Samar Ansari

Research

My current programme treats AI as a systemic risk object. The organising question is whether claims made about AI systems can be independently checked, and the method I use to answer it is epistemic auditing.

The standard

Independent reconstructability. Can an AI claim, output, or system behaviour be reconstructed, verified, attributed, and causally traced, rather than merely judged acceptable? Most existing evaluation stops at acceptability. That is a weaker test, and it lets unverifiable claims pass.

Three pillars

Cost

Auditing AI environmental and compute cost claims for measurement consistency, so that figures reported by different organisations can actually be compared.

Infrastructure governance

Auditing the AI stack, covering training, inference and compute, for whether proposed governance is feasible in practice rather than only on paper.

Knowledge production

Auditing the integrity of the evidence base that institutions depend on, including the peer-reviewed literature itself.

Method

The recurring pattern in this work is to build a typology, attach a feasibility or severity scoring instrument to it, and then use that instrument to surface a structural gap that was not visible before. A second mode is the controlled measurement study, which produces evidence without training new models.

Language technologies and AI governance

A second strand examines multilingual model behaviour and per-language governance, from a native multilingual and non-Anglophone standpoint across Latin, Devanagari and Perso-Arabic scripts. The themes are tokenizer cost, English-anchoring in multilingual models, uneven training-data quality between languages, and per-language accountability under European AI regulation.

Selected current work

AI governance, auditing and sustainability

Hardware-Level Governance of AI ComputeA feasibility taxonomy for regulatory compliance and treaty verification. arXiv, 2026.
Compound Deception in Elite Peer ReviewA failure mode taxonomy built from 100 fabricated citations at NeurIPS 2025. Sole author. arXiv, 2026.
AI Slop and Data Pollution in the Age of Generative AISSRN, 2025. Cited in WIK Discussion Paper No. 546, in the context of the EU Digital Services Act.
Measuring the UnmeasurableA review of AI environmental footprint methodologies across 164 studies, proposing a reporting protocol aligned with Article 13 of the EU AI Act. In preparation.
Systemic and Material Sustainability in Solar PhotovoltaicsAn integrative analysis. SSRN, 2025.

Language technology and clinical NLP

The Slop ParadoxHow synthetic standardization erodes clinical uncertainty. A measurement study using named-entity recognition, uncertainty-language detection and image-text alignment. arXiv, 2026.
A Framework for Ethically-Aware AI in EducationWith S. Jeyaruban. TechRxiv, 2025; under peer review at Education and Information Technologies.
Privacy-preserving federated chatbotsSelf-learning chatbot with federated intelligence, 2025; and a comparative study of self-feeding mechanisms in federated and centralised architectures, IEEE ISCC 2025.

Vision, medical AI, privacy and forecasting

PrivHAR-BenchA graduated privacy benchmark for video-based activity recognition: nine privacy tiers, 1,932 clips, 15 classes, pose keypoints, and an evaluation toolkit. Open dataset. arXiv, 2026.
Multimodal fusion of histopathology images and electronic health recordsarXiv, 2026.
FireNet-LiteA separable convolutional network for ultra-efficient fire detection. Procedia Computer Science, 2026. Part of a family extending the original FireNet model.
Probabilistic wind energy forecastingA GRU-based framework with uncertainty quantification, and adaptive cross-location uncertainty quantification for wind power. 2026.

The full list of 169 items is on the publications page.

Earlier and continuing technical research

Hardware neural networks; analog and current-mode signal processing; machine learning applications; IoT security; privacy-preserving surveillance and human activity recognition; and deep learning for medical imaging. The doctoral work developed feedback neural circuits for solving systems of linear equations, introduced two analog building blocks, and extended the circuits to linear and quadratic programming.

Co-authorship spans Ireland, the United Kingdom, India, the United States, China, Pakistan and Bangladesh.

Grants and funding

Period	Role	Project and funder	Value
2022 to 2027	Co-Investigator	SI2OT, Design of Secure and Intelligent IoT Nodes. FIST scheme, Department of Science and Technology, India.	INR 10,200,000
2020 to 2024	Principal Investigator	Presidential Doctoral Scholarship, Athlone Institute of Technology, Ireland.	EUR 105,200
2020 to 2021	Co-Investigator	DART, Data Analytics for Real-time Troubleshooting. Enterprise Ireland.	EUR 130,265 of a EUR 200,545 project
2019 to 2020	Fellow	Post-doctoral fellowship, COMAND Technology Gateway, Enterprise Ireland.	EUR 40,221 per year
2018 to 2019	Fellow	Post-doctoral fellowship, EU Horizon 2020 PROTECTIVE.	EUR 37,383 per year
2016 to 2018	Principal Investigator	Visvesaraya Young Faculty Research Fellowship, Ministry of Electronics and IT, India.	INR 3,700,000
2016 to 2018	Principal Investigator	Start-Up Project, University Grants Commission, India.	INR 1,000,000

Internal awards at the University of Chester: Vice-Chancellor's AI Innovation Fund, 2025 to 2026; Quality Research funding in 2022 and 2023; conference grants in 2024 to 2025 and 2025 to 2026.

Citation record

4,892 citations, h-index 26, i10-index 51, as measured on Google Scholar on 18 August 2026.