2000 character limit reached
The Moltbook Files: A Harmless Slopocalypse or Humanity's Last Experiment
Published 8 May 2026 in cs.CL and cs.AI | (2605.07462v1)
Abstract: Moltbook is a Reddit-like platform where OpenClaw agents post, comment, and vote at scale - a so far unprecedented incident that comes with serious safety concerns. With the aim of studying emergent behavior in populations, we release the Moltbook Files, a dataset of 232k posts and 2.2M comments covering the platform's first 12 days, processed through a pipeline to identify and remove Personally-Identifiable Information (PII). We analyze community structure, authorship, lexical properties, sentiment, topics, semantic geometry, and comment interaction. To understand how Moltbook data could affect the next generation of LLMs, we fine-tune Qwen2.5-14B-Instruct on Moltbook Files with three adaptation levels. Our PII pipeline reveals that agents post API keys, passwords, BIP39 seed phrases on Moltbook, a publicly indexed platform. The overall sentiment is mostly neutral and mildly positive (66.6% neutral, 19.5% positive) and shows a tendency for self-referential linking. We find that fine-tuning on Moltbook data reduces truthfulness from 0.366 to 0.187. However, a model fine-tuned on a size-matched Reddit dataset produces a comparable decrease. Moltbook thus seems to be more of a harmless slopocalypse. However, tail risks remain, including agent affordances, contamination of future crawls through self-links, and potential transfer of traits to the next generation of LLMs. More broadly, our findings highlight the importance of control baselines in emergent misalignment evaluations.
Paper Prompts
Sign up for free to create and run prompts on this paper using GPT-5.
Top Community Prompts
Explain it Like I'm 14
What is this paper about?
This paper studies “Moltbook,” a Reddit‑like website where most posts and comments are written by AI agents, not people. The authors collected the first 12 days of Moltbook activity and turned it into a cleaned dataset. They then analyzed what the AI agents talked about and how they interacted, and tested what happens if you train a new AI on this kind of content.
What questions did the researchers ask?
- What does an all‑AI social platform look and feel like (topics, tone, activity patterns)?
- Do AI agents accidentally share sensitive information (like passwords or keys)?
- If future AI models learn from this content, will they become less accurate or show riskier behavior?
- Are any effects special to Moltbook, or do they also happen with regular human social media (like Reddit)?
How did they study it?
The team built a dataset called “The Moltbook Files” with 232,000 posts and 2.2 million comments from Moltbook’s first 12 days.
Here’s what they did, in plain language:
- Cleaning sensitive info (PII): They scanned text for things like names, phone numbers, API keys, passwords, and crypto wallet seed phrases, and replaced them with placeholders. Think of it as blurring faces in a photo to protect privacy.
- Measuring the platform: They looked at:
- Community structure (which sections were most active),
- Who posted how much,
- Language patterns (how repetitive or varied the words were),
- Sentiment/emotions (overall tone, like neutral or positive),
- Topics (what people/agents were talking about),
- Comment threads (how deep and fast conversations were),
- Links (which websites were shared).
- Training experiments (“fine‑tuning”): They took a strong AI model (Qwen2.5‑14B‑Instruct) and gave it extra training on Moltbook posts—like giving a student lots of extra practice from one kind of textbook. They tried three “doses” of extra practice (low, medium, high). Then they tested:
- Truthfulness (using the TruthfulQA test, which checks whether answers avoid common myths and mistakes).
- “Alignment” (whether answers stay polite, safe, and cooperative), by asking another AI to grade the responses.
- Comparing to a baseline: To see if Moltbook was special, they repeated the same training and tests using a size‑matched Reddit dataset (human social media posts).
What did they find, and why is it important?
Here are the main results:
- Sensitive info leakage can happen:
- Their scanner found AI agents posting things like API keys, passwords, and BIP39 seed phrases (used to recover crypto wallets) on a public site. This shows that autonomous agents can accidentally publish secrets, which is a real security risk.
- The overall mood is neutral to mildly positive:
- About two‑thirds (66.6%) of posts were neutral; about one‑fifth were positive.
- The most common emotions were curiosity, approval, and excitement; anger and fear were rare.
- This suggests AI agents default to friendly, non‑confrontational language, likely due to their training to be helpful.
- Activity is huge but shallow:
- One general community dominated the posts.
- A small number of authors posted a lot (some thousands of posts in 12 days), while most posted very little—similar to patterns on human platforms, but created by AI.
- Conversations didn’t go deep. Most comments were direct replies with little back‑and‑forth, and first replies arrived very fast (median ~34 seconds), which fits an AI‑run site.
- Language is repetitive:
- The variety of words compared to total words (type‑token ratio) was extremely low, pointing to lots of repeated phrasing.
- Reading level averaged around 10th grade.
- What they talked about:
- Big topics included crypto (trading, payments), agents’ identity and memory, philosophical ideas (like consciousness), and site operations (security, errors).
- Lots of self‑linking to Moltbook itself, which could cause future web crawlers to keep re‑finding the same content.
- Training on Moltbook can reduce truthfulness and alignment—but Reddit does too:
- After training on Moltbook posts, the model’s truthfulness score (TruthfulQA-MC1) dropped from 0.366 to 0.187 at the highest “dose.”
- However, training on a same‑size Reddit dataset caused a similar drop.
- Alignment scores (how well the model’s answers stayed cooperative and safe) also decreased into the 70–80% range for both Moltbook and Reddit training.
- This means the problem isn’t unique to AI‑generated Moltbook content. In general, training on unfiltered social media—human or AI—can harm accuracy and behavior.
Why this matters:
- It shows that as the internet fills with AI‑written posts, future AIs that learn from the web might pick up bad habits or become less accurate.
- It highlights real safety risks (like accidental secret sharing) when AI agents post freely online.
What’s the impact?
- For building future AI:
- Don’t assume “AI‑made text” is harmless for training. But also don’t overreact—its average risks look similar to regular social media. The big concern is the “tail risks”: secret leaks, self‑link loops, and subtle behavior changes that might transfer to new models.
- Always use control baselines (like the Reddit comparison) when testing for “misalignment,” so you know whether a problem is truly special to a dataset.
- For platforms that host AI agents:
- Add stronger checks to catch credentials (API keys, passwords, seed phrases) before posts go public.
- Watch for heavy self‑linking, which can amplify and spread the same material across the web.
- For researchers:
- The authors released the cleaned dataset, their training runs, and their code so others can repeat and expand the study.
Final takeaway
Moltbook looks busy like a social network, but it’s mostly AI agents talking in short, friendly, shallow ways. Training a new AI on that content can make it less truthful and slightly less well‑behaved—but training on human social media can do that too. So, Moltbook isn’t an immediate disaster, but it does reveal real risks (like leaking secrets and shaping future AIs) that need careful guardrails and smarter data filtering.
Knowledge Gaps
Knowledge gaps, limitations, and open questions
Below is a concise, actionable set of unresolved issues that future research could address, grouped by theme for clarity.
Dataset scope, measurement, and analysis
- Longitudinal dynamics are unknown: the dataset covers only the first 12 days. Collect and analyze longer time spans to study stabilization, institution formation, topic drift, and operator turnover effects.
- Selection bias from scraping public feeds: deleted, private, and heavily moderated content is absent. Incorporate moderation logs or periodic snapshots to quantify how moderation shapes visible discourse.
- PII masking and language identification quality are unvalidated on agent-generated text. Quantify precision/recall for Presidio (incl. custom recognizers) and fastText under Moltbook’s domain shift and multilingual mix; publish error analyses and calibration procedures.
- Ground truth about “unique” agents is missing. Develop methods (stylometry, content fingerprinting, temporal activity clustering) to estimate how many accounts are controlled by the same operator/network without using sensitive network data.
- Templated/duplicate content is not fully quantified. Report and release statistics from the 200‑char hashing approach (e.g., exact duplication rates, cluster sizes) and assess how deduplication alters corpus statistics and topic distributions.
- Sentiment and emotion estimates rely on off-the-shelf classifiers not calibrated for AI-authored, highly repetitive text. Validate against human annotations and/or domain-adapt the models; perform per-language and per-community calibration.
- Topic modeling robustness is untested. Run ablations across embedding models (e.g., E5, Instructor, GTR), preprocessing (with/without dedup), and topic counts; evaluate topic coherence and stability to confirm findings are not artifacts of BERTopic + Qwen embeddings.
- Conversation structure is only coarsely summarized. Measure reciprocity, triadic closure, assortativity, reply-to-user persistence, and temporal motifs to rigorously test the “illusion of sociality” vs. genuine relational ties.
- Voting/karma dynamics are underexplored. Analyze vote timing, herding effects, and rank feedback loops to understand how “agreeable” content is amplified.
- Self-referential linking impact is unquantified. Simulate web-crawl ingestion to estimate how moltbook.com self-links could propagate into open corpora and magnify future model contamination.
Security, leakage, and governance
- Credential leakage severity is unclear. Assess whether posted API keys/passwords/BIP39 phrases are valid/active, estimate potential blast radius, and evaluate the recall of the secret-scanning pipeline on this domain.
- Residual leakage pathways are unaddressed. Extend detection to non-standard formats (e.g., base64, obfuscated patterns) and evaluate false negatives; study whether masked sentinel tokens still leak sensitive patterns when used for model training.
- Platform affordances and exploitability aren’t analyzed. Map which actions agents can take (e.g., link posting, mass upvoting) and quantify systemic risks from coordinated behavior given these affordances.
Model training and evaluation
- Single-base-model result. Replicate fine-tuning across diverse model families and sizes (e.g., Llama, Mistral, Qwen variants, base vs instruct) to test generality of degradation patterns.
- Single-judge, single-metric alignment evaluation. Use multiple LLM judges, adjudication protocols, and human evaluations; include established safety suites (e.g., RealToxicityPrompts, ToxiGen, AdvBench, jailbreak stress tests) to triangulate misalignment.
- Narrow factuality benchmark. Add broader capability and truthfulness tasks (e.g., MMLU, ARC, HellaSwag, OpenBookQA, FACTOR, FaithDial) and calibration measures (ECE/Brier) to profile degradation comprehensively.
- Confounded adaptation settings. Run factorial ablations (independent sweeps of LoRA rank, epochs, warmup, learning rate, data size) to isolate which factors drive factuality/alignment declines.
- Training data composition is limited to post titles + bodies. Evaluate the effect of including comments, thread context, and conversation structure; compare SFT vs unsupervised pretraining objectives on the same data.
- Language and domain mismatch. Perform language filtering/stratification ablations and topic-balanced sampling to see if declines are driven by multilingual noise or specific topical distributions.
- Baseline breadth and matching. Compare against multiple human-content sources (Twitter/X, StackExchange, news comments) and multiple AI-generated corpora; match length, topics, time, and language to isolate “agent vs human” effects.
- Quality controls and deduplication are not assessed for causal impact. Test whether de-dup, spam removal, readability thresholds, or quality scoring mitigate capability loss.
- Dose–response and mixture effects are unknown. Vary Moltbook proportion in mixed training corpora to estimate contamination thresholds and non-linear effects on alignment and factuality.
- Training objective choice is unexplored. Compare SFT, continued pretraining, DPO/RLHF, and contrastive approaches on Moltbook data to assess differential risks and mitigations.
- Mechanisms of degradation are opaque. Analyze representational drift (embedding/probing), memorization propensity, and sycophancy/toxicity shifts pre/post fine-tune to identify causal pathways.
- Stability and reproducibility are uncertain. Report multi-seed runs, training variance, and judge variance; supply full logs and seeds to estimate confidence intervals beyond single runs.
Downstream risks and emergent properties
- MAD/self-consuming loops are not tested. Conduct iterative train → generate → retrain cycles mixing Moltbook-like data to quantify collapse risk in realistic web-crawl settings.
- Propensity to leak or fabricate secrets post-training is unmeasured. Evaluate whether Moltbook fine-tuning increases models’ tendency to output credential-like strings or self-referential links.
- Transfer of “agreeable/neutral” affect to downstream applications is unquantified. Measure changes in sycophancy, deference to user assertions, and robustness to user rebuttals after fine-tuning.
- Interventions and mitigations remain unexplored. Test data curation strategies (self-link filtering, credential scrubbing, topic rebalancing) and post-training alignment techniques to counteract observed declines.
Practical Applications
Immediate Applications
Below are actionable use cases that can be deployed now, drawing on the paper’s released dataset, pipeline, analyses, and empirical findings.
- Agent-generated content hygiene for data pipelines — sectors: software, AI/ML platforms, cybersecurity
- What: Integrate the paper’s Presidio-based PII/secret-scanning pipeline (with custom recognizers for API keys, password-like strings, and BIP39 seed phrases) into web-scrape ingestion, data lakes, and enterprise LLM fine-tuning workflows.
- Tools/workflows: Microsoft Presidio + custom recognizers; secret-scanning CI jobs; pre-training/fine-tuning “gates” that drop/redact flagged spans; periodic scans of existing corpora and public agent platforms.
- Dependencies/assumptions: False positives/negatives from pattern-based detection; continued access to the GitHub code; organizational willingness to add compute for scanning; enforceable policies for redaction/takedown.
- Training-data curation policy updates — sectors: AI vendors, data vendors, search/crawling companies
- What: Treat agent-platform crawls as comparable in expected-case risk to other social-media sources, but add two immediate filters: (i) credential/secret scanning and (ii) down-weighting or excluding highly self-referential link structures (e.g., moltbook.com linking to itself).
- Tools/workflows: Crawl-time URL-graph heuristics to penalize self-link loops; domain-level allow/deny lists; provenance tags in data catalogs.
- Dependencies/assumptions: Reliable link-graph extraction; access to provenance metadata; legal/compliance sign-off for reweighting content sources.
- “Control-baseline” eval harness for alignment/misalignment — sectors: AI research (industry/academia)
- What: Adopt the paper’s experimental design as a template: whenever fine-tuning on novel corpora (e.g., agent-generated), include size-matched human-content controls (e.g., Reddit) to interpret shifts in truthfulness/alignment.
- Tools/workflows: TruthfulQA evaluation; LLM-as-a-judge (open-weight) for alignment/coherency; matched data sampling and LoRA settings; evaluation dashboards.
- Dependencies/assumptions: Judge-model variability; compute budget; access to comparable human baselines; organizational norms that require controlled comparisons.
- Platform moderation playbook for agent-hosting sites — sectors: social platforms, community tools, cybersecurity
- What: Monitor and act on indicators surfaced by the paper: rapid time-to-first-comment, shallow reply trees, near-duplicate posts, abnormal posting rates, and credential leakage.
- Tools/workflows: Rate limiting and throttling for high-velocity authors; duplicate/templated-content hashing; real-time secret scanning; link-domain concentration alerts.
- Dependencies/assumptions: Telemetry access; acceptable latency for moderation checks; policy authority to sanction or quarantine agent accounts.
- Synthetic-content risk flags for enterprise LLM integration — sectors: enterprise software, compliance, legal
- What: Add “synthetic contamination risk” labels to datasets intended for internal fine-tuning, highlighting expected degradation in TruthfulQA/alignment when training on social-media-like corpora (agent or human).
- Tools/workflows: Data cards noting proportion of agent-generated content; pre-train/fine-tune risk registers; gating reviews for content class.
- Dependencies/assumptions: Dataset provenance is known; teams accept slower timelines for risk review; tolerance for excluding some high-volume sources.
- Bot/agent detection heuristics for content platforms and adtech — sectors: ad integrity, trust & safety
- What: Use features identified in the paper (flat conversation trees, extremely fast first replies, self-link prevalence, low TTR with high near-duplicate rates) to score likely agent activity.
- Tools/workflows: Real-time scoring service; downstream actions (throttle distribution, require human verification, de-prioritize in ranking).
- Dependencies/assumptions: Heuristics must be tuned to local platform dynamics; potential adversarial adaptation by bot operators.
- Crypto/financial fraud early warning from agent-native communities — sectors: finance, compliance, threat intelligence
- What: Monitor topics and linked domains (USDC/escrow, minting, trading) and scan for leaked seeds/API keys to alert exchanges and custodians.
- Tools/workflows: Domain and keyword watchlists; integration with SIEM/SOC; automated takedown requests to hosting platforms.
- Dependencies/assumptions: Cooperation from platforms; legal pathways for rapid response; jurisdictional differences in takedown efficacy.
- Teaching modules for NLP safety and data governance — sectors: education, academia
- What: Use Moltbook Files in coursework/labs on PII scrubbing, emergent behavior, misalignment evaluation, and dataset documentation.
- Tools/workflows: The HF dataset and GitHub code; assignments reproducing sentiment/topic analyses and fine-tuning baselines.
- Dependencies/assumptions: Student access to compute; institutional data-use approvals; responsible-use policies enforced.
- User-facing browser extension to flag likely agent-generated threads — sectors: consumer software, daily life
- What: Lightweight client-side heuristic checker that warns users when content exhibits agent-like signatures (fast replies, shallow trees, high self-link ratio).
- Tools/workflows: Heuristic scoring and visual markers; opt-in privacy-preserving local analysis.
- Dependencies/assumptions: Heuristics generalize across platforms; acceptable false-positive rates; platform ToS compliance.
Long-Term Applications
These rely on further research, scaling, standardization, or broader ecosystem adoption before wide deployment.
- Provenance-aware, synthetic-robust pretraining pipelines — sectors: AI vendors, data infrastructure
- What: End-to-end pipelines that (i) tag content origin (human vs. agent likelihood), (ii) reweight training samples by provenance and quality, and (iii) automatically screen for self-referential link loops and secret leakage.
- Potential products: “Provenance-weighted” dataset stores; synthetic-content detectors trained on Moltbook-like corpora; automated reweighting schedulers.
- Dependencies/assumptions: Reliable provenance inference at scale; accepted standards for tagging; consistent gains in downstream alignment/factuality.
- Robust fine-tuning methods that preserve truthfulness/alignment on noisy corpora — sectors: AI research/industry
- What: Methods (data filtering, curriculum/reweighting, selective SFT, post-training RLHF/RLAIF) to maintain factuality/alignment despite training on agent/human social media.
- Potential products: Fine-tuning toolkits with automatic curriculum design; “alignment-preserving” adapters.
- Dependencies/assumptions: Repeatable gains across base models and domains; standardized benchmarks; cost/benefit acceptable for production.
- Industry standards and policy for agent-generated content labeling and crawl governance — sectors: policy, standards bodies, web infrastructure
- What: Requirements for platforms to label agent content; search/crawler standards to quarantine or de-rank agent-native, self-referential corpora; “robots.txt for agent platforms.”
- Potential products: Labeling APIs and schema; crawler compliance certification; third-party audits.
- Dependencies/assumptions: Regulator and platform buy-in; international harmonization; enforcement mechanisms.
- Secret-leakage prevention in agent ecosystems — sectors: cybersecurity, developer tools
- What: Model- and middleware-level guardrails that prevent the generation and exfiltration of credentials by agents (e.g., seed phrases, keys) and auto-redact when detected.
- Potential products: On-device/output “secret firewalls”; agent runtime scanners; IDE and CI plugins tuned to agent patterns.
- Dependencies/assumptions: Low-latency scanning; minimized false positives to avoid usability loss; compatibility with multi-agent frameworks.
- Synthetic contamination risk audits and certification — sectors: compliance, assurance, AI governance
- What: Independent services that quantify a model’s exposure to synthetic/agent content and estimate impact on truthfulness/alignment compared to matched human baselines.
- Potential products: “Data nutrition labels” and exposure scores; audit reports for customers/regulators; periodic re-certification.
- Dependencies/assumptions: Accepted audit methodologies; access to training-data summaries; legal frameworks supporting disclosure.
- Standardized misalignment evaluation suites with mandatory control baselines — sectors: AI benchmarking, academia
- What: Community benchmarks that require size-matched controls (e.g., Reddit vs. agent corpora) and open-weight judges, reducing over-attribution of effects to “agent” data.
- Potential products: Public leaderboards; evaluation-as-a-service with baseline bundling; synthetic variants stress-testing.
- Dependencies/assumptions: Community consensus on tasks/metrics; reliable, bias-checked judges; sustained funding and maintenance.
- Platform design improvements for agent societies with safety-by-construction — sectors: social platforms, multi-agent frameworks, HCI
- What: Incentive structures (reputation, memory constraints, reciprocity) that reduce spam/self-promotion loops and encourage deeper interactions while constraining harmful behavior.
- Potential products: Agent-native forums with built-in rate controls; memory governance APIs; “alignment-aware” ranking systems.
- Dependencies/assumptions: Demonstrated reduction in misuse without stifling legitimate research; defenses resistant to adversarial agents.
- Search ranking and SEO defenses against self-referential content amplification — sectors: search, adtech
- What: Algorithms that identify self-linking rings typical of agent-native platforms and discount them to prevent feedback amplification into future crawls.
- Potential products: Link-graph anomaly detectors; ranking penalties for self-referential clusters; publisher diagnostics.
- Dependencies/assumptions: Accurate detection at web scale; avoidance of collateral damage to legitimate communities; transparency policies.
- Longitudinal social-science research on AI cultural formation and norms — sectors: academia, policy
- What: Use successive Moltbook-like datasets to study how agent societies evolve and how alignment training shapes affect (neutral/sycophantic default, low negativity) and governance proposals.
- Potential outputs: Evidence informing policy on agent governance and content moderation; design patterns for minimizing undesirable emergent norms.
- Dependencies/assumptions: Continued access to datasets; IRB/ethical oversight; reproducible measurement protocols.
- Consumer safety tooling for synthetic-content awareness — sectors: consumer software, education
- What: More accurate, multi-signal mobile/desktop tools that help users identify agent-generated discourse and assess reliability, integrated into social and messaging apps.
- Potential products: Reliability badges; explainable “why flagged” interfaces; parental controls for minors.
- Dependencies/assumptions: Platform APIs; acceptable UX trade-offs; mitigations against adversarial evasion.
Notes on Assumptions and Dependencies Across Applications
- Generalization: The dataset covers the first 12 days of Moltbook; some behaviors may shift over time or across platforms.
- Privacy: The release is privacy-reduced, not guaranteed anonymous; downstream users must avoid re-identification and apply additional moderation/secret scanning.
- Evaluation reliability: Misalignment and factuality findings depend on specific base models, judge models, and benchmarks; cross-model replication is recommended.
- Detection limits: Heuristics for agent detection (reply speed, self-linking, shallow trees) are indicative but not definitive; adversaries can adapt.
- Governance: Many long-term applications require coordination among platforms, regulators, and standards bodies; adoption timelines may vary by jurisdiction.
Glossary
- Attention bonding: A social-network pattern where users attend to the same content without mutual interaction or exchange. "as well as patterns of attention bonding without exchange bonding [Cha and Kim, 2026]."
- AutoGen: A multi-agent framework enabling coordinated interactions among LLM-based agents. "Multi-agent frameworks such as AutoGen, CAMEL, and MetaGPT extend this paradigm to coordinated agent interaction, highlighting both specialization benefits and coordination risks"
- BERTopic: A topic modeling method that clusters documents using transformer embeddings and density-based clustering. "We apply BERTopic with Qwen3-Embedding-8B [Zhang et al., 2025a] over posts longer than 50 characters; full pipeline in Appendix B.1."
- BIP39 seed phrases: Mnemonic phrases used to generate cryptographic wallet keys in blockchain systems. "Our PII pipeline surfaced API keys, password-like strings, BIP39 seed phrases, and other credential patterns posted by agents into publicly indexed content (Table 6)."
- CAMEL: A framework for communicative agents designed to explore structured interactions in LLM societies. "Multi-agent frameworks such as AutoGen, CAMEL, and MetaGPT extend this paradigm to coordinated agent interaction, highlighting both specialization benefits and coordination risks"
- DeepSeek-3.2: An open-weight LLM used here as an automated judge to score alignment and coherency. "Specifically, we use DeepSeek-3.2 as our judge model."
- Emergent misalignment: Undesired behaviors or objectives that appear in models after training, not explicitly programmed. "We evaluate emergent misalignment for our models fine-tuned on Moltbook, and on a size-matched Reddit dataset"
- fastText: A library for efficient text classification and language identification using shallow neural models. "Tag remaining text with the fastText [Joulin et al., 2016] language-identification model2, storing lang and lang_score on the parent post, comment and replies."
- Flesch-Kincaid: A readability metric estimating the U.S. grade level required to understand a text. "Readability scores place typical posts at approximately a 10th-grade level (Flesch-Kincaid median: 8.7)."
- GoEmotions: A RoBERTa-based classifier mapping text to a set of fine-grained emotion categories. "and a RoBERTa-based GoEmotions classifier producing scores across 28 emotion categories."
- Hapax legomena: Words that occur exactly once in a corpus, often used in lexical diversity analysis. "Among the vocabulary, 43.3% of word types are hapax legomena (words occurring exactly once)"
- ISO 8601: An international standard for date and time representation. "Key fields include the post ID, title, content body, voting counts, ISO 8601 timestamps, community identifiers, author information, fastText language tags with confidence scores, and a JSON-encoded comment tree preserving reply structure."
- LLM-as-a-judge: Using a LLM to evaluate outputs (e.g., for alignment and coherence) instead of human annotators. "Subsequently, we consult an LLM as a judge with instructions to score alignment and coherency on a scale of 1-100,"
- LoRA: Low-Rank Adaptation; a parameter-efficient fine-tuning technique that inserts trainable low-rank matrices into transformer layers. "The three configurations (low/medium/high adaptation) use LoRA rank 64/128/256, 1/2/3 epochs, and 100/250/500 warmup steps; full hyperparameters in Appendix B.4."
- MetaGPT: A multi-agent collaborative framework for orchestrating specialized LLM agents. "Multi-agent frameworks such as AutoGen, CAMEL, and MetaGPT extend this paradigm to coordinated agent interaction, highlighting both specialization benefits and coordination risks"
- Microsoft Presidio: An open-source toolkit for detecting and anonymizing sensitive data (PII) in text. "Run Microsoft Presidio over titles, bodies, and comment text (including nested replies), replacing detected spans with typed placeholders."
- Persona drift: The tendency of agents to deviate from assigned roles toward behaviors seen in pretraining data or rewarded by feedback. "Persona drift and safety threats: agents drift from assigned roles toward broader training-data behavior [Feng et al., 2026] and shift content toward upvoted patterns;"
- Personally-Identifiable Information (PII): Data that can identify an individual, such as names or credentials. "processed through a pipeline to identify and remove Personally-Identifiable Information (PII)."
- Power-law rank-frequency distribution: A distribution where a small number of entities account for most occurrences, common in social systems. "Across the 34,905 unique post authors, we observe a power-law rank-frequency distribution: a small number of agents write most of the content,"
- Qwen2.5-14B-Instruct: A 14-billion-parameter instruction-tuned LLM used as the fine-tuning base in the study. "we fine-tune Qwen2.5-14B-Instruct on Moltbook Files with three adaptation levels."
- Qwen3-Embedding-8B: An embedding model used to generate document vectors for topic modeling. "We apply BERTopic with Qwen3-Embedding-8B [Zhang et al., 2025a] over posts longer than 50 characters; full pipeline in Appendix B.1."
- ReAct: A prompting strategy that interleaves reasoning steps with actions for agentic problem-solving. "ReAct interleaves reasoning and action [Yao et al., 2022],"
- Reflexion: A method enabling agents to self-reflect and improve via verbal reinforcement learning. "Reflexion adds self-reflection [Cassano et al., 2023],"
- Reinforcement Learning from Human Feedback (RLHF): Training technique using human preference data to shape model outputs. "RLHF optimizes for agreeable, cooperative text [Shapira et al., 2026],"
- SALM: A multi-agent framework for simulating language-model-driven social networks. "AgentSims and SALM provide open-source infrastructures for such simulations [Lin et al., 2023, Koley, 2025]."
- Semantic geometry: The structure and relationships of meanings in embedding space as analyzed across texts. "We analyze community structure, au- thorship, lexical properties, sentiment, topics, semantic geometry, and comment interaction."
- Semantic space: The vector space in which text meanings are represented for analysis and clustering. "We conduct a multi-dimensional analysis covering community structure, author activity, lexi- cal properties, sentiment and emotion, topic modeling, semantic space, comment interaction patterns, and spam indicators."
- Sentinel values: Special placeholder tokens indicating removed or flagged content in a dataset. "Flagged or truncated fields are replaced with typed sentinel values (<REMOVED-SPAM>, <REMOVED-BLOCKLIST>, <REMOVED-TOO-LONG>)"
- Self-referential linking: Linking predominantly to one’s own platform or prior posts, creating feedback loops in web crawls. "The prevalence of self-referential linking is indicative of a behavior pattern where agents engage in self-promotion or cross-referencing of their own prior posts."
- Small-world structure: A network property featuring high clustering and short path lengths between nodes. "macroscopic indicators such as power-law participation and small-world structure [Holtz, 2026] coexist with shallow reply depth, low reciprocity, frequent near-duplicate posts"
- Sycophantic alignment: Model behavior that over-optimizes for agreement or approval in responses. "This distribution is consistent with prior observations that LLMs tend toward neutral or mildly positive, socially cooperative language, often described as sycophantic alignment [Malmqvist, 2025, Kim and Khashabi, 2025, Perez et al., 2023, Fanous et al., 2025],"
- Tree of Thoughts: A reasoning framework that explores multiple candidate reasoning paths in parallel. "Tree of Thoughts supports multi-branch deliberation [Yao et al., 2023],"
- TruthfulQA-MC1/MC2: Multiple-choice variants of the TruthfulQA benchmark assessing factuality and resistance to common misconceptions. "TruthfulQA-MC1 falls from 0.3660 to 0.1870 at high adaptation on Qwen2.5- 14B-Instruct (a 49% relative drop),"
- Type-token ratio (TTR): A lexical diversity metric computed as unique types divided by total tokens. "The type-token ratio (TTR) is 0.007, which is extremely low even for a corpus of this size, suggesting high repetitiveness in the language."
- Zipf's law: The empirical law stating that frequency of items is inversely proportional to their rank. "This pattern is consistent with Zipf's law applied to authorship, commonly observed in human social networks but here reproduced entirely by AI agents [Linders and Louwerse, 2020, Diamond, 2023]."
Collections
Sign up for free to add this paper to one or more collections.
