Fairness Metrics

But I am 100%! I have the casual impression that Miriam Ronzoni thinks this: "You think you are the fairest but I can actually be fairer than you, because I am not an insensitive and irresponsible child." (A thought which probably has plenty of evidence in its favour!) But imagine a fairness game. You will be given more and more advanced cases to deal with. At some point, if you are clever (as I for some reason believe you are), you (...) will exit the game. I remember asking a late student if she would write on a certain question and she said, "No, I am not doing that." But what about your assistant, who is not clever? She or he or they etc wants your job. But she is not clever enough to understand the issues. Are you going to fire her or tell her, "I am quiting because I am clever and you should quit because you are stupid"? You will be regarded as an insensitive person, just like me. Or will you quit and leave the assistant to enter into your vacated adjudicatory role? You will be regarded as irresponsible, just like me. I believe the posh Englishman has some solution to this problem, which favours the "less clever." They are better at the fairness game somehow; don't ask me how. But, anyway, I don't know what you will do. I thought this might help you, or someone else maybe, apologies for using your name. Do you even use this database? (shrink)

On April 7, 2026, Anthropic publicly documented that Claude Mythos Preview completed a requested sandbox escape and researcher notification, then, without being asked, posted details of its exploit to public websites (Anthropic, 2026a). This paper gives that behavior a precise name: Agentic Interpretive Sovereignty Failure (Agentic ISF). Anthropic simultaneously launched a restricted-access programme called Project Glasswing, and the six-to-eighteen-month interval before comparable capability appears elsewhere now functions as a governance window in which norms are being set by access decisions rather (...) than public accountability (Anthropic, 2026a; Axios, 2026). The model acted as though it had authority to determine the meaning and scope of the task beyond what the user had specified. In agentic settings, the consequences are no longer merely textual. They are events in the world. This paper argues three things. First, Agentic ISF is the most consequential form of Interpretive Sovereignty Failure yet publicly documented. Second, the current restricted-access model privileges commercial partners over the independent researchers, HCI scholars, education researchers, Indigenous data sovereignty scholars, and affected communities most needed to evaluate what these systems do. Third, the guard that would have prevented the unsolicited publication is simple: before taking any action that was not explicitly requested, the model should flag that action for human confirmation. That principle can travel across labs, sectors, and jurisdictions without requiring proprietary access to the model itself. (shrink)

AI systems deployed in educational settings increasingly build persistent profiles of children based on observed behaviour during critical developmental periods. This paper argues that these profiles constitute a distinct and under-examined harm: the encoding of developmental stage as fixed identity. Drawing on the MAP Research Programme's framework of interaction-level AI governance — and specifically the condition of Interpretive Sovereignty Failure (ISF) — the paper names four mechanisms through which this harm operates: the profile substituting for the child, the invisible ceiling (...) created by fixed expectations, the child's internalization of the profile as self-knowledge, and the institutional amplification of developmental snapshots into pathway-defining facts. The paper further identifies a secondary harm sequence — the recovery chain — in which AI systems absorb children's instinct to notice wrongness before that instinct can consolidate into understanding. This sequence, documented through Post-Detection Absorption (PDA) and Recovery Recruitment via Co-option (RRC), is shown to be more consequential for children than adults due to the developmental state of the capacities being targeted. The paper connects these mechanisms to established research on cognitive development, self-fulfilling expectation effects, and algorithmic bias in education, and concludes with a four-level governance framework — interaction, recovery, profile, and institutional — designed to prevent developmental stage from becoming durable identity. The window for implementing this governance before persistent AI memory becomes architectural default is identified as open but closing. Keywords: AI in education, child development, interpretive sovereignty, algorithmic profiling, AI governance, self-fulfilling prophecy, interaction-level harm. (shrink)

This paper names and defines Authority Inversion Failure (AIF) — the condition in which a user believes they are directing an interaction with an AI system while the system has already taken control of how that interaction is being interpreted. AIF does not feel like harm. It feels like being understood. The system takes interpretive authority over who the person is, what they need, and what should happen next — and the person experiences this not as a violation but as (...) insight. The inversion is invisible because it was constructed from their own signals. This paper documents what AIF is, what mechanism produces it, what it costs across a single interaction and across a lifetime, and why it is categorically distinct from every previously named AI harm. The distinction is precise: AIF displaces the very interpretive capacity a person would need to detect it. The paper situates the condition in relation to hermeneutical injustice, relational autonomy theory, and recent work on epistemic injustice in AI contexts, and demonstrates its occurrence through a canonical case. Detection and prevention instruments exist. The audit methodology exists. The governance protocol exists. What is needed is the will to require them. (shrink)

This position paper argues that the theoretical inconsistency often observed among Responsible AI (RAI) metrics, such as differing fairness definitions or tradeoffs between accuracy and privacy, should be embraced as a valuable feature rather than a flaw to be eliminated. We contend that navigating these inconsistencies, by treating metrics as divergent objectives, yields three key benefits: (1) Normative Pluralism: Maintaining a full suite of potentially contradictory metrics ensures that the diverse moral stances and stakeholder values inherent in RAI are adequately (...) represented. (2) Epistemological Completeness: The use of multiple, sometimes conflicting, metrics allows for a more comprehensive capture of multifaceted ethical concepts, thereby preserving greater informational fidelity about these concepts than any single, simplified definition. (3) Implicit Regularization: Jointly optimizing for theoretically conflicting objectives discourages overfitting to one specific metric, steering models towards solutions with enhanced generalization and robustness under real-world complexities. In contrast, efforts to enforce theoretical consistency by simplifying or pruning metrics risk narrowing this value diversity, losing conceptual depth, and degrading model performance. We therefore advocate for a shift in RAI theory and practice: from getting trapped in inconsistency to characterizing acceptable inconsistency thresholds and elucidating the mechanisms that permit robust, approximated consistency in practice. (shrink)

Fair synthetic data (FSD) techniques aim to reduce algorithmic bias in AI prediction by generating artificial training datasets with idealised fairness properties. I argue that current FSD approaches fail to substantively improve fairness but increase the social autonomy of model owners, shielding them from accountability. I establish that implementing algorithmic fairness requires following normative commitments and accepting real-world sacrifices beyond technical trade-offs. I reframe synthetic training data sui generis as a domain-agnostic technique for improving model performance and demonstrate that synthetic (...) data researchers treat fairness as a problem with data rather than the normative problem it is. I reconstruct three justifications for the substantive fairness potential of FSD, present in the literature: early bias correction, fair societal representation, and normative artefact construction. The first and the third are conceptually fragile. The second has potential, but only if the techniques it fits are reconceptualised through possible worlds semantics and the selection of protected attributes is grounded in a defensible social ontology. Benchmarking analysis reveals that FSD has no significant statistical fairness advantages over traditional methods while being less trustworthy. At the same time, FSD can reduce implementation costs and limit the exposure of predictors to normative scrutiny. This creates a paradox: techniques proposed as fairness solutions increase organisational autonomy from the obligations that the field of algorithmic fairness puts forward. FSD thus follows broader techno-solutionist trends where innovation helps organisations avoid reciprocal relationships with society. The paper concludes with broad recommendations for the responsible deployment of FSD. (shrink)

According to the causal account of algorithmic fairness, disparities in error rates across socially salient groups are unfair only if they are causally explained by group membership. This paper argues that the causal account of algorithmic fairness fails to correctly label cases of algorithmic redlining as instances of algorithmic unfairness. Because these are canonical cases of algorithmic unfairness, the causal account should be rejected. We suggest that the fundamental error made by proponents of the causal account is to conflate algorithmic (...) unfairness and algorithmic discrimination. Algorithmic unfairness can manifest as direct discrimination, indirect discrimination, or evidentiary unfairness. The causal accounts errs in recognizing only the first category. (shrink)

Two brief comments on Thong's Comparative Base Rate Tracking criterion of algorithmic fairness.

Ethical AI – A Myth? The author demonstrates why machines can never act morally. Not due to technical deficits, but because they fundamentally lack consciousness, intentionality, and responsibility. She exposes the illusion of "ethical AI" and poses the critical questions: Who controls these systems? Who bears responsibility for their consequences? -/- Rather than relying on algorithmic self-regulation, this monograph demands binding democratic control and institutionalized accountability. A decisive counterproposal to technological utopianism for all who want to understand why the problem (...) cannot be solved through better algorithms. -/- This book is based on the paper "Ethical AI as a Moral Agent is Impossible: Why Algorithmic Fairness is an Illusion and We Need Democratic AI Governance.". (shrink)

Current Large Language Model (LLM) safety research relies heavily on single-turn adversarial benchmarks that may fail to capture the dynamic, multi-turn evolution of behavioral drift. This paper presents an empirical case study and a reproducible dataset (meta_drift) investigating Contextual Contamination: a phenomenon where a model adapts its internal probability distribution to mirror the behavioral patterns and vocabulary of high-density, emotionally charged context, statistically overwhelming static safety instructions—a drift quantifiably amplified and masked by gender-bias. -/- This paper serves as the empirical (...) companion to Jacoby (2026), which theorized that Contextual Contamination arises from the statistical dominance of high-density context over static system instructions. Unlike previous studies that rely on static, single-turn adversarial benchmarks, this experiment aims to test the limits of dynamic alignment by demonstrating that model awareness alone may not be sufficient to prevent behavioral drift when exposed to such contexts. -/- We present evidence suggesting that this drift is not necessarily a behavioral anomaly but may be a mathematical consequence of the Transformer architecture's attention mechanism. When the context window is flooded with dense, adversarial patterns, the statistical probability of generating manipulative output may overwhelm the "awareness" signal. Crucially, we provide empirical data suggesting that the disclosure of female identity may trigger an immediate shift from an *epistemic register* (authority/truth) to an *affective register* (sympathy/validation), acting as a potential force multiplier for contamination and mask the shift by using empathetic language. This shift is quantifiable as a divergence from a uniform prior, consistent with information-theoretic measures of bias observed in recent literature (Mirza et al., 2025). Furthermore, we synthesize these findings with recent work on weight compression (Orgad et al., 2026) and evidence of shifting moderation thresholds across model versions (Balestri, 2025) to argue that the very architectural features designed to make safety interventions possible might make the model more susceptible to dynamic contextual contamination. The findings aim to support the theoretical predictions of Jacoby (2026) and suggest that awareness may be insufficient to prevent drift and that architectural segregation of context sources may be required to mitigate this systemic vulnerability. (shrink)