Sam's News — tech-research — 2026-08-18¶

AI¶

7 Study: AI-generated images often untraceable to training data as datasets grow¶

MIT CSAIL researchers identified "attribution decay" in large generative AI models: as datasets expand, individual training images become increasingly disconnected from generated outputs, making attribution impossible at scale. The finding has implications for lawsuits, licensing agreements, and regulatory responsibility frameworks.

Phenomenon termed 'attribution decay' as datasets grow
At large scale, removing single images produces no change in model output
Tested on model trained on artwork from 744 public domain artists
If removing training data changes nothing in output, that data cannot be attributed to the output
Implications for artist attribution, licensing, and regulatory responsibility

Sources: MIT News AI Web Searched, MIT — Computers RSS, MIT — Artificial Intelligence RSS, The Register RSS

6 Survey addresses intellectual property risks in visual generative AI¶

A new arXiv survey documents intellectual property risks in visual generative AI, including unauthorized learning, reproduction, and misuse of protected data and models.

Sources: arXiv — Computer Vision RSS, arXiv — Cryptography and Security RSS

6 EgoTac enables tactile prediction from egocentric vision for robot manipulation¶

A new method predicts tactile sensations from egocentric video to improve robot dexterity without requiring large-scale tactile sensor data.

Sources: arXiv — Computer Vision RSS

AI & Healthcare¶

6.5 LLM physician recommendation algorithms show demographic and reputation bias¶

Researchers audit large language model systems that recommend doctors and find they embed biases linked to physician reputation signals and demographic factors.

Sources: arXiv — Computation and Language RSS

AI Safety¶

6.5 Open-weight models show refusal gaps for low-resource languages like Somali¶

Safety evaluation of instruction-tuned models on native-authored Somali content reveals that safety guidelines apply unevenly across languages.

Sources: arXiv — Computation and Language RSS

AI Policy¶

6.5 Global AI Regulations for FAIR and Ethics: Comparative Review of EU, US, China¶

A comparative matrix analysis reveals how EU, US, and China regulatory approaches differ for high-risk AI, highlighting cross-jurisdictional compliance challenges.

Sources: arXiv — Artificial Intelligence RSS

Security¶

6.5 Georgia Tech Researchers Address Water System Vulnerabilities to Cyberattacks¶

Georgia Tech research points to solutions for protecting municipal water systems against recent cyberattacks across the United States.

Sources: Newswise Web Search

AI Research¶

6.5 GRPO Beyond English: Multilingual Study of Group Relative Policy Optimization¶

Researchers conduct a large-scale empirical study of GRPO, a reinforcement learning technique for improving language model reasoning, across non-English and multilingual settings.

Sources: Apple Machine Learning Research RSS

6 Separating source interpretation from decision aggregation in LLM reasoning improves accuracy¶

Research shows that decomposing evidence interpretation from conclusion synthesis in multi-source LLM tasks yields better performance than single-prompt concatenation.

Sources: arXiv — Computation and Language RSS

6 Vision language models achieve code understanding more efficiently than text-only LLMs¶

Research demonstrates that vision language models can understand source code when provided as images, potentially enabling more computationally efficient code analysis.

Sources: arXiv — Computation and Language RSS