Synthetic data is not a shortcut. It is a test you must pass.
Guest blog by Anthony Collins, Technical Director Data and Digital Competency Centre; Nathaniel Henman, Data Scientist; John Clay, Machine Learning Engineer and Megan Morrison, Delivery Manager at Faculty
Synthetic data has rapidly become a default tool for many AI teams, especially when access to real-world data is limited, sensitive, or logistically challenging. It allows practitioners to model rare scenarios at scale and enables experimentation without risking exposure of confidential or operationally critical information. However, while synthetic data generation is now more accessible than ever, it is crucial to recognise that producing realistic data is not enough.
Our joint white paper with Faculty sets out why the centre of gravity needs to shift from generation to assurance, and how to build a repeatable way to measure whether synthetic data is fit for purpose.
Why this matters for the defence community, and beyond
Defence is a forcing function for good practice. Data in Defence is more than a technical asset. It can encode tactics, doctrine, force structure, and operational realities. That means there are strong constraints on exposing, analysing, curating, and sharing data, including sensitivity in the operational context, platform limitations on persisting and returning data, and challenges around data ownership and storage.
Those constraints also show up across security, transport, aerospace, and critical national infrastructure. Synthetic data is often the practical route forward. The question is not whether synthetic data can be generated. It is whether it can be trusted.
AI assurance depends on data assurance
The paper’s core argument is simple. If a model is trained or validated using synthetic data, you cannot claim assurance for the model unless you first have an assurance approach for the synthetic data. Current approaches to validation are inconsistent and often lack statistical rigour. Techniques that work for one dataset or purpose may not transfer to another.
This is not a theoretical concern. It shows up in how models behave, and it can show up late, once teams are already committed to a pipeline.
Why “good-looking” data fails
Models do not interpret data like humans do. They learn statistical patterns and establish a decision boundary, the point where the model flips from one class to another. If synthetic data under-represents the messy overlap and variability present in the real world, models can learn a boundary that looks clean in tests but fails when conditions change. The paper sets out this “decision boundary problem” as a key reason why visual realism is not a sufficient check on data quality.
A practical example from the paper uses acoustic-derived imagery data and a binary detection task. This is a demanding setting, with few real positive examples and a high cost of missing a true positive. It illustrates why teams need an evidence-led method for synthetic data, not judgement calls.
What evidence looks like in practice
The paper proposes a structured proof-of-concept approach for validating synthetic datasets against real-world benchmarks, including measures of distribution similarity and related statistical indicators.
Two kinds of evidence are particularly useful.
- Statistical signals that surface synthetic-to-real differences
One example is spectral centroid, a measure that reflects the centre of mass of frequency components in an image. In the study, synthetic data typically had a much lower spectral centroid than real data, across both positive and negative examples. This matters because convolutional neural networks are frequency sensitive. In other words, synthetic and real samples can look distinct to the model even if they look similar to a person.
The point is not that the synthetic data is automatically unusable. It is that differences like this need to be measured, understood, and then addressed through improvements in generation and augmentation.
- Model-based tests that show whether real and synthetic are interpreted similarly
The paper sets out a set of model training experiments that vary the mix of real and synthetic data in the training set, then compare performance on real versus synthetic test sets. It uses PR AUC and an F2 score, which weights recall more than precision. It then treats the difference between real and synthetic test scores as a quality signal. For high-quality synthetic data, those differences should be low, close to zero.
In the reported results, score differences were not near zero when synthetic data was included in training. The paper interprets this as evidence that the model was not processing real and synthetic data in the same way, which is exactly the kind of risk assurance needs to uncover early.
Avoiding a common failure mode, shortcut learning
The paper also highlights a failure mode that many AI teams will recognise in different forms. Models can learn shortcuts. When synthetic data is introduced, one shortcut risk is that the model learns “real versus synthetic” separation as strongly as the actual target classes.
The paper uses embedding analysis to show this effect, with distinct clusters forming by both class and data source. In practical terms, that is a warning sign that the model may be learning the wrong thing, and it undermines confidence that the model will generalise to real-world data.
What techUK members can do now
If you are using synthetic data as part of training, evaluation, or testing, there are a few steps that can improve confidence without slowing progress.
- Define “good data” in terms of the downstream task
Synthetic data needs to cover the distribution your system will encounter, not an idealised version of it.
- Measure differences, do not assume equivalence
Use statistical indicators to identify where synthetic and real diverge, then decide whether those differences matter for the use case.
- Compare performance on real and synthetic test sets
Treat the difference as a signal. If the model behaves very differently across domains, you have a risk that needs resolving.
- Check for shortcut learning
Use embedding diagnostics or similar techniques to see whether models are separating real and synthetic as an unintended proxy.
- Make assurance repeatable
Build these checks into the workflow so you can benchmark quality over time and show continuous improvement, rather than producing one-off assessments.
- Be explicit about limitations
Synthetic data can be valuable even when it is imperfect, but only when teams are clear about where it is strong, where it is weak, and what that means for deployment decisions.
Where this goes next
The paper is a proof-of-concept and a starting point. Its recommendation is to advance both synthetic data generation methods and automated assurance pipelines, with the goal of benchmarking data consistently and reducing the burden on individual projects to invent their own validation approach.
For the defence community, this supports faster, safer progress in AI adoption. For the wider techUK community, the lesson is broader. Synthetic data can be an enabler, but only if the evidence for its quality is treated as a first-order engineering requirement.
Read the paper and download the white paper here.
Defence Programme activities
The Defence programme works to help the UK’s defence technology sector align itself with the MOD. techUK members are able to navigate and better understand the UK Defence sector to successfully align their own investment and resources to take advantage of business opportunities. Visit the programme page here.
Upcoming events
Latest news and insights
Learn more and get involved
Defence updates
Sign-up to get the latest updates and opportunities from our Defence programme.
Meet the team
Fred Sugden
Associate Director, Defence and National Security, techUK
Fred is responsible for techUK's activities across the Defence and National Security sectors, working to provide members with access to key stakeholders across the Defence and National Security community. Before taking on the role of Associate Director for Defence and National Security, Fred joined techUK in 2018, working as the Programme Head for Defence at techUK, leading the organisation's engagement with the Ministry of Defence. Before joining techUK, he worked at ADS, the national trade association representing Aerospace, Defence, Security & Space companies in the UK.
Fred is responsible for techUK’s market engagement and policy development activities across the Defence and National Security sectors, working closely with various organisations within the Ministry of Defence, and across the wider National Security and Intelligence community. Fred works closely with many techUK member companies that have an interest in these sectors, and is responsible for the activities of techUK's senior Defence & Security Board. Working closely with techUK's Programme Head for Cyber Security, Fred oversees a broad range of activities for techUK members.
Outside of work, Fred's interests include football (a Watford FC fan) and skiing.
- Email:
- [email protected]
- Phone:
- 07985 234 170
Read lessmore
Jeremy Wimble
Senior Programme Manager, Defence, techUK
Jeremy manages techUK's defence programme, helping the UK's defence technology sector align itself with the Ministry of Defence - including the National Armaments Directorate (NAD), UK Defence Innovation (UKDI) and Frontline Commands - through a broad range of activities including policy consultation, private briefings and early market engagement. The Programme supports the MOD as it procures new digital technologies.
Prior to joining techUK, from 2016-2024 Jeremy was International Security Programme Manager at the Royal United Services Institute (RUSI) coordinating research and impact activities for funders including the FCDO and US Department of Defense, as well as business development and strategy.
Jeremy has a MA in International Relations from the University of Birmingham and a BA (Hons) in Politics & Social Policy from Swansea University.
- Email:
- [email protected]
- LinkedIn:
- https://www.linkedin.com/in/jeremy-wimble-89183482/
Read lessmore
Authors
Anthony Collins
Technical Director Data and Digital Competency Centre, Faculty
John Clay
Machine Learning Engineer, Faculty
Nathaniel Henman
Data Scientist, Faculty
Megan Morrison
Delivery Manager, Faculty
