Background
Trustworthy AI systems lie at the intersection of software
engineering, machine learning, and real-world application. Across a wide range of
domains, contemporary services increasingly depend on software systems in which machine
learning supports consequential decision-making—triage, diagnostics, risk prediction,
scheduling, and clinical documentation in healthcare, for example—and, like any
mission-critical software, such systems must be subjected to testing, monitoring,
validation, and maintenance. Yet existing academic work predominantly focuses on model
performance measured on retrospective data at development time, while reporting far less
on how a system behaves once deployed—namely, how it performs when the data distribution
shifts, when real data are used and captured, and when surrounding workflows evolve. The
same class of generative methods that underpins these systems also produces content at
scale on social media platforms, where provenance is difficult to establish and where
users encounter such content alongside evidence-based material. Trustworthiness in AI is
therefore a property of the entire sociotechnical system, comprising its engineering,
its models, and the application environments in which they are embedded, rather than a
quality inherent to the model alone.
Goal/Rationale
A model that performs well on a held-out dataset may still fail in service, because its
performance depends on the data, the workflow, and the people around it, and all three
change over time. Privacy-preserving architectures allow multi-site learning without
pooling identifiable records, and dataset documentation makes the training population
explicit. Model cards and assurance frameworks record provenance and intended use.
Finally, human factors evaluation examines whether users can interpret an output and act
on it safely, and continuous monitoring detects drift. Regulatory developments provide
further examples of the need for robust assurance. In the medical-device domain, the
FDA, Health Canada and the MHRA have published joint principles for pre-authorised
change control in machine-learning-enabled devices. This symposium asks what evidence is
needed to establish and maintain trustworthiness in AI systems across different
application settings, and how that evidence should be produced.
Scope
We invite original research, systematic and scoping reviews, methodological papers and
evaluated case studies from academia, industry and the public sector. Contributions are
welcome from software engineering, machine learning, informatics and human factors. We
particularly encourage multidisciplinary teams. The following themes are of particular
interest:
- Software architecture, testing and maintenance for machine-learning-enabled systems,
including those in safety-critical domains
- Verification, validation, benchmarking and assurance methods for trustworthy AI
- Explainable and human-centred AI, including how domain experts interpret and act on
model output
- Federated and privacy-preserving learning architectures for distributed and
multi-site data
- Large language models in documentation, decision support and user-facing
communication
- Detection of and resilience to AI-generated misinformation
- Decision support design, human-computer interaction and workflow integration
Trustworthy AI Systems: Software Engineering,