Generation starships may be the only way humans travel to other stars. These hypothetical spacecraft would travel at sub-light speed and take generations to reach their destination. Over the hundreds or even thousands of years, generations of human beings would be born, live, and then die on these ships. Even if that awkward arrangement could be made to work, how would everything else function for so long? What about the spacecraft? What about the AI?
AI will undoubtedly play a huge role if we ever build generation starships. AI already provides autonomy for current spacecraft and rovers for tasks like navigation. During decades of space travel, generation ships will have a greater need for autonomy. According to new research, current AI systems simply can't provide the needed capabilities. The research shows that new hybrid forms of AI will be needed to keep these ships running.
The research is titled "The state of hybrid artificial intelligence for interstellar missions." It's published in the journal Progress in Aerospace Sciences, and the sole author is Alex Ellery from the Centre for Self-Replication Research (CESER), Department of Mechanical & Aerospace Engineering at Carleton University in Ottawa, Canada.
Ellery isn't concerned with Large Language Models (LLM), the retail-level AI used in products like Gemini and Claude. He's mainly talking about neural networks. While LLMs are a specific type of neural network designed to understand and generate human languages, other neural networks have more powerful functions.
"Interstellar flight represents an unforgiving environment for autonomous operations with many unknowns imposing the need for advanced artificial intelligence (AI)," Ellery writes. "Evidently, interstellar distances preclude any human intervention from Earth, as diagnostic telemetry downloads and software workaround uploads would take years for the round-trip cycle."
The transit portion of a mission isn't too demanding for AI. Existing AI can, for the most part, handle autonomous guidance, navigation, and control. However, things get more difficult once the cruise portion is over. "The complexity of the flight phase of an interstellar mission pales with that of interception with the target destination," Ellery explains.
NASA's Perseverance rover uses AI to navigate and identify minerals, proving AI's value in space missions. However, the AI needed for a starship would be vastly more complicated. Image Credit: NASA
"For in-situ exploration at the destination extrasolar system, near-human level capabilities will be required for full autonomy following deceleration to encounter the target extrasolar system," Ellery writes, and more powerful AI is needed.
Interstellar missions that last 100 years or more will heavily rely on autonomous systems and AI, not only for operating the spacecraft during travel but also for inevitable maintenance. "High system availability demands that interstellar spacecraft be self-repairable, imposing significant demands on onboard intelligence," Ellery writes.
His work reviews AI's current status and capabilities, and whether it can meet these demands. "In particular, we focus on hybrid AI methods as these appear to offer the richest capabilities in offsetting weaknesses inherent in paradigmic <sic> approaches," he explains.
Generation ships need more than the autonomy that AI provides. They need autonomicity. "Autonomicity involves self-management, including self-reconfiguration to accommodate onboard failures, self-optimization to respond to mission evolution, self-healing to repair damage and self-protection to prevent environmental insults," Ellery explains.
Repairing a generation spaceship on a journey to another star is a massively complex undertaking for AI. The raw materials needed to repair a spacecraft and the manufacturing capabilities required are extremely complex. Think of all the different structural, thermal, and magnetic properties required for materials in different parts of a spacecraft. Managing this and performing repairs while keeping a ship running is an almost overwhelming task. Is current AI up to it?
Ellery identifies a tension between symbolic AI and neural networks that's at the heart of the problem. Symbolic AI is logical and predictable, and operates by logical deduction. But it suffers from brittleness when asked to do too much. Neural networks can adapt by learning, which separates them from symbolic AI. This ability to learn is critical to starships, since the situation is so complex. However, they're considered opaque to analysis. That means that humans, even the engineers who build neural networks, don't really understand how they work.
Ellery's main point is that hybrid AI, based on symbolic AI and neural networks, offers the best of both worlds and might be what's needed for generational starships. "Hybridizing symbolic processing techniques with artificial neural networks appears to offer the advantages of both," he writes, partly because that's how we think. "Human cognition appears to implement both neural learning and symbolic processing."
In his paper, Ellery describes different types of hybrids, like fuzzy neural networks and Markov logic networks. An in-depth discussion of all of these types is beyond the scope of this article.
In his conclusion, Ellery explains that while we can outline what AI needs to do for an interstellar mission, we don't have AI that can do it yet.
This illustration shows a Stanford Torus-based generation ship. It could house tens of thousands of people on a long journey to another star. An AI system that could manage a spacecraft like this and keep it running would have to be almost unimaginably complex. Image Credit: By Heineken11 - Own work, CC BY-SA 4.0, https://commons.wikimedia.org/w/index.php?curid=99940475
A critical part concerns learning. "Artificial neural networks currently implemented suffer from several deficiencies – they cannot learn from single examples and require large datasets," Ellery writes. "Deep learning, in particular, requires extensive sets of training data, which will be scant in interstellar environments."
Another barrier is simply the complexity of an interstellar mission. "Finally, as in all AI systems, problems grow exponentially with the dimensionality of the state space of the real environment," he writes.
For Ellery, it boils down to two main issues. The first is that current AI is weak because it's comprised of different approaches with "narrow specialist capabilities." The second is that we lack a deep understanding of the general mechanisms of intelligence, including our own. These prevent us from building a robust hybrid AI to handle a complex, long-duration interstellar mission.
"Unless these issues are addressed, it is unlikely that we can implement a successful interstellar mission beyond a rapid flyby, assuming that no self-repair will be required," Ellery writes. "This is highly problematic because it severely limits the scientific returns that we can gain from interstellar missions," he concludes.
It's unclear what will happen to a powerful hybrid AI on an extended starship mission, and it's also unclear if we can describe a complex hybrid AI as sane or insane. However, once the system is isolated from the rest of human civilization on a long-duration generation starship mission, it may suffer and break down. This is one of the psychological risks for humans, and AI may suffer similarly.
Though entirely fictional, the HAL 9000 sentient AI from the Space Odyssey might strike a cautionary note regarding AI for space missions. Image Credit: By Tom Cowap - Own work, CC BY-SA 4.0, https://commons.wikimedia.org/w/index.php?curid=103068276
HAL 9000, the AI from the Space Odyssey science fiction series, might provide a cautionary tone for the whole endeavour, even though it's clearly just fiction. As the mission goes on, HAL 9000 begins to malfunction. It starts with minor errors during chess games, and it eventually kills a crew member because it thinks that's the only way to protect the mission.
While this is a dramatized version of sentient AI on a space mission, the underlying issue is real. Even if we can build a hybrid AI that can handle the complexity of an interstellar mission, how will it respond during the long trip? There's no way to test that.
Is it possible it could go "insane"? Since we don't even understand how our intelligence works, we can't answer that.
