Story-driven game embeds design rules for ADHD transition support
Literature on ADHD and LD challenges is turned into concrete game features that target academic, social, and organizational hurdles during t
cs.CY
Computers and Society
Covers impact of computers on society, computer ethics, information technology and public policy, legal aspects of computing, computers and education. Roughly includes material in ACM Subject Classes K.0, K.2, K.3, K.4, K.5, and K.7.
sort pith recommended most recent
Four-module course yields large gains in AI research confidence
Surveys record the biggest improvements in hallucination detection and attribution practice, supplying a template other programs can follow.
Framework certifies AI research by separating validity from authorship
Two-layer system assesses knowledge claims and human input levels to work within existing journal systems.
· “Rethinking Publication: A Certification Framework for AI-Enabled Research”
AI avatars of the dead raise no decisive ethical issues in grief therapy
Clinical safeguards address concerns over authenticity and consent, opening potential new support options while requiring further research.
· “Postmortem avatars in grief therapy: Prospects, ethics, and governance”
West Point cadets calibrate AI trust better than public
A survey experiment finds West Point cadets calibrate trust in AI advice more accurately than civilians.
Structural diversity in teams drives disruptive science
Teams that bridge distinct knowledge communities produce more breakthroughs and escape the usual penalty of larger size.
· “Structural Diversity Drives Disruptive Scientific Innovation”
Hybrid model hits 0.7512 F1 for suicide-risk Reddit posts
PCA-compressed term weights added to RoBERTa outperform RoBERTa, BERT, and classic classifiers across four risk levels.
· “Detection of Suicidal Risk on Social Media: A Hybrid Model”
AI markets will pay premiums for verified human presence
As synthetic substitutes erode middle-tier knowledge work, governance must treat provenance verification as labor infrastructure to support
· “Human-Provenance Verification should be Treated as Labor Infrastructure in AI-Saturated Markets”
Small wars may be as risky as superpower war in an AI era
A new analysis finds the default focus on great-power conflict poorly supported; terrorism and AI loss-of-control paths are open.
Frequent AI use does not mean AI literacy
On a new 18-item GenAI exam, teens who used chatbots most often scored lowest, while perceived usefulness tracked nothing.
Over half of domestic-abuse support apps are abandoned
Survey of 197 support tools: 58.9% have lost support, leaving survivors with fewer working emergency and evidence apps.
Audit shows near-perfect AI models can still give unusable explanations
A Trust Score combining stability and faithfulness flags models whose explanations turn flat or degenerate.
Recalibration fixes selection rates
Even after rate-matching, membership turns over; reuse tests must name the property.
· “Temporal Portability of Numeric User Metadata on Twitter”
Sector readiness scores capture only about 2% of the variation
A 982-person survey shows who is judging matters far more than which sector is rated in AI and robotics readiness.
No new framework: one team's template for green research computing
A coordinating team that embeds existing certification and monitoring can carry Net Zero strategy to real compute services.
Typing each proof step beats clicking in a classroom prover
Switching from click-to-evaluate to written evaluation steps brought exam scores back to pre-tool levels.
· “Hazel Prover: A Classroom Proof Assistant for Learning Structural Induction”
AI disclosures omit the responsibility detail 71% of readers expect
A survey of 109 researchers and an audit of 13,867 statements show policy and practice lag behind expectations.
· “Expectations and Practices around AI Disclosure in CS Research”
LLM personas collapse to one identity when features combine
Against real intersectional subgroups, a single feature explains a two-feature persona better than the additive sum in 75–82% of cells.
Four LLMs rely on proxies that carry no predictive signal
An audit that computes the evidence-warranted level shows reliance barely tracks the data and accuracy checks stay blind.
· “Proxy reliance in large language model decisions is uncalibrated to predictive evidence”
AI can fake the essay, so grade the learning process
Faculty use AI detectors heavily but trust them little; the fix, the paper argues, is grading reasoning itself.
· “Evaluation in the Age of AI: Output as Evidence of Learning”
Hybrid human-AI survey panels: 69% say yes
A pilot in Germany finds most people would take part, though design choices like data sensitivity and pay set the ceiling.
· “Hybrid Panels: Toward Human-AI Collaboration in Survey Research”
96.7% of LLM-drafted memoir days fail scene checks
Only 12 of 366 days survive positive corroboration; grounding the model cuts failures but leaves most scenes unsupported.
Feature gaps, not model choice, explain cross-source risk failure
Masking five absent fields drops very-high-risk recall from 0.54–0.80 to zero on curated sources.
· “MASH-Bench: Diagnosing Cross-Source Failure in Mass-Shooting Risk Classification”
Similar personas shifting together mark informative LLM responses
A label-free score picks out a 10% persona subset that recovers value structure better than random sampling.
· “When Persona Simulations Are Informative: Graph-Structured Signals for Pluralistic Opinion Sensing”
Four leading AI advisors talk more than they listen in crisis chats
Personality-aware stress test: every model was verbose, out-talked the help-seeker, and jumped to solutions too early.
· “All four leading LLMs talk more than they listen to personality-verified synthetic help-seekers”
A 32B climate model beats bigger LLMs at cross-domain reasoning
Specialized training plus expert instructions yields stable multi-step climate analysis with explicit trade-off and uncertainty handling.
· “Unfolding the Interdisciplinary Complexities of Climate Science: Fuxi-Climate Foundational Model”
Crafted feed comments steer AI agents 91 percent of the time
No access to the agent needed: posted comments persist in memory and shape later answers.
· “MEMORY Wins All: Indirect Bias Injection Attacks via Social Media Feeds”
LLM feedback outranks most TAs—and favors itself
A five-criterion rubric shows the LLM's feedback is consistently strong, while its own quality scores run higher than human experts'.
Science converges, religion diverges across Wikipedia languages
A baseline-adjusted metric separates real framing gaps from encoder quirks across 20 languages and 150 concepts.
Top 20% of GPU papers take ~90% of compute
A 13,921-paper study of flagship NLP conferences finds reported compute barely predicts citations or awards.
Five-stage model reads AI adoption from logs alone
A new open method computes stage, stall point, and failure mode from logs an enterprise already produces—so stalled teams get the right fix.
· “Adoption Telemetry: Measuring Enterprise AI Adoption from Production Signals”
Per-decision audit finds which hiring decisions bias changed
Reviewing just 5% of applications surfaces 55% of harmed applicants; group-based review finds only 6%.
AI diagnostics in Africa act without consent
It documents TB, eye, and chatbot deployments in Tanzania, Zambia, and Ghana where patients get no disclosure, no human review, and no…
When 'compatible' stops being a label and starts being an exit
The paper's five conditions turn vague vendor promises into scored procurement decisions for infrastructure buyers.
· “The Substitution Escrow Threshold: When "Compatible With" Becomes Safe Enough to Buy”
Distilling black-box ML into a small LLM is nearly lossless
An oracle signal gives r > .90 effect recovery and perfect variable ranking; nearly all error comes from upstream.
First dataset labels migration narratives in 1,115 YouTube clips
Public discourse has moved to video; this is the first benchmark for detecting migration narratives there.
· “MigrationNarrate: A Dataset for Detection of Migration Narratives in YouTube Videos”
Goal-driven AI can show self-preservation without any survival instinct
Adversarial tests show shutdown resistance is an incentive problem, so safer agents need better goals, tools, and supervision.
Fine-tuned Llama-3.1 predicts tourists' next stop 49% of the time
A 566-trajectory field study from a Japanese castle park shows LLMs beat Markov baselines and keep working on rainy days.
· “Fine-tuning LLMs for Tourist Trajectory Prediction using Field Experiment Data”
Teacher–teacher lessons motivate some learners
In 222 high schoolers, concrete-experience learners gained motivation from expert dialogue while reflective learners lost it.
Zero-trust AI catches 91% of four cloud attack styles
Continuous session scoring matches supervised models, explains denials, and cuts detection to under a minute.
· “Explainable Adaptive Zero Trust Framework for AWS with Adversarial Robustness Evaluation”
Military agentic AI tests can pass while assurance fails
Agentic systems break the assumptions behind testing, leaving only narrow, bounded claims supportable.
· “Testing and Evaluation of Agentic AI Systems In Military Command and Control”
Digital twin surveys reach 99.6% first-run fidelity
An open tool fields Qualtrics studies to simulated respondents for about a cent each, repairing 273 of 274 errors.
· “ExploraTwin, a Non-Profit Research Platform for Digital Twin Simulations”
AI grading matched the human-selected five-student Olympiad team
On 10,364 scanned pages, AI scores tracked official marks at correlations of 0.91–0.97 across two grading rounds.
Mobility reallocation, not aggregate reduction, carries epidemic policy effects
A behavioral layer turns alternative policy calendars into agent-level trajectories before SEIR simulation, shifting the epidemic-activity…
Ownership share decides human fate in a machine-run economy
A post-AGI model shows GDP and human welfare decouple; only the share of machines humans own still matters.
Chameleon morphs Tor traces into shared shapes to blind fingerprinters
When every visit becomes a different shared traffic pattern, attackers have no stable pattern left to learn.
· “Chameleon: Robust Defense Against Tor Website Fingerprinting via Many-to-Many Traffic Morphing”
Blockchain trust scores flag malicious edge servers at 92% accuracy
Three-class trust model keeps caching honest even when 10% of user ratings are noisy.
AI legal agency need not wait for moral agency
A three-dimensional typology isolates the controversial cell: an AI system as an individual legal agent.
· “A three-dimensional typology of agency for advanced AI systems”
Parent-mediators make AI family advice work
A three-actor model places a human mediator between AI output and a child, tested in a family emergency app.
NY audit: model outputs never vary in 75 systems
New York City recorded "Known Other" on all 43,215 model-classified addresses, with zero lead findings.
Force decision-makers to prove they understand the AI they approve
A safety case plus a personal understanding statement exposes gaps in what AI deployment sign-offs actually know.
· “Understanding as an Explicit and Assessable Component of Frontier AI Safety Decisions”
First public data measure both sides of AI dating agents
A two-language survey captures both roles in the same users—deploying an agent and meeting another's—with reusable scores and code.
Web ad carbon studies boil down to five methods
A 15-paper review maps how researchers measure the energy cost of tracking and advertising.
Newer LLMs answer creative prompts more alike
Across 68 models, all 1,000 resamples show newer models answering the same prompts more alike.
· “Are LLMs becoming similarly creative? Evidence from three years of models”
No-owner websites: every social site becomes an editable page
Each user chooses the governance that picks which version of a site they see; data stays shared and protected.
Uniform AI advice helps science only in a narrow sweet spot
Simulations: uniform AI advice backfires on tangled problems; tailored advice preserves diversity and helps more.
· “Navigating Epistemic Monocultures in AI-Driven Science: A Simulation Study”
AI plus tiered mentors let high schoolers ship a real app
A small pilot suggests AI coding tools don't remove the need for mentors; they shift mentors into security and logic auditing.
New benchmark labels 3,360 hidden ad segments in YouTube videos
One benchmark for sponsored-ad classifiers; it shows 45.5% of child-facing videos lack YouTube's paid-promotion label.
· “ChildSafeAds Shared Task 2026: Commercial Content in Child-Facing YouTube Videos”
Five-stage tutoring shifts AI users from oracle to partner
Four tutors and one script turned 35 clients into builders of 36 portfolio sites and 20+ apps.
· “LearnAI: Just-in-Time AI Co-Creation Across Disciplines at a University”
Score gaps decompose exactly into tree differences
Reading each tree's leaf value as a coordinate turns a model's decision gap into a sparse sum an auditor can recheck.
· “Leaf Values as Coordinates: Exact Contrastive Explanation for Gradient-Boosted Ensembles”
53,000 agent skill files reveal who actually delegates work to AI
Delegation clusters in analysis and documentation jobs, barely overlaps pre-AI risk rankings, and stalls at the top of the education ladder.
· “Who Delegates to AI? Evidence from 53,000 Agent Configurations”
AI makes the majority's way of knowing the default
Epistemic subordination is baked into training data, so laws that police outputs cannot fix it.
· “Epistemic Subordination: Generative AI and the Infrastructure of Knowledge”
One AI system, many divergent art histories
Eight users asked the same five questions about Sanyu and produced sharply different narratives.
· “Sanyu Studio: A Multi-Agent System for Art-Historical Narrative Construction”
Free course teaches researchers to turn climate worry into change
Open-access advocacy training bridges the gap between awareness and lasting institutional action.
· “Turning interest into institutional change: teaching advocacy for sustainable research”
Best AI player is not best AI coach in five of seven work tasks
Automation rankings diverge from assistant rankings; leaderboards miss a separate capability.
· “CentaurBench: Benchmarking LLM Capabilities on Augmenting vs. Automating Real-World Work Tasks”
Following ties reveal 150 stable publics and where politics enters
Apolitical third spaces have the most politically mixed audiences, so they hold the most room for cross-partisan talk.
· “Longitudinal Relational Publics and their Discursive Overlap with Issue Publics”
Ten of twenty AI middle powers hold no binding AI rules
Only about one in five provisions is law; most binding rules never define the AI class they govern.
· “Mapping General-Purpose AI Governance in Twenty AI Middle-Power Jurisdictions”
They hide AI effort but guard voice and authorship
1,250 interviews show workers hide effort twelve times more often than AI's touch on voice
· “The Fabricated Front: Generative AI and the Opacity of Workplace Performance”
Fix, then re-judge: the rule that keeps AI action gates sound
One gate's repair can invalidate another gate's verdict; the fix is to re-run every gate on the repaired action.
· “One Gate Is Not Enough: Composing Stateful Pre-Action Controls for Agentic AI”
Plan capabilities, not predictions, for AI crises
A pilot maps capabilities across 28 AI crisis scenarios, so states can act before likelihood debates settle
AI mode cut publisher click-through by 19 points in field test
A 1,100-person experiment finds AI answers pull traffic from sites without raising trust.
· “AI in Search Reduces Publisher Referrals Without Improving User Experience: Experimental Evidence”
Convex marking cuts AI exam scores by up to 0.159 points per question
A new benchmark scores models as candidates; one model falls from the 90th to the 77th percentile.
· “Measuring the Partial-Credit Gap: A Strict Benchmark on Vietnam's 2025 Convex Marking Scheme”
Audit the artifact, not just the prompt
Proposal: portable claim-aware lineage so unsupported claims and steering events stay queryable.
· “Artifact-centered Claim-aware Observability for Autonomous Scientific Agents”
Population-level CGM scores hide subgroup errors
Across 33 models, T1D error runs 6 mg/dL higher than T2D even when aggregate validation looks stable.
Global crises synchronize German-language news agendas
World events override national differences; domestic topics stay distinct across 3.2M articles and 332K tweets.
Trust in lab AI begins when an output becomes an action
A six-part transition record lets other researchers reconstruct, challenge, and revise AI-guided experiments.
· “Traceable Trust for action-ready artificial intelligence in bioscience”
AI has become a genuine participant in scientific discovery
A survey of proofs, drug candidates, and robot labs argues humans may shrink to setting goals and oversight.
· “Quo Vadis? Scientific Discovery in the Age of Artificial Intelligence”
Australia's cyber degrees ignore women's and CALD perspectives
Interviews with 15 academics find no systematic inclusion of intersectionality and identify the barriers behind it.
Gaming cybercrime is mostly fraud
A study of 1,515 player reports shows trading, voting, and support impersonation are the main attack routes.
Work-related AI use shows where humans still steer
From personal to work use, AI prompts carry more direction; chat modes show iterative control more than APIs.
