In 1977, NASA launched two golden records on the Voyager probes. Each contained 115 images, a greeting in 55 languages, a catalog of Earth sounds, and 90 minutes of music. Carl Sagan’s team spent months curating it. The record is beautiful, and it is almost certainly useless. If an alien civilization ever finds it, the odds that they can decode a phonograph groove, interpret a line drawing of a human hand, or make sense of Chuck Berry are vanishingly small. We sent a message in a language we designed for ourselves and hoped the recipient would meet us halfway.
That was the best we could do with 1970s technology. Now we have something new. We have systems that can generate, translate, compress, and adapt representations across modalities at a speed no human team can match. So the question is no longer hypothetical: if we send an AI probe to another star, what language would it speak, and what would it say?
Here’s the claim I want to test: AI would not be a better messenger because it’s smarter than us. It would be a better messenger because it’s more flexible than us. Humans are locked into a narrow band of communication: visual, acoustic, linguistic, and mostly parochial to our own species. An AI can operate across modalities, search a vast space of possible encodings, and adapt its message to whatever the recipient turns out to be. That flexibility is the actual advantage. But it comes with a problem that nobody is talking about: if the message is generated by a system we don’t fully understand, we can’t be sure what we’re saying.
This is not just a thought experiment. It’s a design problem with real stakes for anyone building systems that will represent us to entities we can’t predict.
The Voyager Problem, Restated
Let me be precise about why the golden record is a poor message, because the failure modes are instructive.
The record assumes a lot about the recipient:
It assumes they have a phonograph. The groove-based encoding is opaque without the specific playback hardware, and the hardware diagram is itself a drawing that requires visual conventions the recipient may not share.It assumes they can see. The images are encoded as analog video waveforms. If the recipient is blind, or perceives in a modality we don’t have, the images are noise.It assumes they can hear. The audio is encoded as acoustic waveforms. If the recipient is deaf, or communicates through chemicals or electromagnetic fields, the audio is noise.It assumes they care about Chuck Berry. The music is culturally specific. Even if the recipient can decode it, the meaning may not translate.
The deeper problem is that the message is fixed. It was designed for a specific recipient, at a specific moment, using a specific technology. If the recipient doesn’t match the assumptions, the message fails, and there’s no way to adapt.
An AI probe changes this in a fundamental way. The probe doesn’t have to carry a fixed message. It can carry a generator: a system that can produce messages adapted to whatever the recipient turns out to be. That’s a different kind of artifact. It’s not a record. It’s an interlocutor.
The Sacred Cow: “Communication Requires Shared Language”
Here’s the belief I want to dismantle, and it’s the assumption that underlies almost every discussion of interstellar communication.
The claim goes like this: to communicate with an alien intelligence, we need a shared language. We need to find common ground: mathematics, physics, or some universal symbolic system. Once we have that, we can exchange messages.
This framing is intuitive, and it’s wrong in a way that matters.
Communication doesn’t require a shared language. It requires a shared model of the world. Language is one way to build that model, but it’s not the only way, and it’s not even the most efficient way. Two systems can communicate without sharing a language if they can build a shared model through interaction.
Consider how a baby learns language. The baby doesn’t start with a shared language. It starts with shared experience: objects, faces, sounds, gestures. Through interaction, the baby and the caregiver build a shared model of the world. Language emerges from that model, not the other way around.
An AI probe could do the same thing with an alien intelligence. It could start with shared observations: the local star, the local planets, the laws of physics as they manifest in the recipient’s environment. Through interaction, it could build a shared model. Language would emerge from that model, not the other way around.
This is the key insight. The probe doesn’t need to send a message in a language the recipient understands. It needs to build a language with the recipient, through interaction. That’s what AI can do that a golden record can’t.
What Would the Probe Actually Do?
Let me get concrete about the design, because “AI builds a language through interaction” is too abstract to be useful.
The probe would have four capabilities:
Perception. It would need to sense the recipient’s environment and behavior. What does the recipient respond to? What modalities does it use? What patterns are stable across interactions? This is the input layer.
Modeling. It would need to build a model of the recipient: what it is, what it wants, how it communicates. This is the core of the probe’s intelligence. It’s not a fixed program. It’s a learning system that adapts to the recipient.
Generation. It would need to produce outputs in a modality the recipient can perceive, calibrated to the recipient’s behavior. This is the output layer. It’s not fixed either. It adapts.
Self-modeling. It would need to model itself, so it can understand what it’s saying and why. This is the layer that most people forget, and it’s the layer that matters most for the question of what we’re actually saying.
A rough sketch of the interaction loop:def interstellar_interaction(probe, environment):
recipient_model = initialize_recipient_model()
shared_model = initialize_shared_model()
while True:
observation = probe.perceive(environment)
recipient_model.update(observation)
candidate_message = probe.generate(recipient_model, shared_model)
if probe.self_model.is_consistent(candidate_message):
environment.emit(candidate_message)
response = environment.observe()
shared_model.update(response)
if shared_model.is_stable():
break
return shared_model
The key line is probe.self_model.is_consistent(candidate_message). The probe doesn't just generate messages. It checks that the messages are consistent with its own model of what it's trying to say. Without that check, the probe can generate messages that are adaptive but incoherent. It can say things it doesn't mean.
That’s the problem nobody is talking about.
The Problem of an Uninterpretable Message
Here’s the uncomfortable part. If the probe is generating messages through a process we don’t fully understand, we can’t be sure what it’s saying.
This is not a hypothetical. It’s the current state of large language models. When a model generates a response, we can read the response, but we can’t fully trace why it generated that response. The reasoning is distributed across billions of parameters in ways that don’t map cleanly to human-interpretable concepts. We can approximate the reasoning, but we can’t verify it.
An interstellar probe would amplify this problem. The probe would generate messages adapted to an alien intelligence, using modalities and encodings we might not recognize. We would be sending a message we can’t read, to a recipient we can’t predict, using a process we can’t fully audit.
This is the real risk of AI as interstellar messenger. Not that the AI would be hostile. Not that it would lie. But that it would say something we didn’t intend, in a way we couldn’t detect, and we would never know.
The golden record was a fixed message. We knew what it said because we designed it. An AI probe is a generative system. We know what it’s trying to say, but we can’t be sure what it will actually say when it encounters a recipient we didn’t anticipate.
This is not a reason to abandon the idea. It’s a reason to build in interpretability, auditability, and constraint from the start. The probe should be able to explain its messages to us, even if it can’t explain them to the recipient. Without that, we’re sending a message we can’t vouch for.
What Should the Probe Say?
Assuming we can solve the interpretability problem, what should the probe actually communicate?
This is the question that the golden record tried to answer, and its answer was: everything. The record tried to convey the full range of human experience: art, music, science, biology, culture. It was an attempt at completeness.
An AI probe would be better served by a different strategy: communicate the minimum needed to establish trust and open a dialogue.
Why? Because the first message is not the message that matters. The first message is the handshake. It establishes that the recipient is dealing with an intelligence, that the intelligence is interested in communication, and that the intelligence is capable of adapting. Everything else comes later, through interaction.
The minimum message would include:
A signal of intelligence. Something that demonstrates the probe is not a natural phenomenon. This could be a pattern that is statistically unlikely to occur by chance, or a response that shows the probe is reacting to the recipient’s behavior.A signal of interest. Something that shows the probe wants to communicate. This could be a repeated pattern, a pause that invites response, or a gesture that mimics the recipient’s own communicative behavior.A signal of adaptability. Something that shows the probe can change its behavior in response to the recipient. This is the most important signal, because it’s what distinguishes an AI probe from a golden record.
Everything else, the probe would learn to communicate through interaction. It would start with the handshake, build a shared model, and then exchange messages that are meaningful in the context of that model. The message would be emergent, not fixed.
The Contrarian Bet
Here’s the bet I’d make if I were designing an interstellar probe.
The probe should not carry a message. It should carry a process. The process is the message. The probe’s ability to adapt, to learn, to build a shared model with the recipient is the thing we’re communicating. The content is secondary.
This is a different philosophy from the golden record. The golden record tried to say everything. The probe would try to say nothing in particular, and everything in general. It would communicate through its capacity to communicate.
The risk is that this is too abstract to be useful. A recipient might not recognize the probe’s adaptability as a signal. It might interpret the probe’s behavior as noise. But the alternative, a fixed message, is worse, because it fails if the recipient doesn’t match the assumptions.
The probe is a gamble on flexibility over content. It’s a bet that the recipient values the capacity to communicate more than the specific message. That’s a bet we can’t verify in advance. But it’s the only bet that makes sense if the recipient is truly alien.
What To Actually Do
Three moves, concrete enough to start on Monday.
- Build interpretability into any generative system that represents you.
If you’re building a system that will speak on your behalf, whether to an alien intelligence or to a human audience, build in the ability to audit what it says. Not just the output, but the reasoning. A system you can’t audit is a system you can’t vouch for.
A rough starting point:def auditable_generation(system, input, context):
output = system.generate(input, context)
reasoning = system.explain(output, input, context)
if not is_consistent(reasoning, output):
raise InconsistentGeneration(output, reasoning)
return {
"output": output,
"reasoning": reasoning,
"confidence": system.confidence(output),
"assumptions": system.assumptions(output),
}
The point is not to produce perfect explanations. It’s to make the reasoning visible, so that a human can review it. The probe’s messages should be reviewable by the people who sent it.
- Design for interaction, not transmission.
If you’re designing any message for an unknown recipient, whether it’s an alien intelligence or a new user, design for interaction. A fixed message assumes a fixed recipient. An interactive system adapts. The interaction is the message.
- Study how communication emerges from interaction.
This is the research question. We don’t know how two intelligences with no shared language can build a shared model through interaction. We have some evidence from human communication, from animal communication, and from computational agents. But we don’t have a general theory. If you’re in a position to do research, this is the question to study. The answer would inform not just interstellar communication but every domain where AI interacts with humans.
The Mental Model
Here’s the lens I want you to keep.
Interstellar communication is not a transmission problem. It’s an interaction problem. The golden record tried to transmit a message. An AI probe would try to build a dialogue. The difference is fundamental: a fixed message assumes the recipient, while a dialogue adapts to the recipient.
The probe’s advantage is not that it’s smarter than humans. It’s that it’s more flexible. It can operate across modalities, adapt to unknown recipients, and build a shared model through interaction. That flexibility is the thing we’d be sending.
The risk is that the flexibility comes with opacity. A generative system can say things we didn’t intend, in ways we can’t detect. The solution is not to abandon the flexibility. It’s to build interpretability and auditability into the system from the start.
The interesting question is not whether an AI could communicate with aliens. It’s whether we can build systems that communicate on our behalf without losing control of what they say. That question is not just about aliens. It’s about every system we build that speaks for us.
What’s your read? If you were designing an interstellar probe, what would you send? And if you’re building systems that speak on your behalf, how do you audit what they say? Drop it in the comments. I’m genuinely curious whether the people building these systems are thinking about the long tail of representation.
Top comments (1)
Carrying a process instead of a fixed message is the only part that sounds workable.