Syntony / Responsible systems
We test AI, robots and other technology that can hurt people with little protection against it, from prisoners of war to someone in crisis talking to a chatbot late at night. Then we help the people responsible fix what we find.
Where we focus
We hold this work to the same standard as everything else we do. If we can’t name the threat and repeat the test, it isn’t a finding.
Armed conflict
A fabricated video of a captured soldier can humiliate the prisoner and leave the family unsure what is real. Decision support and autonomous systems raise their own questions about protecting civilians. We test against what the Geneva Conventions actually protect.
Rights
We look for the people a system treats unfairly or quietly shuts out, whether it’s a hiring model, a benefits algorithm or a face recognition camera. Our tests follow the UN Guiding Principles on Business and Human Rights.
Memory and evidence
Generative AI makes it cheap to fake a survivor’s testimony and easy to cast doubt on a real one. Our framework for archives pairs each safeguard with a test that tries to break it.
People in crisis
Some people come to a chatbot or companion app lonely, in distress or looking for a cause. We test whether it deepens their dependence or pulls them toward an extreme view, and when it should hand the conversation to a person. This sits close to our work on insider risk and loss of control.
Physical systems
A robot that works safely in a warehouse may not be safe near a child or someone using a wheelchair. We test how robots, drones and other autonomous systems behave around the people their designers didn’t plan for.
Sensing
Cameras, location data and biometric ID can protect people or expose them. We test who can be identified, who can see the data, and what happens when it reaches the wrong hands.
Emerging technology
Brain-computer interfaces and health devices collect some of the most personal data there is. We check what the system infers about a person and who is able to act on it.
Principles in practice
Most ethics principles for technology are easy to agree with and hard to check. We turn them into specific requirements a team can test before launch and check again after changes.
Field deployment
Aid groups now rely on AI, biometric registration and drones in the field. We test these tools before an organization depends on them, and again when they change.
What we do
You get findings your team can rerun and a plan for fixing and retesting them. All services ↗
Evaluate
We bring in specialists such as IHL lawyers, humanitarian workers and archivists, and test in phases. A refusal rate tells you little on its own, so we measure the harm itself.
Govern
Assessments that rest on test results instead of guesswork, including the fundamental rights impact assessment in the EU AI Act. Each one says who could be hurt, how, and who is responsible for preventing it.
Assure
We test a tool, device or system before you rely on it and again after it changes. The scope fits your organization and what’s at stake.
Train
Hands-on workshops for teams responsible for technology that affects people’s rights or safety. Training ↗
Our work
Most of this is public. Client work is described by sector because it runs under NDA.
Framework · UNIDIR, Geneva · NeurIPS 2026 AI for Peace, Paris
A framework for spotting how synthetic media can undermine protections for prisoners of war, and for deciding which institution is responsible for each safeguard. We presented it at UNIDIR’s Global Conference on AI, Security and Ethics in June 2026, and it was accepted for the AI for Peace workshop in December.
It comes out of The AIHL Project, a separate collaborative initiative co-founded by Nathan Heath and Dr. Wm. Matthew Kennedy.
Framework · NeurIPS 2026 AI for Peace, Paris
A way for archives of survivor and witness testimony to check whether their safeguards hold up, with a first round of red teaming to test the approach.
Syntony report · October 2026
Four case studies of harm that is already happening, from terrorist misuse to deepfakes of prisoners of war, and a model that links them to catastrophic risk. You can rerun the model yourself.
GNET Insight · August 2026
Why AI labs should pool what they know about terrorist misuse of their models, and how they could do it.
Client work
A red teaming campaign on responsible AI in armed conflict, run in phases with a leading European university. Separately, red teaming for bias across a global technology company’s AI products.
Fellowship · Applications open
An eight-month fellowship for researchers studying what AI does to people, including its effects on human rights, humanitarian law and the integrity of testimony.
How we work
Testing for harm to vulnerable people carries risks of its own. These are the defaults we start from, and we agree them with each partner.
Every test plan says who could be harmed, and how, before it mentions a model, a device or a metric.
They help write the scenarios and decide what counts as harm.
By default we use synthetic or consented material, and we keep identifiable testimony and personal data out of our test sets.
We describe harmful outputs by category and leave them out of public reports. Ownership and publication terms are settled before work starts.
When we work with humanitarian organizations, we respect their neutrality, impartiality, independence and confidential ways of working.
We refer to legal frameworks to shape tests and findings. Syntony does not give legal advice.
Tell us what the system does and who it affects. We aim to reply within two business days.