A curated index of 218 research papers on how humans must evolve to use AI well as AI advances. Organized by the four-phase framework — Humans Use AI as Tool · Assistant · Executor · Organization. Each paper is grounded evidence for one phase: what capability humans must sustain at that phase, how it weakens without active effort, and how AI systems can be designed to support that human evolution.
As humans delegate more to AI, the capabilities humans need shift: from critical thinking, to evaluative expertise, to metacognitive monitoring, to systems thinking. When human evolution lags AI advancement, the failure modes pile up: weakened reasoning, polished-but-flawed artifacts approved, autonomous workflows drifting unmonitored, coordinated agent systems running opaque. This index curates the research that documents that human side.
The structured store papers.yaml is the source of truth — the README, statistics, and the website are auto-generated from it. See CLAUDE.md for the schema and contribution workflow.
Phase 1 (61) · Emerging Phase 2 (15) · Phase 2 (52) · Emerging Phase 3 (8) · Phase 3 (25) · Emerging Phase 4 (30) · Phase 4 (0) · Surveys & Position Papers (27)
CC Collaboration & Co-Creation (61) · MA Mutual Adaptation (63) · HF Human Feedback Loops (37) · LH Longitudinal HCI Studies (70) · PS Position & Survey (84)
automation bias (12) · ChatGPT (11) · RCT (9) · trust (9) · governance (9)productivity (8) · RLHF (8) · Copilot (8) · trust calibration (7) · longitudinal (7)agentic AI (6) · alignment (6) · sycophancy (6) · reliance (6) · over-reliance (6)benchmark (3) · dataset (3) · survey (3)
Mor Naaman (6) · Zahra Zahedi (5) · Subbarao Kambhampati (5) · Diyi Yang (4) · Daniel Buschek (4)
Sarath Sreedharan (4) · Dario Amodei (4) · Mina Lee (3) · Kevin Wei (3) · Noam Kolt (3)
Lev Tankelevitch (3) · Advait Sarkar (3) · Sandhini Agarwal (3) · Eric Horvitz (3) · Ethan Perez (3)
Samuel R. Bowman (3) · Amanda Askell (3) · Jeffrey T. Hancock (3) · Jan Leike (3) · Raja Parasuraman (3)
We welcome contributions from the community!
- Missing a paper? Open an issue with the paper title, link, and any relevant details — we'll add it.
- Want to add papers yourself? Edit
papers.yaml, runbash scripts/update_repo.sh, then submit the regenerated diff. SeeCLAUDE.mdfor the YAML schema. - Spotted an error? Open an issue or PR to correct any paper metadata (authors, dates, institutions, etc.).
Full searchable index: https://xli04.github.io/Awesome-Human-AI-Coevolution-Paper-List/. Structured source of truth:
papers.yaml(218 entries). Framework definitions:deep_research/phased_framework.md. Each paper is assigned a single phase, withemerging-phase-Xfor clear bridge cases; the secondary 5-category axis (CC/MA/HF/LH/PS) is also stored per entry.
Humans use AI to answer questions. To use AI well here, humans must sustain critical thinking — comparing AI outputs against their own reasoning rather than absorbing them passively. The capability erodes through uncritical acceptance, and the feedback that erosion produces pushes models toward sycophancy.
Papers that bridge reasoning-level use with artifact production. Humans use AI to prompt their thinking but begin to produce drafts or ideation material that needs evaluation, not just judgement.
Humans use AI to produce bounded artifacts (drafts, code snippets, partial implementations) and verify them. To use AI well here, humans must sustain evaluative expertise — knowing what good work satisfies, including failure modes. The capability erodes when polished output is accepted on surface signals.
Papers that bridge artifact-level assistance with end-to-end autonomy. Humans still drive the workflow but begin to delegate sequences of steps, demanding monitoring on top of evaluation.
Humans use AI to complete end-to-end workflows, setting goals and intervening when execution drifts. To use AI well here, humans must practice metacognitive monitoring — selective inspection of where the workflow can fail. The capability erodes through passive supervision, producing scaled errors humans cannot catch in time.
Papers that bridge autonomous-agent use with system-level coordination. Includes governance-layer interventions on ecosystem feedback loops, constitutional / RLAIF systems, and the model-collapse line of work — contributions that argue toward Phase 4 governance without demonstrating a domain having fully arrived there.
Humans use AI to coordinate systems of work across many agents. To use AI well here, humans must develop systems thinking — shaping the system that produces actions rather than inspecting each action. No domain has officially entered Phase 4 yet, so this section is intentionally empty, and the Emerging Phase 4 section above lists the papers that argue toward this mode.
Surveys, position pieces, and theoretical frameworks that span multiple phases — scaffolding for how to think about humans using AI well, rather than grounded evidence for any one phase.
Some of the design and scaffolding here is adapted from OSU-NLP-Group/GUI-Agents-Paper-List. Thanks for their awesome work!