Topics — The twenty-eight kinds of claim this anthology files against — the primary axis of both the practice registry and the reading list, and the same twenty-eight for each (ADR-026). The filing rule is the point: take a domain topic when the claim is ABOUT the domain; a claim merely discovered in one still takes its kind. The axis organizes the record, it does not bound it: content the list cannot place is evidence the list is short (ADR-059).
Deployment and society — what a system does once it is running among people, and how anyone outside it could find out — audits of deployed platforms and recommenders, information ecosystems, moderation and governance, and the access problems and funding pressures of studying any of it from the outside.
14 of 737 LIT documents. Back to the full index.
| # | Title | Summary | Status |
|---|---|---|---|
| LIT-480 | Auditing Political Exposure Bias: Algorithmic Amplification on Twitter/X During the 2024 U.S. Presidential Election | Ye, Luceri and Ferrara (2024), ARXIV-2411.01852. 120 sock-puppet accounts, six weeks, 9.79M tweets from Twitter/X’s “For You” timeline. The method is the transferable part: puppets that never interact, follow sets controlled to the subject, and exposure weighted by rank rather than counted. Accounts following nobody get the most diverse recommendations, so personalization concentrates rather than broadens. | Active |
| LIT-481 | Industry Influence in High-Profile Social Media Research | Bak-Coleman, West, O’Connor and Bergstrom (2026), ARXIV-2601.11507. 49% of social media papers in Science, Nature and PNAS have a disclosable industry tie and most go undisclosed — from 21% of authors, so a concentrated group rather than broad engagement. Ties are over-represented in research on what users share and sparse in research on platform dynamics. | Active |
| LIT-482 | Moving towards informative and actionable social media research | Bak-Coleman, Lewandowsky, Lorenz-Spreen, Narayanan, Orben and Oswald (2025), ARXIV-2505.09254. Individual-level randomized trials do not estimate collective effects, for four named reasons — non-linearity across scale, hysteresis, feedback in time, and SUTVA violation through the network. A null from such a trial is not evidence of no effect, and the positive recommendation is to evaluate specific interventions instead of net effects. | Active |
| LIT-487 | Introducing AI to an Online Petition Platform Changed Outputs but not Outcomes | Corpus et al. (2025), ARXIV-2511.13949. Change.org’s “write with AI” tool reached the US, GB and Canada eleven weeks before Australia. Across 1.5M petitions, access made petitions 49% longer, more lexically varied, 1.5 grades harder to read, and rated higher on quality and persuasiveness — moving them toward the profile of successful pre-AI petitions — while raising inter-petition similarity 23% and leaving outcomes flat or worse. The features that used to predict success stopped predicting it. | Active |
| LIT-492 | The unintended consequences of large language models as a labor-augmenting technology in science | Duede, Gross, Crockett and Bergstrom (2026), ARXIV-2607.17397. A marginal-value-theorem model of how a time-saving tool changes what its users produce, assuming the tool is perfect — faster with no added error and no cost. Removing time raises the long-run rate of return, which is the opportunity cost of time, so developed work is developed less thoroughly. Which work gets developed at all moves in opposite directions depending on the phase accelerated, and thoroughness reverses sign in the third phase — a split the paper’s own closing sentence does not carry. | Active |
| LIT-503 v2 | First Contact: Unsupervised Human-Machine Co-Adaptation via Mutual Information Maximization | Reddy, Levine and Dragan (2022), ARXIV-2205.12381. Proposes scoring a control interface by the mutual information between the operator’s command and the state transition it induces — no labels, no reward, no knowledge of the task. Validated on 540K examples (Spearman ρ = 0.43, in 4 of 5 domains) and used to learn an interface from scratch in under 30 minutes. The fifth domain correlated negatively, and fixing it needs a horizon hyperparameter whose setting requires the task knowledge the method assumes away. | Active |
| LIT-524 | Ghost in the Transformer: Detecting Model Reuse with Invariant Spectral Signatures | Wang, Ma, Xinyi and Li (2025), ARXIV-2511.06390 — take the singular spectra of W_q W_k^T and W_v W_o per layer. Those products are invariant to the permutation and scaling transforms that hide derivation while preserving function, so the spectra fingerprint the model. F1 0.9867 over 55 model pairs, data-free, against 0.9610 for the best prior method. Read as NOTE-268. | Active |
| LIT-697 | AgentSociety: Large-Scale Simulation of LLM-Driven Generative Agents Advances Understanding of Human Behaviors and Society | Piao et al. (2025), ARXIV-2502.08691. The paper presents a simulator of about 10k LLM-driven agents in a realistic urban, social and economic environment. It claims that the simulator’s outcomes on polarization, inflammatory-message spread, UBI, hurricane mobility shocks and urban sustainability qualitatively match real-world empirical findings, which would make it usable as a testbed for social science. | Deferred — seeded from the abstract and a skim on 2026-09-25; not read in full |
| LIT-716 | The Impact of Generative AI on Social Media: An Experimental Study | Møller et al. (2025), ARXIV-2506.14295. In a randomized experiment with 680 US participants, generative-AI writing aids raised engagement and output on a simulated social platform but lowered perceived quality and authenticity, with negative spill-over into threads, and no single tool improved both producer and consumer experience. | Deferred — seeded from the abstract and a skim on 2026-09-25; not read in full |
| LIT-717 | Your Brain on ChatGPT: Accumulation of Cognitive Debt when Using an AI Assistant for Essay Writing Task | Kosmyna et al. (2025), ARXIV-2506.08872. In a small four-month EEG study, participants who wrote essays with ChatGPT showed the weakest brain connectivity, the lowest sense of ownership and the worst recall of their own essays. Those switched from ChatGPT to no tools stayed under-engaged, which the authors call accumulated “cognitive debt”. | Deferred — seeded from the abstract and a skim on 2026-09-25; not read in full |
| LIT-718 | Automated decision-making as domination | Burrell (2024), DOI-10.5210/fm.v29i4.13630. Defining algorithmic injustice as allocative unfairness is a product of computer science’s status and funding, not of the harms themselves. Redefining it as domination, following Iris Marion Young, widens the remedies from code fixes to law, participation, resistance and bans. | Deferred — seeded from the abstract and a skim on 2026-09-25; not read in full |
| LIT-740 | Role-Play with Large Language Models | Shanahan, McDonell and Reynolds (2023), ARXIV-2305.16367 — a conceptual paper, with no experiments. A base LLM behind a dialogue prompt is best described as role-playing the character the context implies. More precisely, it keeps a superposition of characters consistent with the conversation so far and commits to none of them, and every reply narrows the set. The framing lets folk-psychological words be used without ascribing beliefs or goals to the model. It also gives a behavioural test for kinds of falsehood: invented answers vary under regeneration, “good faith” errors do not, and “deliberate” deception shows when the same question is asked from different contexts. | Active |
| LIT-741 | Textoshop: Interactions Inspired by Drawing Software to Facilitate Text Editing | Masson, Kim and Chevalier (2024), ARXIV-2409.17088 (CHI 2025). A GPT-4o-backed text editor that replaces typed prompts with drawing-software operations, each of which sends an engineered prompt. Resizing a selection shortens or expands it, rotating it reorders the words, a colour picker sets tone, boolean operations merge passages, and layers hold versions. In a within-subject study with 12 participants, against a text editor plus a ChatGPT replica on the same model, participants rated themselves more successful (4.6 vs 3.9), gave a SUS of 90 against 69, and finished tasks 55 s faster. The quality of the resulting text was not measured. | Active |
| LIT-752 | Position: AI/ML Influencers Have a Place in the Academic Process | Weissburg, Arora, Wang, Pan and Wang (2024), ICML 2024 position paper. Papers tweeted by the two ML paper-sharing accounts @_akhaliq and @arankomatsuzaki have median citations of 24 and 31 against 14 and 12 for controls matched on venue, year, topic embedding and author prominence, while OpenReview scores of tweeted papers and their controls are indistinguishable. A negative-outcome-control estimate keeps a positive causal effect on being highly cited at every threshold tried, largest for less-cited papers. | Active |