Mykhaylo Bogachov

This paper extends the notion of performative prediction, or the capacity of models to influence the phenomena they predict, to large language models (LLMs). I argue that LLMs necessarily influence what and how humans write and speak, and that their performance depends on this influence. While certain LLM outputs can be verified formally or against ground truth, most are evaluated against common human practice. For such cases, I propose an interpretation of LLMs as predictive role-players, where LLM outputs are approximations of human linguistic behaviour appropriate to social roles or recognizable tasks. Exposure to such outputs affects action-guiding beliefs in humans, so LLM-generated “predictions” of typical language use become interventions into how members of society communicate or act. The epistemic habits underlying machine learning make such influence both likely and difficult to recognize, yet emerging computer-science frameworks can be extended to study performativity in LLMs specifically. Evidence that LLMs already change the style and content of human writing and speech, together with the difficulty of screening AI influence out of the human-authored datasets for next-generation models, indicates that LLMs and human behaviour have already formed a feedback loop. Recognizing performativity in LLMs calls for a reassessment of a core idea underlying AI evaluation. For tasks with no objective success criteria, “human-level” intelligence is not a fixed target, but something LLMs can reshape, since the data they learn from is the very data they influence. Thus, some LLM capability gains can be ambiguous: either LLMs are imitating humans better, or humans are becoming like LLMs, or both. I discuss the technical and political stakes of performative LLMs and call for the development of better methods to quantify performative prediction.

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.