Our mission
Help people from technical and non-technical backgrounds develop the knowledge, judgement and practical skills to evaluate AI systems, recognise their limitations and contribute to safer development and use.
Learning across disciplines
AI Safety
in Practice
Better questions. Clearer evidence.
Meaningful human oversight.
C Combinator brings people across disciplines together to explore how AI systems behave, where they fail and how people can stay responsible for consequential decisions.
Our proposed AI safety learning programme connects practical exercises with deeper questions about evaluation, authority and accountability as AI systems become more capable.
Why this matters
A convincing answer
is only a starting point.
An AI system can produce a convincing answer while missing important evidence. When it also uses tools, the questions extend beyond whether its answer is correct: what is it allowed to do, who can intervene, and what record will explain its actions?
These are practical engineering and governance problems. They also connect to broader concerns about the risks of increasingly capable AI.
Workshops can help people understand those problems and begin useful work. They cannot establish that a model is safe, replace expert review or prove that a particular safeguard will prevent every failure.
Our purpose
Learning that strengthens
human oversight.
Our vision
A community in which people can question AI systems, test consequential claims and turn learning into useful work that strengthens human oversight as AI capabilities grow.
What guides us
Our values
Public benefit
Choose learning activities for their contribution to people and society. Make their educational purpose clear and distinguish it from commercial promotion.
Evidence and honesty
Show sources, assumptions and limitations. Learn from failed tests as well as successful demonstrations.
Human agency and accountability
Make permissions, responsibilities and human decisions explicit. Ask who can intervene and what happens when something goes wrong.
Care with capability
Use controlled, defensive exercises. Avoid exposing sensitive data or creating unnecessary opportunities for harm.
Accessible, rigorous learning
Welcome different disciplines and levels of experience. Explain unfamiliar ideas clearly while maintaining standards of evidence.
Collaboration and independence
Invite diverse perspectives, disclose relevant interests and evaluate ideas on their merits.
The work to explore
Questions worth testing.
What supports this conclusion?
Explore source quality, uncertainty and ways to test an AI-generated claim. Identify what evidence is missing and what would justify a different conclusion.
What is the system allowed to do?
Map tool access, permissions and human-review points. Distinguish an instruction to an AI system from an independently enforced limit on its actions.
How could the workflow fail?
Examine outdated information, misleading instructions inside retrieved material, insecure integrations and tests that reward the wrong behaviour.
What would make the decision reviewable?
Consider the records needed to understand an action, investigate an incident and improve the workflow without retaining unnecessary personal information.
These learning areas connect reliability, security, governance and safety. They are related, but they are not interchangeable.
From interest to useful work
A practical learning pathway
Understand
Start with a concrete failure mode and the people it could affect.
Test
Use a bounded example to examine what the system does and what the evidence supports.
Review
Explain the test, identify its limitations and invite challenge from another discipline.
Continue
Choose a useful next step: deeper study, a better evaluation, a carefully scoped project or collaboration with relevant expertise.
Confirmed · 25 November 2026
AI Workflow Review
AI Workflow Review is confirmed for our event on 25 November 2026. See the existing event page for details.
We are developing a facilitated exercise using a synthetic AI-assisted workflow. Participants would examine an AI-generated finding, trace its evidence, identify actions that need permission and define points where a person should review or stop the process.
The proposed output is a short workflow-review worksheet containing:
- The claim being tested and its supporting evidence.
- A failure case and a way to test it.
- The actions permitted to the system.
- A human-review or escalation point.
- Remaining uncertainties.
See the event page for the latest programme and attendance information.
View the 25 November eventOur event record
A place to start.
Event records show activity, not independently verified impact. Venue use and individual participation do not imply institutional endorsement.
London · 6 November 2026
AI Safety in Practice at Barclays
Our archive records a C Combinator gathering at Barclays Innovation Hub as part of Barclays Entrepreneurs Festival. See the event record for available notes and photographs.
View the event archiveCambridge · 25 November 2026
AI Showcase & Awards
A standalone C Combinator event. Consult the event page for the latest confirmed programme, times and attendance information.
View event detailsLearning from the programme
Look beyond
attendance.
A useful evaluation could examine understanding, completed review worksheets and feedback about accessibility and usefulness. Attendance, confidence and competence are different measures.
The evaluation approach for the exercise is still in development. No learning gains or risk-reduction outcomes are claimed here.
Further learning resources can help participants explore the field in greater depth. BlueDot is one external resource for AI safety learning and related pathways.
BlueDot is an external resource, not a stated partner or funder of C Combinator.
Support practical learning
Make useful work possible.
Support could help a specific educational activity through facilitation, accessible materials, a suitable venue or evaluation resources. We welcome conversations about a clearly scoped contribution and the costs it would cover.
Commercial sponsorship has a separate prospectus. Sponsorship benefits should not be confused with evidence of safety or an endorsement of a product.
Taking part
A few practical questions.
Do I need AI experience?
We welcome different disciplines and experience levels. Individual exercises may have prerequisites, which should be stated in their programme.
Do I need a Cambridge affiliation?
Our main events welcome people without a Cambridge affiliation. Check the relevant event page for attendance details.
Does an exercise certify that my AI system is safe?
No. These are educational activities. Testing provides evidence within the scope of a test, not a universal safety guarantee.
Who organises the activities?
C Combinator is an independent community founded by Legis Labs. Venue use, individual participation and linked resources do not imply institutional endorsement.
Can I propose a session?
Send a title and outline through the speaker form. A proposal does not confirm a speaking slot.
How is participant information handled?
Read our privacy notice and conduct guidance. Use public or synthetic material in exercises and keep confidential records, credentials and proprietary data in their authorised systems.