TLDR: The paper proposes a techno-philosophical framework, drawing on Gilbert Simondon’s concept of individuation and ‘futurity,’ to understand AI systems as continuously evolving rather than static. It argues that current regulations like the EU AI Act overlook this dynamic nature, particularly how data generates recursive value chains. Using the Google AI stack as a case study, the authors illustrate how user interactions are continuously leveraged to refine AI, leading to power asymmetries. They propose new governance mechanisms like temporal traceability and a right to contest recursive reuse to address these regulatory blind spots and ensure more effective, lifecycle-based AI oversight.
Artificial intelligence systems are not static tools; they are constantly evolving entities, learning and adapting from every interaction. This dynamic nature, often overlooked by current regulations, is the core focus of a new research paper titled Futurity as Infrastructure: A Techno-Philosophical Interpretation of the AI Lifecycle by Mark Coté and Susana Aires.
The paper introduces a fresh perspective on the long-term dynamics of data within AI systems. It argues that the journey from data collection to model deployment creates self-reinforcing cycles of value, which existing regulatory frameworks for Responsible AI struggle to address. To understand this, the authors propose a ‘techno-philosophical’ approach, drawing inspiration from Gilbert Simondon’s philosophy of technology.
Understanding AI’s Evolution: Individuation and Technicity
At the heart of this new framework are two key concepts: individuation and technicity. Individuation describes AI as a process of ‘becoming’ rather than a fixed object. It outlines three phases:
- The pre-individual milieu: This is where raw data, model architectures, and parameters exist as untapped potential.
- The process of individuation: During this phase, these potentials cohere through training, tuning, and integration, leading to a functional AI model.
- The individuated AI: The functioning model, which crucially, retains a ‘residual pre-individuality’ – an ongoing capacity for adaptation, retraining, and transfer to new domains.
Technicity, the second concept, refers to the inherent potential for new functionality present in all technical objects. For AI, this means that data and models are not exhausted after initial use; they carry a potential that can be continuously transformed and recombined to create novel functionalities and applications.
Futurity: The Self-Reinforcing AI Lifecycle
To translate these philosophical ideas into practical terms, the paper introduces ‘futurity.’ This concept describes the self-reinforcing lifecycle of AI, where increased data availability improves model performance, deepens personalization, and opens up new areas of application. Futurity highlights how data in deep learning systems is ‘non-rivalrous’ (can be reused infinitely without depletion) and how infrastructures like feature stores enable real-time feedback and continuous adaptation. This dynamic process, the authors argue, leads to escalating power imbalances, concentrating value and decision-making power within a few tech giants.
The Google AI Stack: A Case Study in Futurity
To illustrate futurity in action, the paper examines the Google AI stack, tracing the recursive lifecycle of data across seven interconnected stages:
- Data Generation and Capture (App to Firebase): User interactions are captured and formalized into machine-readable data.
- Data Structuring and Preprocessing (Firebase to BigQuery): Raw data is cleaned, standardized, and prepared for both historical training and real-time inference.
- Model Training and Orchestration (BigQuery to TFX): Data is processed to form a coherent, operational model capable of generalizing patterns.
- Deployment and Personalisation Infrastructure (TFX to Vertex AI): The trained model is deployed, becoming an adaptive system that personalizes responses based on historical data and emergent segmentation.
- Personalised Inference (Real-Time Prediction): The model uses current user context to generate tailored predictions, influencing user behavior before they even act.
- Adaptive Feedback Loops (Recursive Learning): User responses to predictions are captured and fed back into the system, refining future inferences in near real-time.
- Reintegration (Recursive Infrastructure and Value Accumulation): Feedback data re-enters the pipeline for both immediate inference and future model retraining, closing the loop and continuously compounding value for the platform.
This case study demonstrates how the Google AI stack functions as a ‘closed-loop’ system where user data is continuously transformed into predictive and financial capital, operationalizing and monetizing futurity.
Addressing Regulatory Blind Spots in the EU AI Act
The paper identifies three major blind spots in the EU AI Act from this techno-philosophical perspective:
- Temporal Recursivity: The Act treats AI systems as static, bounded objects, failing to account for their continuous evolution through feedback loops and post-deployment refinement.
- Ex Ante Regulatory Logic: The Act’s focus on pre-market assessment struggles with recursive AI systems that adapt and evolve after deployment, necessitating a shift to lifecycle-based regulation.
- Asymmetries of Value Capture: The Act does not address how AI systems extract value from time, enclosing user agency and distributing benefits unevenly, particularly for large platforms.
Also Read:
- Understanding and Governing AI Agent Behavior in the Digital Era
- Understanding AI Governance in Practice: Insights from the Fabric Repository
Proposals for Futurity Governance
To address these challenges, the authors propose several governance mechanisms:
- Temporal Traceability: Tracking how predictions and model outputs change over time and what data transformations shaped them.
- Feedback Accountability: Making visible how user interactions feed back into model updates and personalization strategies.
- Recursion Transparency: Identifying which inputs contribute to which model outputs at various stages of training or inference.
- Right to Contest Recursive Reuse: Allowing individuals to challenge or opt out of the continuous use of their behavioral data for refinement and personalization.
Beyond these, the paper suggests broader regulatory proposals, including ‘temporal value disclosure’ (reporting how much model performance is due to post-deployment user data), ‘infrastructure transparency requirements’ (mandating clear reporting on data flows within proprietary systems), and an ‘AI windfall tax’ to fund a public ‘Futurity Value Redistribution Fund’ or a federated AI Data Commons. These proposals aim to reorient the flow of AI futurity towards public value and ensure more equitable distribution of benefits from AI’s continuous evolution.
