We study how AI systems fail people, and build the evidence to fix them.

SRLab was founded by Dr. Syed Raza Bashir and Dr. Shaina Raza, and incorporated under the Canada Not-for-profit Corporations Act. The research programme is led by Dr. Raza and carried out with a small group of scientists, graduate students, and interns from partner institutions in Canada, the United States, and Europe.

Our earliest work was on bias in news text, published at AAAI and ACM AIES and released as open-source software that is still in use. From there the lab grew into multimodal fairness benchmarks, safety evaluation for language models, and, most recently, the governance and evaluation of AI agents. We are a partner in the European Union's Horizon Europe programme through the AIXPERT project on trustworthy AI, working alongside the Vector Institute.

Recent

• June 2026 Our bias detection and rewriting toolkit, unbias-plus, was demonstrated to the Prime Minister of Canada during a visit to the Vector Institute.

• June 2026 Presented HumaniBench and SONIC-O1 at HAICON 2026 in Munich, in connection with the new Vector Institute and Helmholtz Munich research agreement.

• May 2026 Data and Impact Accounting accepted as a spotlight position paper at ICML 2026.

• January 2026 Dr. Raza joined the AI governance panel at AAAI 2026 in Singapore alongside researchers from IBM, Google DeepMind, and Microsoft.

How we work

We release what we build. Benchmarks, datasets, and code are published under open licences so that results can be reproduced and criticised.

We keep clear boundaries. Our members hold positions at other institutions; work done there is attributed there, and lab resources are not mixed with external projects. Our governance policy is below.

We prefer measured claims. Much of our research is about whether AI evaluations mean what they appear to mean, so we try to hold our own findings to the same standard.

Where we would like to go

Over the next few years we want to grow from a volunteer-led lab into a stable Canadian research organisation with funded positions for early-career researchers, particularly on the evaluation science of agentic AI, synthetic-media forensics, and the environmental cost of AI systems. We are seeking public, philanthropic, and industry partners who share that aim.

AI systems are trained on the world as it has been recorded, and the record is uneven. Women, racialised people, newcomers, and people outside English-speaking, high-income settings are the groups most often misdescribed or overlooked by these systems, and least often in the room when they are designed. A large part of our research exists to make that visible and measurable.

What our research has found

Image generators change who they draw depending on the grammatical gender of a word, and assign occupations along stereotyped lines. Vision-language models rate the same content differently depending on the apparent gender or ethnicity of a face. Audio-video models perform unevenly across the demographic groups in our SONIC-O1 benchmark. Bias-detection tools themselves give different verdicts depending on how they are run. These are not abstract concerns; they shape hiring screens, content moderation, health information, and news feeds.

How the lab is built

SRLab's research programme is led by a woman. Our board includes a researcher whose full-time work is AI for women's health. Our founders are immigrants to Canada who built their careers here, and most of our students and interns come from groups that remain underrepresented in AI research. We do not think that makes our work better by itself, but it does mean the questions we ask start from lived experience rather than from a checklist.

What we are committing to

Every benchmark we release reports results by demographic group, not only in aggregate. We give priority in student and intern placements to women and to people from underrepresented communities, and we pay for the work wherever funding allows. We publish our methods so that community organisations, regulators, and journalists can check AI systems themselves rather than take a vendor's word. And we are seeking partners, including through Women and Gender Equality Canada and provincial programmes, to turn these findings into changes in how institutions procure and audit AI.

Our work falls into four themes: fairness and bias in text and multimodal models; evaluation and governance of AI agents; synthetic media and deception; and the environmental cost of AI. Below are the projects that are public. Work under anonymous review is not listed until it is accepted.

Fairness and bias

• SONIC-O1

  A benchmark for how well multimodal models understand real audio and video together. 231 videos, about 60 hours, and 4,958 human-verified questions across 13 conversational domains, with analysis of performance by demographic group. Built with the Vector Institute.

  Project pagePaperDatasetLeaderboard

• HumaniBench

  A human-centred benchmark for large multimodal models, organised around seven principles including fairness, empathy, robustness, and multilingual ability. Presented at HAICON 2026 (Helmholtz Munich) and under revision for ACM Transactions on Intelligent Systems and Technology.

• Gender and stereotype in image generation

  Two studies of text-to-image models: how the grammatical gender of a prompt changes who appears in the picture (EMNLP 2025 Findings), and how occupations are assigned along gender and ethnic lines (NeurIPS 2025 workshop).

  Grammatical gender paper

• RESPECT

  A framework for detecting and reducing disrespectful and exclusionary language in online conversation, published in the Natural Language Processing Journal (2025).

  Paper

• BEADs and ViLBias

  BEADs is a dataset for evaluating bias across domains in text; ViLBias extends bias detection to paired images and captions in news. Both are released for public use.

  Datasets and models

• Reproducibility of bias-detection tools in progress

  An audit of whether span-level bias detectors give the same verdict when the sampling seed, inference platform, or surrounding text changes. We argue that unstable verdicts presented with confidence are themselves a fairness harm.

Agents: evaluation and governance

• Transparency in agentic AI

  A survey and framework for explaining what an AI agent did and why, organised along five questions (what, why, how, when, who) and proposing the Minimal Explanation Packet as a record produced at the end of every task.

  Project pagePaper

• TRiSM for agentic AI

  A review of trust, risk, and security management for multi-agent systems built on language models, published in AI Open (2026).

  Paper

• Responsible agentic reasoning (R²A²)

  A framework for agents that carry fairness, privacy, and auditability checks through each reasoning step rather than only at the final output.

  PaperCode

• From features to actions

  Trace-based, trajectory-level evaluation that locates where an agentic system went wrong, in cases where feature-attribution methods cannot.

  Paper

• FairSense-AgentiX

  An agent-based platform for fairness and risk analysis of text, images, and datasets, built with tool selection and self-critique loops. Developed under AIXPERT.

  Project page

• Harness-aware evaluation of agents in progress

  A survey arguing that an agent's benchmark score belongs to the whole evaluation set-up (model, harness, environment, and evaluator), not the model alone. Part of our contribution to the AIXPERT evaluation framework.

Synthetic media and deception

• VLDBench

  A benchmark for detecting disinformation that combines images and text, published in Information Fusion (2025).

  Paper

• Model immunization against falsehoods

  Fine-tuning language models on paired false claims and corrections so that they resist repeating known falsehoods, by analogy with vaccination. Accepted at IJCNN / IEEE WCCI 2026.

  Paper

• F-DPO

  Factuality-aware preference optimisation that cuts hallucination rates in language models without a separate reward model. ACL 2026 Findings.

  Paper

• Deepfake forensics in progress

  A research programme on audio-visual deepfake detection that generalises across generators: a survey of real-world distribution, forensics, and provenance; a trace-labelled forensic benchmark; and detection models that route on signal reliability. Conducted with the Vector Institute under AIXPERT.

Environmental cost of AI

• Data and Impact Accounting (DIA)

  A method for tracking the cumulative carbon and water footprint of open-source models and their derivatives, with a public dashboard. ICML 2026 spotlight.

  Project pagePaper

• Does cheaper evaluation change the conclusion? in progress

  A pilot study on whether quantisation, batching, and benchmark reduction preserve conclusions about accuracy, fairness, and bias, and what each saves in energy and water.

A short selection. Lab members have published more than 100 peer-reviewed papers; full lists are on Google Scholar under each author.

1. Who is responsible? The data, models, users or regulations? A comprehensive survey on responsible generative AI for a sustainable future.Raza, Qureshi, Zahid et al. ACM Computing Surveys, 2025.
2. TRiSM for agentic AI: a review of trust, risk, and security management in LLM-based agentic multi-agent systems.Raza, Sapkota, Karkee, Emmanouilidis. AI Open, 2026.
3. VLDBench: benchmarking vision-language disinformation detection.Raza, Bashir, Emmanouilidis, Shah et al. Information Fusion, 2025.
4. Developing safe and responsible large language models.Raza, Bashir et al. Machine Learning, 2025.
5. Just as humans need vaccines, so do models: model immunization to combat falsehoods.Raza et al. IJCNN / IEEE WCCI, 2026.
6. Data and Impact Accounting for open-source AI derivatives.ICML 2026, spotlight.
7. A jamming attack detection technique for opportunistic networks.Singh, Woungang, Dhurandher, Khalid. Internet of Things, 2022.
8. Reinforcement learning-based fuzzy geocast routing protocol for opportunistic networks.Khalid, Woungang, Dhurandher, Singh. Internet of Things, 2021.

Recognition

• Responsible AI Leader of the Year, Women in AI Summit and Awards 2025 (North America).

• The Peak's Emerging Leaders 2026 in artificial intelligence (Canada).

• Stanford/Elsevier list of the world's top 2% most-cited scientists, 2024 and 2025.

• CIHR Health System Impact Fellowship and AI4PH award, for responsible AI in health equity and public health.

• Partner in AIXPERT, one of three projects selected from 137 proposals under Horizon Europe's trustworthy AI call (grant 101214389).

• Six acceptances at NeurIPS 2025 and a spotlight at ICML 2026, across fairness, sustainability, and agent safety.

• Editorial and review service: reviewer for Nature; editorial board of Springer Discover Computing; NeurIPS main-track programme committee; guest editor for Elsevier journals. Co-organiser of the 2025 AI Governance Workshop.

Scientific Research-AI Lab, operating as SRLab, is incorporated under the Canada Not-for-profit Corporations Act (corporation number 1697413-9, incorporated 7 May 2025) and is governed by its board of directors. The board approves the research programme, budget, and any partnership or funding agreement.

Disclosure of affiliations

Every member's external affiliation, academic or industrial, is listed on this page. Work carried out at another institution is attributed to that institution in publications and on this site.

Separation of resources

Lab funds, data, and computing are used only for lab projects. Members do not use lab resources for work belonging to their employers, and do not bring employer resources into lab projects without a written agreement.

Conflicts of interest

Before a grant application or partnership is signed, the board reviews potential conflicts. A director with a conflict does not take part in the decision. Conflict declarations are recorded in the board minutes.

Open outputs

Publications are made available as open-access preprints. Datasets and code are released under permissive licences unless a data-provider agreement prevents it, in which case we say so.

Questions about our governance can be sent to contact@srlab.ai.