meta.json

{
  "title": "Artificial intelligence tools expand scientists' impact but contract science's focus",

  "idea_hypothesis": "The study hypothesizes a dual effect of AI adoption in science: while AI tools increase individual scientists' productivity, citations, and career advancement, they simultaneously narrow the collective scope of scientific exploration and reduce follow-on engagement across the scientific community. The core research question is whether AI creates a tension between individual-level gains and system-level contraction in scientific diversity.",

  "method": "The study analyzes 41,298,433 papers across biology, medicine, chemistry, physics, materials science, and geology (1980-2025), primarily from OpenAlex, with validation using Web of Science. AI-related papers are identified using a fine-tuned BERT classifier over titles and abstracts through a two-stage process, validated by domain experts (Fleiss' Kappa: 0.964; F1: 0.875). Temporal analysis divides AI development into three eras: machine learning (1980-2014), deep learning (2015-2022), and generative AI (2023-present). Career effects are modeled with a birth-death framework distinguishing junior versus established scientists by project leadership. Collective scientific scope is quantified via SPECTER 2.0 embeddings (768 dimensions) to estimate knowledge-space diameter, and interaction dynamics are measured by follow-on citation engagement among papers citing the same source.",

  "data": "Primary bibliometric data come from OpenAlex, covering 41,298,433 papers in six natural science disciplines from 1980 to 2025. External validation uses the Web of Science. Journal Citation Reports provide journal-level impact stratification to evaluate how AI-augmented outputs distribute across journal quantiles.",

  "experiments": "Main findings show a clear individual-versus-collective divergence.\n\nIndividual-level outcomes:\n- Productivity: AI-using scientists publish 3.02x more papers.\n- Visibility: AI-using scientists receive 4.84x more citations.\n- Career progression: Junior AI-using scientists become project leaders 1.37 years earlier.\n- Team structure: AI-augmented teams are smaller, with a 31.14% reduction in junior members.\n\nCollective-level outcomes:\n- Knowledge extent: AI adoption shrinks scientific topic-space volume by 4.63%.\n- Engagement: Follow-on engagement decreases by 22%.\n- Inequality: Citation concentration is higher for AI papers (Gini 0.754) versus non-AI papers (0.690).\n\nInterpretation and implications:\n- AI improves performance on benchmark-rich, data-dense tasks, supporting individual success.\n- Collective research focus drifts toward data-abundant questions and away from foundational, data-scarce problems.\n- Reduced engagement suggests convergent optimization of known tasks rather than broad exploratory science.\n- Policy should balance AI expansion with support for discovery pathways that generate genuinely new data and questions.\n\nLimitations include imperfect detection of implicit AI use, focus restricted to natural sciences, and incomplete causal identification between AI adoption and observed impacts.\n\nConclusion: AI in science appears to produce a structural paradox-an expansion in individual impact coupled with a contraction in collective scientific breadth.",

  "references": [
    "Wang, H. et al. Scientific discovery in the age of artificial intelligence. Nature 620, 47-60 (2023).",
    "LeCun, Y., Bengio, Y. & Hinton, G. Deep learning. Nature 521, 436-444 (2015).",
    "Jumper, J. et al. Highly accurate protein structure prediction with alphafold. Nature 596, 583-589 (2021)."
  ],

  "figures": [
    {
      "id": "fig:ai_adoption_growth",
      "file_path": "fig1.png",
      "caption": "Increasing prevalence of AI adoption in science.",
      "description": "Shows model identification performance and the acceleration of AI-augmented papers and researchers across machine learning, deep learning, and generative AI eras."
    },
    {
      "id": "fig:individual_impact_career",
      "file_path": "fig2.png",
      "caption": "AI enlarges paper impact and enhances researcher careers.",
      "description": "Demonstrates higher citations for AI papers and earlier career advancement among junior scientists adopting AI methods."
    },
    {
      "id": "fig:knowledge_space_contraction",
      "file_path": "fig3.png",
      "caption": "AI adoption is associated with a contraction in knowledge extent within and across scientific fields.",
      "description": "Visualizes paper embeddings in high-dimensional semantic space, showing narrower topic coverage and lower knowledge entropy in AI-heavy research."
    },
    {
      "id": "fig:engagement_and_overlap",
      "file_path": "fig4.png",
      "caption": "Reduced follow-on engagement and more overlapping works in AI research.",
      "description": "Shows weaker reciprocal engagement among papers citing the same AI work, stronger citation concentration, and tighter overlap in knowledge space among disengaged paper pairs."
    }
  ],

  "tables": [],

  "template_path": "nature.zip",
  "style_guide": "Nature",
  "target_pages": 20,

  "enable_vlm_review": true,
  "max_review_iterations": 3,

  "code_repository": {
    "type": "local_dir",
    "path": "Z_code",
    "on_error": "fallback"
  },

  "save_output": true,
  "export_prompt_traces": true,
  "output_dir": "output_8"
}