AI Policy

SocArXiv AI Policy

March 8, 2026

Context

This policy aims to protect the epistemic commons by distinguishing research that is part of advancing social science knowledge from that which merely dilutes our work – with much less effort than peer review. It continues a shift away from our earlier “accept almost everything” approach and is designed to support three main stakeholder groups: (a) scholars who rely on a more reliable research ecosystem, (b) moderators whose volunteer labor cannot bear excessive discretionary burdens, and (c) independent researchers who may benefit from clearer signals when their submissions fail to meet scholarly standards. With the emergence of other outlets devoted to AI-generated research (for example, aiXiv accepts work produced by large language models, or LLMs), including preprints, it is increasingly important that authors know where different kinds of work should be submitted so that our shared infrastructure remains effective and sustainable.

Because the technological landscape continues to shift quickly, we expect to revisit and update this policy as needed.

We are researchers, and part of a human research community. We make human assessments based on our perception, judgment, and experience – which we do not defer to automated systems. Because of our human nature, we do not promise an algorithmically pure implementation of an AI policy.

In general, we want to implement policies that don’t burden good actors in order to filter out bad actors, that don’t devolve into an arms race we can’t win, and that aren’t too cumbersome for our moderators – all while improving the moderation process and making it as efficient as reasonably possible.

The fact is that we are getting overrun with machine-generated slop, some of which is manipulative or fraudulent and some of which is very low quality. This is a common phenomenon across many platforms. At the same time, some real researchers are using LLMs and other automation tools to help do research. 

Also, LLMs are increasingly integrated into everything, so the boundaries are unclear. For example, people may use LLMs to do language translation, find and format references, or copy edit, and in the process automation may alter the text, with or without human oversight.

This policy process is complicated because any rules we write will likely be ingested into LLM training datasets, which may then produce submissions that comply with the letter of our rules.

Goals

With the above motivation and context, our goals with the policy are the following:

• Help disseminate valuable research better and faster
• Delineate types of work that are unacceptable
• Increase barriers to the circulation of fraudulent research
• Reduce moderation effort to keep up with submission flow and avoid burnout of our moderators
• Implement moderation policies that we can enforce consistently
• Reduce the dissemination of LLM-generated content
• Encourage author disclosure of acceptable LLM use
• Signal that we are discovering LLM usage to discourage at least some bad actors

Policy

• There are acceptable uses for LLMs and other machine-assistance tools in research on our site, as long as they are disclosed and documented. Failure to disclose, or implausible disclosures, are grounds for rejection. These are acceptable:
  • Language translation (we may ask for the original to review)
  • Pre-writing work such as literature searches, idea generation and organization, when the paper also reports on original research and the author attests to thorough human supervision (e.g., verifying sources).
  • Copy-editing and formatting
  • Machine-assisted content or data analysis
  • Dictation software

• These are unacceptable uses:
  • Generating text which is used verbatim (including whole paragraphs and sections)
  • Generating fake human subjects data (including simulated, in silico samples)
  • LLM as co-authors or interlocutors as if human (e.g., “interviewing” or “dialogue” with LLMs)
  • Generating false information to mislead moderators (or readers)
  • Submitting AI-generated content which the author has not thoroughly reviewed and confirmed (this includes reviewing the cited sources to confirm they exist and are accurately characterized, and verifying images)
  • Entire papers produced by AI generation with no human-generated components beyond prompts

• Other work we do not allow, whether created by machines or not:
  • Technical assessment or design of AI programs or algorithms performed by LLMs or similar programs
  • Superficial, high-level theory synthesis and systemic reviews (e.g., “proposed framework” papers with big claims and superficial write-ups)
  • Overviews of prior research or theory with no contribution to scholarship (comparable to undergraduate course assignments) 
  • Exceptions to the rules in this category may be made for work that has been peer-reviewed, or based on individual appeal.

• We evaluate papers using various procedures:
  • Moderator assessment of minimal contribution to social science
  • Identification of author as a human with a consistent identity (as marked by ORCiD or other sources)
  • Moderator detection of likely-AI slop, with the following red flags:
    • Literature reviews or synthesis without substantive contribution
    • Unsubstantiated grand theories, which may have:
    • Fanciful equations
    • Integration of multiple major theories into definitive new “paradigms,” “proposed frameworks,” or similar, without substantive engagement
    • Papers written in many short sections, often including bullet-point lists and numbered paragraphs.
    • Fake empirical research
    • Assessment, whether by machine or by moderator perception, of papers with:
      • High truth-indifference or phoniness (e.g., no realistic grounding)
      • Enterprise misrepresentation (e.g., false or misleading description of the research project)
      • Constraint evasion (e.g., implausible or facile assumptions without evidence)
      • Low falsifiability (e.g., untestable claims without justification)
      • Absence of methodological accountability (e.g., extensive methods and data discussion without any research output)

• We allow moderators to place papers on hold if they require further examination and move on to those that look more human-generated.

• We allow appeals by email to socarxiv@gmail.com. However, appeals end at our discretion. Posting on SocArXiv is a privilege, not a right.

• We ban authors for fraud, plagiarism (including LLM-generated plagiarism), abuse, misrepresentation, or repeated violations.