Essay
Why a satellite engineer cares about AI safety
I work on satellite technology. People sometimes ask why someone in my field spends evenings thinking about AI safety. My honest answer is that the habits satellite work teaches you are exactly the habits AI needs.
I'm excited about what AI could do: help end diseases, grow more healthy food with less waste, and give people everywhere access to expertise that used to be rare. But I've also learned, from building systems that have to work far from home, that capability is only half the job. The other half is making sure the system does what you meant, for everyone it touches, even when someone is trying to make it fail.
Here are four lessons from my side of engineering that shape how I think about AI.
01You can't patch it after launch
Once most satellites are in orbit, nobody can reach them with a screwdriver. That changes how you work. You don't ask "does it work?" You ask "how could this fail, and have we tested every one of those ways?" You test on the ground, under conditions harsher than space, because a mistake after launch is very hard to undo.
AI models are different in one way: they can be updated. But once millions of people rely on a model, a failure has already reached them by the time anyone patches it. That's why I care about evaluations: careful tests, run before deployment, that ask how the model could fail and for whom. A model that is safe in the cases nobody tested hasn't been shown to be safe.
02Watch the telemetry
A satellite talks to the ground all the time: temperatures, voltages, orientation, error counts. Teams watch that telemetry for anything unusual, because small anomalies are often early warnings of big failures.
AI systems need the same discipline. Some of the most important safety work isn't about the model at all. It's about noticing how the model is being used: spotting patterns that suggest fraud, harassment or other abuse, and acting before harm spreads. You can't stop harmful actors you never see.
03Everyone under the footprint counts
A satellite's footprint covers whole regions: many countries, many languages, many ways of living. You design for everyone under it, not just the people who happen to be near the engineers.
AI is the same, except that most safety testing still happens in standard English. That creates two quiet failures. Honest users who speak differently, like the many millions of Nigerians who use Pidgin every day, can get refused or lectured for harmless messages. And a guardrail that only works in English may not stop a harmful request written in Pidgin. Either way, the people under the footprint get a worse, less safe product. That's the problem my project Salecta sets out to measure.
04Expect someone to jam the signal
Satellite links face deliberate interference: jamming, spoofing, people trying to make a system believe something false. You can't design only for accidents. You also have to assume an adversary.
AI safety works the same way. Scammers, fraudsters and abusers will look for the phrasing, language or framing that slips past a model's guardrails. A safety measure has to hold up against someone actively trying to defeat it, without punishing the honest majority. Getting that balance right is, to me, the heart of the problem.
What I'm working toward
AI that helps end disease, feeds people well, and earns a relationship of trust, safety and security with every human it serves.
That's a long road, and I'm starting with something small and concrete. Salecta tests whether today's AI models make the same safety decisions in Nigerian Pidgin as in English: whether they help honest people and refuse harmful ones, whichever way someone speaks. It's open source, checked by native Pidgin speakers, and I'll publish the results here, including whatever doesn't go the way I expect.
If you work on evaluations, multilingual safety, or trust and safety, I'd love to compare notes. You can find me on GitHub.