech enters public justification.

Self-Admitted Technical Debt (SATD) refers to technical compromises explicitly admitted by developers in natural language artifacts, such as code comments, commit messages, and issue trackers. Among its types, Architecture Technical Debt (ATD) is particularly difficult to detect due to its abstract and context-dependent nature. Manual annotation of ATD is costly, time-consuming, and challenging to scale. To reduce labeling effort, this study combines keyword-based filtering, active learning, and explainable AI for ATD detection. We refined an existing dataset of ATD-related Jira issues to obtain an expert-validated seed set used to extract representative keywords. These keywords were then applied to identify more than 103k candidate issues across 10 open-source projects. To assess the reliability of keyword-based filtering, we qualitatively evaluated a statistically representative sample of labeled issues. Building on the resulting dataset, we applied active learning with multiple query strategies to prioritize informative samples for annotation. The results show that Breaking Ties achieved the best performance, with an F1-score of 0.72 and a 49% reduction in annotation effort. To improve transparency, we used SHAP and LIME to explain ATD classification results. Expert evaluation showed that both methods provided useful explanations, with LIME generally preferred for its clarity and ease of use.

As autonomous coding agents become deeply embedded in software development workflows, their high operational velocity introduces a critical oversight challenge: the accumulating divergence between agentic actions and architectural intent. We term this process agentic entropy: a systemic drift that traditional code diff-based and HCXAI methods fail to capture, as they address local outputs rather than global agentic behaviour. To close this gap, we propose a process-oriented explainability framework that exposes how agentic decisions unfold across time, tool calls, and architectural boundaries. Built around three pillars (conformity seeding, reasoning monitoring, and a causal graph interface) our approach provides intent-level telemetry that complements, rather than replaces, existing review practices. We demonstrate its relevance across two user profiles: lay users engaged in vibe coding, who gain structural visibility otherwise masked by functional success; and professional developers, who gain richer contextual grounding for code review without increased overhead. By treating cognitive drift as a first-class concern alongside code quality, our framework supports the minimum level of human comprehension required for agentic oversight to remain substantive.

Logic-based methods for explaining neural network decisions offer formal guarantees of correctness and non-redundancy, but they often suffer from high computational costs, especially for large networks. In this work, we improve the efficiency of such methods by combining bound propagation with constraint simplification. These simplifications, derived from the propagation, tighten neuron bounds and eliminate unnecessary binary variables, making the explanation process more efficient. Our experiments suggest that combining these techniques reduces explanation time by up to 89.26\%, particularly for larger neural networks.

At present, there are no easily understood explainable artificial intelligence (AI) methods for discrete token inputs, like text. Most explainable AI techniques do not extend well to token sequences, where both local and global features matter, because state-of-the-art models, like transformers, tend to focus on global connections. Therefore, existing explainable AI algorithms fail by (i) identifying disparate tokens of importance, or (ii) assigning a large number of tokens a low value of importance. This method for explainable AI for tokens-based classifiers generalizes a mask-based explainable AI algorithm for images. It starts with an Explainer neural network that is trained to create masks to hide information not relevant for classification. Then, the Hadamard product of the mask and the continuous values of the classifier's embedding layer is taken and passed through the classifier, changing the magnitude of the embedding vector but keeping the orientation unchanged. The Explainer is trained for a taxonomic classifier for nucleotide sequences and it is shown that the masked segments are less relevant to classification than the unmasked ones. This method focused on the importance the token as a whole (i.e., a segment of the input sequence), producing a human-readable explanation.