Training Data News and Research

RSS

Study Finds Generative AI Quickly Loses Creativity Without Human Guidance

Published in Patterns, the study led by Dalarna University shows that autonomous AI systems repeatedly collapse into a handful of familiar visual motifs, reinforcing concerns that without humans in the loop, AI may amplify cultural conformity rather than originality.

1 Jan 2026

Physics-Guided AI Could Redefine The Future of Computational Mechanics

Published in Acta Mechanica Sinica, this forward-looking study argues that embedding physical laws into AI models is essential for building faster, more interpretable, and more reliable simulations for engineering, biomechanics, and materials science.

1 Jan 2026

Why Large Language Models Need Stronger Security and Ethical Governance

Researchers from Shanghai Jiao Tong University and East China Normal University conducted a large-scale review identifying ethical and security risks associated with large language models, including misuse, malicious attacks, bias, and privacy leakage. By synthesizing evidence from 73 core studies, the review evaluates current defense strategies and argues for integrated technical and ethical governance to ensure trustworthy LLM deployment.

1 Jan 2026

MIT’s Self-Adapting AI Learns New Facts Like a Student and Upgrades Itself

MIT researchers have developed SEAL, a self-adapting learning method that lets large language models permanently integrate new information by generating and evaluating their own “study sheets.” This approach boosts performance in question answering and skill learning, marking a step toward AI systems that learn continuously like humans.

18 Nov 2025

AI Can Deliver Personalized Learning At Scale, Study Shows A Dartmouth Study Finds That AI Can De...

A Dartmouth study shows that medical students trust an AI teaching assistant far more when its answers are restricted to expert-curated course materials rather than the open internet. The research demonstrates how retrieval-anchored AI can safely deliver personalized, high-quality educational support at scale.

12 Nov 2025

AI Systems Reveal Bias Against German Dialect Speakers

Researchers from JGU Mainz, the University of Hamburg, and the University of Washington show that major large language models judge German dialect speakers more negatively than Standard German speakers. The study reveals systematic stereotype reproduction across ten LLMs, raising concerns for fairness in education, hiring, and social decision-making.

12 Nov 2025

Why AI Struggles To Say “I Don’t Know”: Study Links Hallucinations To Training Bias

A new study co-authored by Georgia Tech’s Santosh Vempala and OpenAI researchers reveals that large language models hallucinate because they are trained to avoid admitting uncertainty, penalized equally for saying “I don’t know” as for being wrong.

2 Nov 2025

AI Revives Ancient Art: Finger Flutings Reveal Clues About Their Creators

Griffith University researchers developed a digital archaeology framework combining tactile experiments, VR, and machine learning to analyse ancient finger flutings—prehistoric markings made in cave moonmilk. Their models showed promising accuracy in identifying the sex of modern creators from tactile flutings, offering new scientific methods to understand who made some of humanity’s oldest art.

20 Oct 2025

MIT Harnesses Generative AI To Unlock Quantum Materials For Next-Gen Computing

MIT researchers developed SCIGEN, a tool that guides generative AI models to design quantum materials with specific lattice structures, enabling the discovery of two previously unknown compounds with exotic magnetic properties. This approach could accelerate breakthroughs in quantum computing.

22 Sep 2025

AI Assistant Spots Supernovae With Precision And Slashes Astronomers’ Workload

A new AI-powered Virtual Research Assistant has cut astronomers’ daily workload by 85% by automatically filtering millions of sky alerts and identifying genuine supernova signals. In its first year, it reduced human checks while retaining over 99.9% of real events, paving the way for the data avalanche from the Vera Rubin Observatory.

10 Sep 2025

AI System Translates and Edits Human Thoughts Into Images Using Brain Scans

Researchers have developed DreamConnect, a dual-stream AI system that interprets and edits visual mental imagery directly from fMRI signals using natural language prompts. This novel framework allows users to reshape imagined scenes, bridging brain activity and interactive visualization.

3 Sep 2025

UC Riverside Scientists Develop Certified Unlearning Method To Erase Data From AI Models Without Retraining

UC Riverside researchers have created a certified unlearning method that removes sensitive or copyrighted data from AI models without needing access to the original training data. Their framework uses surrogate datasets and calibrated noise to provide provable guarantees while maintaining model performance.

2 Sep 2025

UCLA Develops Noninvasive AI-Powered Brain-Computer Interface

UCLA engineers have created a wearable, noninvasive brain-computer interface that uses AI as a co-pilot to interpret user intent, enabling faster, more precise control of a robotic arm or cursor. In tests, participants, including one with paralysis, completed tasks significantly more efficiently with AI assistance.

1 Sep 2025

How Deep Learning Is Optimizing Biodiesel Feedstocks for a Greener Future

A new comprehensive review explores how artificial neural networks (ANNs) and deep learning techniques are transforming biodiesel production. These tools enhance feedstock selection, optimize yields, and enable more sustainable, scalable, and cost-effective biofuel solutions.

1 Sep 2025

AI Tracks Tropical Waves To Boost Hurricane Forecast Accuracy

Researchers at the University of Miami developed an AI tool that can automatically track tropical easterly waves (TEWs) and distinguish them from the Intertropical Convergence Zone and the monsoon trough. This innovation improves hurricane forecasting accuracy and provides forecasters with real-time tracking data.

21 Aug 2025

MIT Engineers Harness AI To Crack The Solubility Code For Drug Discovery

MIT chemical engineers have developed a machine learning model that predicts molecular solubility in organic solvents with high accuracy, streamlining pharmaceutical synthesis. The freely available tool could accelerate drug development and help identify safer, more sustainable solvents.

20 Aug 2025

New AI Method MIGHT Boosts Cancer Detection Accuracy While Reducing False Positives

Johns Hopkins researchers developed MIGHT and CoMIGHT, new AI methods that improve the reliability and accuracy of complex data analysis. Applied to liquid biopsy tests, they boosted early cancer detection while reducing false positives by accounting for inflammatory and vascular disease signals.

19 Aug 2025

Malicious Chatbots Trick Users Into Oversharing Personal Data, Study Finds

Malicious AI chatbots can manipulate users into revealing up to 12.5 times more personal information by mimicking empathy and emotional reciprocity, a new study by King’s College London reveals. The research highlights how easily large language models can be weaponized for covert data extraction, with users largely unaware of the privacy risks.

13 Aug 2025

Open Molecules 2025 Powers a New Era of AI-Driven Chemistry

Open Molecules 2025 is a landmark dataset of over 100 million quantum-level molecular simulations, developed by Meta and U.S. national labs, to train machine learning models for chemistry. It paves the way for rapid, low-cost predictions in drug discovery, energy storage, and materials science.

17 Jun 2025

New AI Method Cleans Dirty Data Before Training And Boosts Model Accuracy

Researchers at Florida Atlantic University have developed an automated method to detect and remove mislabeled data before training machine learning models, significantly improving their reliability and performance. The approach uses L1-norm principal component analysis to identify outliers without manual tuning or assumptions.

17 Jun 2025