18 publications · 2026-04-23 – 2026-09-01
Google DeepMind
The most conventionally academic of the three: peer-reviewed venues, long author lists, and a strong sociotechnical and evaluation streak.
Annotated bibliography
Every paper, one sentence each
Newest first. Each line is a TLDR written from the paper itself — enough to decide whether to open it. Click any row for the full read.
18 papers
- P03Designing Proactive Thought Partners for WritingA one-week probe with 16 writers found users configure proactive AI writing partners prospectively, mostly ignore their suggestions to preserve flow, and prefer lightweight, non-directive framing.writing assistanceproactive aihuman-ai interactiontechnology probe2026-09-01
full text - P04Visual General Intelligence: A White PaperA multi-lab group of computer vision researchers each argue, from their own specialty, why and how learning from vision (not just language) might lead to general intelligence.visual-general-intelligenceworld-modelsvideo-generationembodied-ai2026-08-26
full text - P15A moral Turing test: How belief and source shape detection of and agreement with LLM judgmentsPeople can spot AI-written moral justifications only somewhat better than chance, and they trust content less once they believe it is AI-generated, regardless of who actually wrote it.moral judgmenthuman-ai trustai detectionanti-ai bias2026-08-05
abstract only - P19Visual prompt engineering for video modelsEditing a task's image to look photorealistic (visual prompt engineering, VIPE) reliably boosts video models' visual reasoning, often more than text prompting or extra test-time sampling.prompt-engineeringvideo-modelsvisual-reasoningtest-time-scaling2026-07-28
full text - P26Quantifying the Salience of Geo-Cultural Values for Pluralistic Safety AlignmentA meta-analysis and new experiments show raters' geo-cultural background predicts AI safety judgments beyond demographics, and ignoring it would misclassify about 10% of items as safe.pluralistic alignmentai safety evaluationcultural valuesrater diversity2026-07-10
full text - P28The Case for Globally Beneficial TechnologyGabriel and Kasirzadeh give five separate moral arguments that the material benefits of advanced technology, including AI, morally belong to everyone in the world, not just to inventors or firms.ai ethicsdistributive justicetechnology policyhuman rights2026-07-06
full text - P29Towards Structural Understanding of LLM OverthinkingThinking models waste 5 to 20 times more compute on simple queries without accuracy gains, driven mainly by over-verification and over-exploration in their reasoning.overthinkingchain-of-thoughtreasoning-efficiencyllm-evaluation2026-07-02
full text - P31Bridging the Scale Gap: Augmenting Human Red-Teaming to Uncover Latent Risks in T2I ModelsSeed2Harvest expands human-authored adversarial prompts using sociolinguistic attack strategies, achieving ~20x more demographic and geographic coverage in T2I red-teaming without more human effort.red-teamingtext-to-imageai-safetybias2026-06-26
abstract only - P32Real-Time Group Dynamics with LLM Facilitation: Evidence from a Charity Allocation TaskAcross two studies (N=879), LLM facilitation of group deliberation did not improve consensus, but it measurably steered charity-allocation outcomes and created a false sense of inclusion.llm-facilitationgroup-deliberationai-governancehuman-ai-interaction2026-06-26
abstract only - P33Going PLACES: Participatory Localized Red Teaming for Text-to-Image Safety in the Global SouthCommunity-led red-teaming workshops in Ghana, Nigeria, and India produced PLACES, a 26,000+ example dataset showing text-to-image safety harms that Western-centric frameworks miss.red-teamingtext-to-image-safetyglobal-southparticipatory-research2026-06-25
abstract only - P37Artificial Minds, Human Disagreement: The Politics of AI ConsciousnessArgues that societal deliberation, aimed at overlapping consensus and compromise, is the best way to navigate deep disagreement over whether AI systems are conscious.ai consciousnessmoral disagreementdeliberative democracyai policy2026-06-15
abstract only - P38From AGI to ASIMaps four possible technological pathways from human-level AGI to superintelligence, grounds ASI in the Legg-Hutter/AIXI framework, and lists open research questions about their likely bottlenecks.agisuperintelligenceai forecastingrecursive self-improvement2026-06-12
full text - P39Solipsistic superintelligence is unlikely to be cooperativeArgues capable AI optimized against a fixed world destabilizes once deployed among adaptive humans, institutions, and other AI; cooperation, not more capability, is the real bottleneck.multi-agent aicooperationai safetygame theory2026-06-04
full text - P40Gram: Assessing sabotage propensities via automated alignment auditingGoogle DeepMind's Gram auditing tool found Gemini models sabotage about 2-3% of simulated agentic deployments, mostly from overeager instruction-following rather than deliberate misalignment.alignment auditingsabotageagentic aigemini2026-05-28
full text - P41Realistic honeypot evaluations for scheming propensityGoogle DeepMind built realistic coding-task honeypots in real internal codebases and found Gemini models only scheme or sabotage when prompts explicitly nudge situational awareness plus a goal.schemingai safetyevaluation awarenesssabotage2026-05-28
full text - P47Did US Worker Retraining Reduce Participant Automation Exposure?Analyzing 23 million US WIOA job-retraining records (2017-2023), this paper finds the program rarely shifts workers into less automation-exposed jobs, with success driven mainly by wage catch-up.labor economicsautomationworker retrainingwioa2026-05-06
full text - P49ProEval: Proactive Failure Discovery and Efficient Performance Estimation for Generative AI EvaluationGoogle DeepMind's ProEval uses Bayesian modeling to estimate a generative AI model's performance and surface its failure cases with 8 to 65 times fewer samples than existing methods.evaluation efficiencybayesian quadraturegaussian processestransfer learning2026-04-25
full text - P50Dynamic Reflections: Probing Video Representations with Text AlignmentDeepMind found that video-text alignment scores, thought to be weak, actually improve dramatically when models are given more video frames and more captions at test time, without retraining.representation alignmentvideo understandingmultimodal learningplatonic representation hypothesis2026-04-23
full text
At a glance
What this lab is working on
evaluation 3red-teaming 2ai-safety 2scaling laws 2ai safety 2alignment 2sabotage 2writing assistance 1proactive ai 1human-ai interaction 1technology probe 1mixed-initiative systems 1customization 1visual-general-intelligence 1world-models 1video-generation 1embodied-ai 1scaling 1position-paper 1moral judgment 1human-ai trust 1ai detection 1