Video summary

L'intelligenza artificiale sta diventando un problema per la ricerca

Main summary

Key takeaways

News and Commentary

Summary

The video argues that language-model AI (e.g., GPT) is increasingly harming the culture and methodology of academic research. The harm is described as coming less from “breaking rules” and more from creating a flood of low-effort, poorly vetted papers that damages how scientific value is recognized.

Key points and reasoning

  • Rising volume of AI-generated submissions

    • The presenter tracks a yearly count of emails from people asking him to review “amateur physics articles.”
    • The numbers reportedly increase sharply:
      • 2–6 per year (2018–2022)
      • 14 (2023)
      • 32 (2024)
      • 70 (2025)
      • 187 already by June 2026 (mid-year)
    • Most of these articles are said to be written using language models, especially GPT.

  • “Noise” problem in research publishing

    • The surge clogs inboxes and review channels.
    • With many low-quality or semi-sensical AI-generated manuscripts, genuinely original work becomes harder to notice—especially for independent researchers without strong academic credentials.

  • No structural containment

    • The presenter argues there aren’t enough reviewers, or effective systems, to filter the growing volume.
    • As a result, selection processes and incentives will likely change in unpredictable ways.

  • Methodological and cultural critique: papers as an end product

    • A major conceptual complaint is that some people treat writing and publishing a paper as the goal.
    • AI is framed as a shortcut that enables this inversion.
    • By contrast, the presenter emphasizes proper research as a process where the paper is the final output of months (or longer) of study, analysis, and understanding.

  • Effort and understanding can’t be skipped

    • Using an example from a master’s thesis, the presenter explains that copying code (or—by analogy—outsourcing understanding to AI) fails unless you can explain the code line-by-line and understand the underlying methods.
    • Real work is described as involving learning (e.g., Monte Carlo methods), testing, running simulations, and doing in-depth reasoning before writing.

  • Misleading “revolutionary physics” claims

    • The presenter describes receiving papers claiming to overturn all physics—such as revisions to the “entire nature” of space/time.
    • These claims are described as outside scientific logic, suggesting AI can produce polished narratives without real comprehension.

  • Delegating writing isn’t the same as delegating analysis

    • The presenter distinguishes:
      • Serious AI use: researchers delegate narrowly defined tasks they understand (e.g., searching data).
      • Shortcut use: prompting AI to generate a complete publishable paper.
    • The latter is framed as the opposite of scientific rigor.

  • Practical consequence: papers people can’t even defend

    • If authors don’t understand what they wrote, they can’t properly discuss it or answer questions.
    • Publication becomes more performative than scientific.

Tone / position

The presenter says he isn’t just complaining that AI is “illegal” or trendy. He wants a rational, structural critique of how AI changes incentives and the meaning of research.

Presenters or contributors

  • The video presenter (unnamed in the subtitles)

Original video