The Eye That Never Closes and Never Asks Permission

An AI system in orbit is now deciding what on Earth deserves attention — without a human analyst, without a regulatory framework, and without the consent of the countries being observed.

Sign up for my Substack daily AI newsletter here.

See my AI Training course portfolio for corporate Business Leaders here.

Article link: https://open.substack.com/pub/johanosteyn/p/the-eye-that-never-closes-and-never?r=73gqa&utm_campaign=post&utm_medium=web&showWelcomeOnShare=true

Follow me on LinkedIn: https://www.linkedin.com/in/johanosteyn/

In April 2026, for the first time in history, an Earth observation satellite identified what it was looking for without being told where to look by a human analyst on the ground. The satellite, Yam-9, was built by Loft Orbital and carried a software package called NAVI-Orbital, developed by NASA’s Jet Propulsion Laboratory, running Google DeepMind’s Gemma 3 vision-language model. Researchers issued natural language queries — identify areas where natural environment meets human development, locate infrastructure around railway hubs — and the model processed the satellite’s sensor data in orbit, returning results without the information ever being downloaded to Earth for human review. The demonstration was published by TechCrunch on 15 June 2026 as the first reported use of a vision-language model in.

The technology works. The governance architecture that should define what it is permitted to do, against whom, for what purpose, and with what accountability to the people being observed does not exist. The gap between those two facts is the most important thing about this story — and it is almost entirely absent from the coverage it is receiving.

CONTEXT AND BACKGROUND

Earth observation satellites have been watching the surface of the planet since the 1970s. The governance question they raised — who has the right to collect imagery of sovereign territory, under what conditions, and with what obligations regarding privacy and use — was managed, imperfectly but meaningfully, by the human bottleneck in the system. Data was collected, downloaded, processed by analysts, and turned into intelligence products that required human judgment at every stage. The time, cost, and expertise required to extract meaningful insight from raw satellite imagery served as a practical constraint on the scale and speed of surveillance.

Yam-9 has changed that architecture. The vision-language model running in orbit does not need an analyst to tell it what to look for. It accepts natural language instructions, processes imagery autonomously, and returns results in real time. Paul Lasserre, Loft Orbital’s head of AI, described the implication directly: “It opens the door to always-on, patrol layers in space. If you have a VLM, you can have logic — like ‘monitor this border for me, and let me know when something is suspicious,’ and interact back and forth with the satellites.”

The commercial context amplifies this. Loft Orbital’s business model is infrastructure-as-a-service. It builds, launches, and operates satellites for third-party customers who analyse and market the data. Kepler Communications operates the largest cluster of GPUs in space and has confirmed undisclosed use cases since its spacecraft launched in January 2026. Planet Labs flies satellites with the same Nvidia Jetson Orin processors used in Yam-9 and has research underway on vision-language model applications. The capacity to deploy always-on autonomous satellite intelligence is not a single company’s capability. It is becoming a commercially available service — available to any government, corporation, or organisation that can afford to task it.

INSIGHT AND ANALYSIS

The governance vacuum that surrounds autonomous satellite AI is not a gap that will fill itself through market forces or technological development. It requires deliberate action by the same international institutions that govern other forms of surveillance, weapons deployment, and cross-border data collection — institutions that have so far been absent from the conversation about what vision-language models in orbit are permitted to do.

The origin of NAVI-Orbital is where the governance argument becomes most precise. The software package was built by a JPL team led by technical lead Juan Delfa Victoria, inspired by an idea from JPL researcher Taran Cyriac John. The original concept was not surveillance. It was assistance — specifically, a voice-and-vision digital assistant for astronauts exploring the Moon or Mars who cannot tap on physical keyboards while wearing pressurised spacesuits. “How about we provide an assistant, like in video games and in movies, where you see an AI which is interactive?” Delfa Victoria explained. The same software architecture that was designed to give suited astronauts conversational autonomy in deep space has been commercialised as a terrestrial border patrol layer. No deliberate decision was made to cross a line. The transition happened because no line existed — and that is the most important governance observation in the entire story.

The outer space treaties that govern satellite operations — the 1967 Outer Space Treaty, the 1972 Liability Convention, the 1976 Registration Convention — were written for a world in which satellites collected raw data and humans processed it. None of them address the question of autonomous AI decision-making in orbit. None of them define what constitutes AI-enabled surveillance of a sovereign territory. None of them establish accountability for the consequences of an autonomous system determining that a specific location, activity, or population pattern is suspicious — and alerting a human operator to act on that determination.

The EU AI Act, which is the most comprehensive AI governance framework currently in force anywhere in the world, defines high-risk AI applications across multiple domains including law enforcement, border management, and biometric identification. It does not address AI systems operating in orbit above national jurisdictions. South Africa’s Draft National AI Policy, gazetted in April 2026 and withdrawn weeks later after hallucinated citations were discovered, does not address orbital AI either. No national or international framework currently in existence does — and the technology it needs to govern has already been demonstrated.

I have previously written about the specific challenge that AI surveillance technology poses for South Africa — examining how the country is not paying adequate attention to the civil liberties, accountability, and regulatory implications of AI-enabled monitoring systems that are being deployed in other jurisdictions and that will inevitably arrive here. The facial recognition story was about surveillance from street level. The Yam-9 demonstration is the same argument from orbital altitude — the same absence of governance, the same asymmetry between what the technology can do and what the regulatory framework is equipped to address, and the same implication for the populations being observed without their knowledge or consent.

The specific capability Lasserre described — monitoring a border and alerting when something is suspicious — carries an accountability question that no existing framework answers. When an AI system in orbit makes an autonomous determination that something at a border is suspicious and alerts a human operator, a consequential decision about a real person or a real situation has been made by a system that nobody elected, nobody can appeal to, and nobody outside the operating company can audit. The downstream human action that follows — whether that is a military response, a law enforcement intervention, a commercial transaction, or a policy decision — is predicated on an algorithmic assessment that the affected parties had no knowledge of, no opportunity to contest, and no recourse to challenge.

IMPLICATIONS

For South African boards, governments, and institutions, the Yam-9 milestone raises three questions that most have not yet encountered and all need to engage with.

The first is about what is being observed. South Africa’s Transnet rail network is being stripped by organised criminal syndicates who know the network’s patterns, schedules, and vulnerabilities better than the operators do. Agricultural land is being degraded by drought and misuse in ways that satellite imagery can detect more accurately and more cost-effectively than ground-level inspection. Mining operations are generating environmental footprints that regulators, investors, and activist organisations can now monitor from orbit in real time and at commercial scale. The question of what autonomous satellite AI can see about South African infrastructure, agriculture, and industry is no longer hypothetical. It is a current commercial capability available to anyone with the budget to task a satellite.

The second is about who is doing the observing. The organisations most likely to deploy always-on autonomous satellite intelligence over South African territory in the near term are not South African. They are US-based companies operating under US regulatory frameworks, serving customers whose interests may or may not align with South Africa’s own governance priorities. The data those systems collect about South African territory, the inferences they draw, and the intelligence products they generate will be governed by the legal and commercial frameworks of the companies that build and operate them — not by South African law, not by African Union policy, and not by any international framework yet in existence.

The third is about the infrastructure that makes it possible. The Nvidia Jetson Orin processors that run vision-language models in orbit are subject to US export controls. The same semiconductor restriction framework that limits AI capability in adversary states determines which countries can build and operate their own autonomous satellite intelligence infrastructure. The African nations that cannot access the compute required to field their own Earth observation satellites with AI capabilities are not simply behind in a technology race. They are structurally dependent on the countries and companies that control that compute for intelligence about their own territory — a dependency that is the space-based version of the AI infrastructure sovereignty argument that the SpaceX IPO made visible from the orbital data centre direction.

CLOSING TAKEAWAY

The eye that never closes has been a governing metaphor for surveillance states and omniscient authorities across centuries of political philosophy. It has now entered commercial operation, in orbit, running on a vision-language model that began as a concept for helping suited astronauts navigate the surface of the Moon. The distance between that origin and a border patrol layer that autonomously flags suspicious activity from space is not a technological distance. It is a governance distance — and it was crossed without any international treaty, national regulation, or corporate accountability framework being updated to reflect the crossing.

That is not a warning about a future risk. It is a description of the current state. The governance frameworks that should define who can deploy always-on autonomous satellite intelligence, against what targets, for what purposes, and with what accountability to the people and territories being observed are not being built at the speed the technology is being deployed. The organisations and governments that wait for those frameworks to arrive before asking the governance questions will be answering them in retrospect — when the intelligence has already been collected, the decisions have already been made, and the consequences are already visible.

Johan Steyn is a prominent AI thought leader, speaker, and author with a deep understanding of artificial intelligence’s impact on business and society. He is passionate about ethical AI development and its role in shaping a better future. Find out more about Johan’s work at https://www.aiforbusiness.net