Human-AI Coevolution

28 concepts15 papers1 proposal

Traditional machine learning operates under a clean, unidirectional assumption: humans provide data, engineers train a model, and the model serves as a static tool to automate or assist with a task. However, as artificial intelligence becomes deeply embedded in digital platforms, physical spaces, and social institutions, this linear framing breaks down.

As argued by Jacy Reese Anthis et al. in the ICLR 2025 Workshop on Human-AI Coevolution, AI systems are not static tools but evolving agents engaged in continuous, long-term feedback loops with human users. When an AI system is deployed, its outputs alter human behavior, beliefs, and preferences. Humans adapt to the AI's capabilities and constraints, and this modified human behavior generates the very data used to train the next iteration of the AI. This reciprocal process is known as human-AI coevolution.

The Mechanics of the Closed Loop

To understand why human-AI systems coevolve, we must look at how algorithmic outputs become behavioral inputs, which then loop back into the model's training pipeline.

┌────────────────────────────────────────────────────────┐
│                                                        │
▼                                                        │
Algorithmic Output (Recommendations, Content, Decisions)  │
│                                                        │
▼                                                        │
Altered Human Behavior (Shifted preferences, language)   │
│                                                        │
▼                                                        │
New Data Generation (Logs, clicks, text, labels)         │
│                                                        │
└────────────────────────────────────────────────────────┘

The authors of How Algorithmic Confounding in Recommendation Systems Increases Homogeneity and Decreases Utility formalize this cycle through the lens of algorithmic confounding. When a recommendation engine surfaces a specific product or piece of content, it artificially inflates the probability that a user will interact with it. If the system subsequently treats that interaction as independent evidence of user preference during retraining, it creates a self-reinforcing cycle. The system effectively learns from data that its own past actions confounded.

This phenomenon is what Iyad Rahwan et al. term a hybrid feedback loop in Machine behaviour. Because algorithmic outputs continuously shape human choices, the data generated by society is no longer an objective reflection of human nature; it is a product of the human-AI interaction itself. Consequently, machine behavior and human social dynamics can no longer be analyzed in isolation—they form a unified, co-evolving system.

How Humans and Environments Adapt

The adaptation within this loop is truly bidirectional. Humans do not merely consume AI outputs; they actively alter their behavior to navigate, exploit, or accommodate the machine.

Behavioral Accommodation

In Tidying Up the Conversational Recommender Systems' Biases, Armin Moradi highlights that users frequently modify their natural communication patterns—such as simplifying their language or conforming to perceived technical limitations—when they know they are interacting with an artificial agent. This behavioral shift ensures that the AI is continuously trained on unnatural, distorted data, degrading the authenticity of future interactions.

Policy Friction

When an AI's learning policy conflicts with human psychology, the behavioral shifts can be destabilizing. Shashi Suman et al. explore this in Potential Impacts of Smart Homes on Human Behavior: A Reinforcement Learning Approach, demonstrating that when a reinforcement learning system and a human inhabitant have misaligned reward structures, it can inadvertently trigger erratic human behaviors, such as frequent task-switching and fragmented routines.

Environmental Manipulation

It is not just individual humans who adapt; entire digital ecosystems evolve to cater to algorithms. Giulio Frey et al. introduce the concept of mecha-nudging in Mecha-nudges for Machines, showing how digital platforms subtly alter their environments to optimize for the heuristics of autonomous AI agents. This deployment feedback loop risks a gradual disempowerment of humans as the digital world begins to prioritize machine-readable signals over human-centric utility.

Systemic Consequences of the Loop

When these recursive loops run unchecked, they cause the long-term performance and societal impact of an AI system to deviate drastically from its behavior on static, human-only training sets.

  • Homogenization: Allison J. B. Chaney et al. demonstrate that feedback loops drive users to behave more similarly to one another over time, narrowing the diversity of content consumed and reducing overall user utility.
  • Epistemic Rewiring: In AI Systematically Rewires the Flow of Ideas, Zhonghao He et al. characterize AI as an "epistemic technology" that actively rewires human knowledge and beliefs. Because human preferences shape AI outputs and AI outputs modify human preferences, these self-reinforcing cycles can drastically amplify underlying societal biases.
  • Value Lock-in: Aligning AI to a static snapshot of current human values can freeze moral progress. As Tianyi Qiu et al. point out in ProgressGym: Alignment with a Millennium of Moral Progress, because AI outputs actively influence human values, static alignment risks locking in contemporary moral blind spots and preventing the natural evolution of human ethics.

From Static Metrics to Coupled Dynamical Systems

To safely build and evaluate AI in a coevolutionary world, the research community is moving away from static benchmarks toward frameworks that treat human-AI interaction as a coupled dynamical system.

Sarah Dean et al. argue in Accounting for AI and Users Shaping One Another: The Role of Mathematical Models that mathematical models must explicitly account for how algorithmic outputs modify user states over time. This perspective is shared by Dino Pedreschi et al. in Human-AI Coevolution (Abstract Reprint), who advocate shifting the focus of AI evaluation from static performance metrics to the long-term evolution of the system's data-generation process.

This paradigm shift manifests across several design and evaluation methodologies:

Dynamic Personalization Benchmarks

Traditional personalization benchmarks evaluate how well a model adapts to a fixed user profile. To address this limitation, Md Awsafur Rahman et al. developed LikeBench: Evaluating Subjective Likability in LLMs for Personalization, which measures how models maintain alignment and "charm" over long, noisy interactions where both the user and the machine are adjusting to one another.

Relationship-Centric Design

Moving beyond simple user interfaces, Josh Andres et al. propose a scenario-based design framework in Understanding and Shaping Human-Technology Assemblages in the Age of Generative AI that treats the long-term relationship dynamics between users and agentic technologies as a core engineering ingredient.

Scaffolding and Co-Learning

Coevolution can also be steered toward positive outcomes. In Supporting Human and Machine Co-Learning in Citizen Science: Lessons From Gravity Spy, Carsten Østerlund et al. demonstrate a "co-learning" framework where the AI scaffolds human learning and skill acquisition, while human volunteers simultaneously help the AI adapt to novel, unclassified data patterns.

Bidirectional Alignment

Ultimately, managing these feedback loops requires a fundamental rethinking of AI safety. As outlined by Hua Shen et al. in the ICLR 2025 Workshop on Bidirectional Human-AI Alignment, alignment cannot be a one-way street. To prevent systemic misalignment, we must develop adaptive alignment strategies that regulate the mutual influence between humans and machines. This requires experimental frameworks, such as those proposed by Avishek Anand et al. in Effects of Algorithmic Decision-Making and Interpretability on Human Behavior: Experiments using Crowdsourcing, to empirically measure and control how factors like model interpretability alter human trust, fairness, and decision-making within the recursive loop.

Key Foundational Work

Effects of Algorithmic Decision-Making and Interpretability on Human Behavior: Experiments using Crowdsourcing (2018) — contributed 2 concept elements to this direction.
How algorithmic confounding in recommendation systems increases homogeneity and decreases utility (2017) — contributed 1 concept elements to this direction.
History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_human_ai_co_evolution_7d5aec47,
  title = {human-AI co-evolution},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/human-ai-co-evolution/rdr_7d5aec47703f48d69ec6b58ed9800357},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 1, frozen text):

@misc{lacuna_human_ai_co_evolution_7d5aec47_v1,
  title = {human-AI co-evolution},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/human-ai-co-evolution/rdr_7d5aec47703f48d69ec6b58ed9800357/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-20},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake