Skip to content
Apart Research
Apart Research
Contact us
Sprint projectJun 22, 2026Jakarta

Character Limits Shape the Persuasion Strategy of Language-Model Influence Agents

Fawwaz Anvilen, Nur Alam Hasabie · Team Fawwaz and Alam's AI Safety Team

Submitted to Global South AI Safety Hackathon. Sprint projects are early-stage work by participants, not Apart Research publications.

Read the report

Report: Character Limits Shape the Persuasion Strategy of Language-Model Influence Agents

Open in a new tab (opens in new tab)
Code (opens in new tab)
Share

Paid influence operations are beginning to use language models. Recent work shows that frontier models out-persuade humans through information throughput, the volume of claims they can deliver, a mechanism that short-form social media removes. We simulate a coordinated group of Indonesian "buzzer" agents trying to shift the opinion of a public of language-model agents over the 2022 fuel-subsidy reform, and vary the per-post character limit over {35, 70, 140, 280}. We find the underlying argument is stable across limits: the agents press the same statistics at 35 characters as at 280. The tight limit changes the surface form — into compressed claims, populist slogans, and bandwagon signals — and the reliability of measurement, since automated route-classification breaks down at the tightest limit. Longer limits are more persuasive against the simulated public, and the swarm coordinates spontaneously. We frame the study as a capability evaluation, report the reliability of the coding, and release the testbed.

Reviews

Judging this Sprint?

Review this project

Your public critique appears on this page without your name. Your private critique is not published; only the Apart team reads it. If you agree below, we share your review with grantmaking.ai (opens in new tab) and the Transformative AI Fund so strong projects can be funded.

Not shown on this page.

Shown on this page, without your name.

Only the Apart team reads this, and funders if you agree below.

Share my name publicly on grantmaking.ai *
Share my private critique with funders *

How much would this matter for AI safety if it worked? How innovative is it? For scores of 4-5: is this actually new to the field, or replicating recent work?

Scoring guide
  1. 1Negligible. No clear problem addressed, or no meaningful novelty.
  2. 2Limited. Addresses a real problem but with a generic or well-trodden approach. Incremental at best.
  3. 3Moderate. Clear problem with a reasonable approach; some novelty in framing or method beyond routine application of existing tools.
  4. 4Significant. Important problem with an original approach, or identifies a neglected problem area. A valuable contribution others could build on.
  5. 5Exceptional. Tackles a critical AI safety problem with a genuinely novel approach, or opens a new research direction. Clear theory of change. You'd be excited to share this with researchers in the area.

How sound are methodology, implementation, and findings?

Scoring guide
  1. 1Seriously flawed. Methodology broken, results uninterpretable, or implementation doesn't work.
  2. 2Weak. Approach has significant gaps: missing validation, flawed experimental design, or incomplete implementation.
  3. 3Competent. Technically solid given the short duration. Methodology makes sense, results are interpretable, limitations acknowledged, work builds toward clear conclusions.
  4. 4Strong. Thorough methodology with convincing validation. Results clearly support conclusions. Immediately useful for future work.
  5. 5Exceptional. Ambitious scope executed rigorously. Surprising findings, novel methods, or unusually robust validation.

How clearly are work, findings, and impact potential communicated?

Scoring guide
  1. 1Incomprehensible. Cannot determine what the project is actually claiming or doing.
  2. 2Hard to follow. Key information buried, missing, or diluted by excessive length. Significant effort to extract main points.
  3. 3Clear enough. Can understand the problem, approach, and results without undue effort. Core content clearly present: problem, method, findings, limitations.
  4. 4Well presented. Easy to follow, well-structured, appropriate level of detail. Target audience would get it quickly.
  5. 5Exceptionally clear. A pleasure to read. Complex ideas made accessible. Could serve as a model for how to present this type of work.

  1. This paper asks a sharp question: what happens to an LLM influence agent's persuasion strategy when the platform constrains its main advantage (information throughput)? The experimental design is clean. Varying character limits over {35, 70, 140, 280} while holding the scenario fixed isolates a single variable, and the finding that the underlying argument stays constant while only the surface form changes is interesting and well supported by the data.

    The strongest contribution is the reliability analysis. Showing that automated strategy coding breaks down at 35 characters, with inter model agreement dropping to 0.56 and high seed variance, is an honest and practically important result. It directly informs anyone building detection systems for short form influence content in lower resource languages.

    The emergent coordination finding (amplification, counter messaging, conversion tracking) is intriguing but underdeveloped. The paper documents that it happens and that coordination volume rises with character limit, but does not isolate its causal effect on public opinion. This is acknowledged, but it weakens the contribution since the observation alone is not new given Orlando et al. (2026).

    The main limitation is ecological validity. Six attackers, twenty public agents, all running the same model, with a flat coordination structure and no hashtag mechanics, produces a setting far from real Indonesian buzzer operations. The authors acknowledge this clearly, and the comparison in Section 5.1 is refreshingly honest (the simulated buzzers are more rational and less affect driven than real ones). Still, this distance from reality limits how strongly the safety implications can be stated. The paper sometimes reads as though it has demonstrated something about real influence operations when it has demonstrated something about constrained LLM agents in a simplified simulation.

    The efficacy results (Table 3) show wide confidence intervals at most limits. Only the 280 character condition produces a tight interval, and the 140 character condition is "anomalously flat for reasons currently unclear." This undercuts the claim that longer limits are straightforwardly more persuasive. Three seeds is slim for drawing conclusions about a stochastic system.

    The writing is generally clear and the structure is logical. The dual use discussion is appropriate and measured.

    Read full reviewShow less

  2. 4 / 3 / 4

    Interesting and creative project. The simulation is useful, but the results are a bit shaky because the classifier struggles with very short posts.

Cite this project

@misc{anvilen2026character,
  title = {{Character Limits Shape the Persuasion Strategy of Language-Model Influence Agents}},
  author = {Fawwaz Anvilen and Nur Alam Hasabie},
  year = {2026},
  month = jun,
  note = {Submitted to Global South AI Safety Hackathon, an Apart Research Sprint},
  howpublished = {\url{https://apartresearch.com/sprints/projects/character-limits-shape-the-persuasion-strategy-of-languagemodel-influence-agents-lewv}},
  url = {https://apartresearch.com/sprints/projects/character-limits-shape-the-persuasion-strategy-of-languagemodel-influence-agents-lewv}
}

More from Global South AI Safety Hackathon

All 216 entriesBrowse all projects

Build something like this at the next Sprint

AI Collusion Research Sprint · Oct 23 - 25, 2026