We treat the internals of generative models as mechanistic objects rather than black boxes. We introduce \textbf{Attribution Graphs} (AGs), which extend GradCAM++ to circuit-level representations, and \textbf{Causal Probing}, a do-calculus intervention method for identifying causal latent structures, enabling detection and correction of spurious correlations, demographic biases, and misaligned decision circuits during training. We further propose the \textbf{Cognitive Alignment Score (CAS)}, quantifying agreement between model-internal representations and human concepts, a \textbf{saliency-first privacy mechanism} sharing only thresholded attribution nodes, a bias-aware regularizer aligning subgroup statistics, and a Reveal-to-Revise loop integrating attribution signals into parameter updates without separate fine-tuning. Evaluated on CelebA, FairFace, Jigsaw, and HateXplain, our method achieves \textbf{94.1%} accuracy, \textbf{92.3%} macro F1, \textbf{79.4%} IoU-XAI, and \textbf{12.7} FID at 72--76% adversarial robustness, while reducing subgroup disparity Δbias by \textbf{41%}, demonstrating that mechanistic interpretability, fairness, and generative performance can be jointly optimized.
Automated systems built on artificial intelligence (AI) are increasingly deployed across high-stakes domains, raising critical concerns about fairness and the perpetuation of demographic disparities that exist in the world. In this context, causal inference provides a principled framework for reasoning about fairness, as it links observed disparities to underlying mechanisms and aligns naturally with human intuition and legal notions of discrimination. Prior work on causal fairness primarily focuses on the standard machine learning setting, where a decision-maker constructs a single predictive mechanism fY for an outcome variable Y, while inheriting the causal mechanisms of all other covariates from the real world. The generative AI setting, however, is markedly more complex: generative models can sample from arbitrary conditionals over any set of variables, implicitly constructing their own beliefs about all causal mechanisms rather than learning a single predictive function. This fundamental difference requires new developments in causal fairness methodology. We formalize the problem of causal fairness in generative AI and unify it with the standard ML setting under a common theoretical framework. We then derive new causal decomposition results that enable granular quantification of fairness impacts along both (a) different causal pathways and (b) the replacement of real-world mechanisms by the generative model's mechanisms. We establish identification conditions and introduce efficient estimators for causal quantities of interest, and demonstrate the value of our methodology by analyzing race and gender bias in large language models across different datasets.
A central goal of mechanistic interpretability is to identify which internal components causally drive a language model's behavior. Because these importance estimates serve as the evidence for identifying circuits, systematic errors can lead to the misidentification of the underlying mechanisms. While activation patching provides a gold-standard causal metric, its computational cost is prohibitive at scale. Practitioners instead rely on attribution patching, a gradient-based, first-order approximation whose reliability remains poorly understood. In this work, we characterize the source of this unreliability, demonstrating that the dominant error stems from the non-linearities in the downstream network rather than local curvature at the patched component. This insight yields three practical tools: (i) a reliability score to detect untrustworthy estimates, (ii) error bounds quantifying potential attribution mis-specifications, and (iii) a Hessian-vector-product (HVP) correction that eliminates the leading-order error with only one additional backward pass. In evaluations across five model families (124M-9B parameters) and both random-token and naturalistic (name-swap) perturbations, HVP is the only second-order correction feasible at larger scale, where standard baselines like Integrated Gradients become computationally prohibitive. In comparative experiments, a multi-step HVP variant matches or exceeds the accuracy of Integrated Gradients at significantly lower compute, outperforming prior second-order baselines. These improvements lead to higher-fidelity circuit recovery on standard benchmarks and support a Screen-Flag-Fix workflow that targets computational effort only toward the components flagged as unreliable.
As AI generative models evolve at unprecedented speed, image attribution has become a moving target. New diffusion, adversarial and autoregressive generators appear almost monthly, making existing watermark, classifier and inversion methods obsolete upon release. The core problem lies not in model recognition, but in the inability to adapt attribution itself. We introduce IncreFA, a framework that redefines attribution as a structured incremental learning problem, allowing the system to learn continuously as new generative models emerge. IncreFA departs from conventional incremental learning by exploiting the hierarchical relationships among generative architectures and coupling them with continual adaptation. It integrates two mutually reinforcing mechanisms: (1) Hierarchical Constraints, which encode architectural hierarchies through learnable orthogonal priors to disentangle family-level invariants from model-specific idiosyncrasies; and (2) a Latent Memory Bank, which replays compact latent exemplars and mixes them to generate pseudo-unseen samples, stabilising representation drift and enhancing open-set awareness. On the newly constructed Incremental Attribution Benchmark (IABench) covering 28 generative models released between 2022 and 2025, IncreFA achieves state-of-the-art attribution accuracy and 98.93% unseen detection under a temporally ordered open-set protocol. Code will be available at https://github.com/Ant0ny44/IncreFA.