In 2023 I wrote about the possibility of using multimodal generative AI to create a simplified version of the “Gestalt” communication described in Greg Egan’s 1997 hard science-fiction novel Diaspora. The basic idea still seems interesting. In fact, developments in generative AI, neuroscience and AI-to-AI communication since then make parts of it considerably more plausible.
However, my original post—and some of the answers produced by ChatGPT 4 at the time—also overstated what Egan actually describes as Gestalt communication, so it is worth starting again.
The more interesting question is no longer simply: Can we use AI and richer media to communicate more information between people? It is: Can AI help two people construct sufficiently similar internal models of an idea, while transmitting much less explicit explanation between them?
That turns out to be a much more useful way to think about a practical, human version of Gestalt communication.
Gestalt communication in Diaspora
Diaspora begins in the year 2975 CE, when humanity has developed into several very different forms:
- Fleshers — biological humans and their many genetically and technologically modified descendants.
- Gleisner robots — software-based minds inhabiting physical robotic bodies.
- Polis citizens — software-based people who live primarily within virtual communities called polises and experience virtual environments called scapes.
Polis citizens have communication and perceptual capabilities radically different from biological humans. Egan describes two standardized information modalities used by citizens: linear and gestalt. He calls them distant descendants of hearing and vision. That is an important distinction.
Gestalt is not simply telepathy, nor is it described as two minds temporarily merging into one consciousness. That was an embellishment in my original interpretation. Instead, Gestalt appears to be an extremely rich perceptual and representational channel that can combine:
- images, symbolic information and non-visual tags;
- object metadata that would normally require substantial explanation, such as an asteroid’s chemical composition, mass, spin and orbital parameters;
- “colour” extending far beyond the visual spectrum humans normally perceive;
- identity and other information broadcast by citizens themselves.
Rather than thinking of Gestalt as transmitting an entire consciousness, it might be better to imagine a standardized multidimensional representation system in which perception, metadata, symbolic meaning and contextual information can be presented together.
There is another important limitation in Egan’s description: a citizen cannot simply transmit everything happening inside its mind. Yatima’s developing mind contains enormously complicated parallel activity, but only portions of it can gain control of the linear or Gestalt output channels at any particular time.
In other words, Gestalt is higher-dimensional communication, not unlimited communication. That actually makes the concept more interesting as something humans might approximate.
Can generative AI communicate in Gestalt?
When I originally asked ChatGPT 4 whether it could communicate like the citizens of Diaspora, it said no. That remains true in the literal sense: humans do not possess Egan’s standardized Gestalt input channel, and a language model cannot send me an arbitrary multidimensional structure and have my brain directly interpret it.
But the answer becomes more interesting if the question is changed slightly. Could an AI system act as an intermediary that helps one human communicate a rich internal model to another human? That looks much more achievable.
The important transformation would be: Human A → AI → shared semantic representation → AI → Human B
Human A communicates what they are trying to express using whatever media are convenient: conversation, text, diagrams, documents, examples, photographs, equations or other material. An AI builds a structured representation of what Human A appears to mean. Another AI process then reconstructs that representation for Human B according to what Human B already understands.
The objective would not be to reproduce the sender’s exact brain state. It would be to get the receiver to approximately the same conceptual state. This suggests that the central problem is not really bandwidth. It is semantic alignment.
Communication as model alignment
When I explain an idea to another person, I am trying to cause something to happen inside their head. I have some internal model, and I convert a small portion of that model into words. The other person hears or reads those words and reconstructs their own model. If the communication works, the two models become sufficiently similar for whatever we are trying to accomplish.
There is already a body of cognitive research describing conversation in somewhat similar terms. The interactive-alignment model of dialogue developed by Martin Pickering and Simon Garrod proposes that people in conversation progressively align linguistic representations and develop increasingly compatible situation models.
Neuroscience provides another intriguing perspective. Research from Uri Hasson’s lab at Princeton has demonstrated measurable coupling between the brain activity of speakers and listeners during successful communication. Greater speaker-listener coupling has been associated with greater understanding.
More recent work from the same group went further. In a 2024 Neuron paper, researchers used contextual embeddings from large language models to model linguistic information as it passed from the brain of a speaker, through spoken language, and into the brain of a listener. They found that these contextual embeddings could model a shared, context-rich linguistic space involved in communication.
This does not mean that a human brain secretly operates like GPT or contains literal embedding vectors. But it supports an important idea: successful communication appears to involve constructing related semantic representations in different minds. Words are the transmission mechanism; the meaning is something reconstructed at either end.
The bottleneck may not be information bandwidth
My original post concentrated heavily on richer interfaces: augmented reality, sound, haptics and eventually brain-computer interfaces. Those things could certainly increase the available communication bandwidth, but bandwidth is probably not the main constraint.
I can download a 500 MB textbook in seconds. That does not mean I can understand the textbook in seconds. Giving someone ten simultaneous streams of information through an augmented-reality headset could make the problem worse rather than better. Human attention, working memory, prior knowledge and conceptual integration remain constraints.
So the objective of a practical Gestalt system should probably not be maximize information transmitted per second. It should be closer to maximize useful conceptual alignment per unit of human attention.
That is a very different problem, and generative AI is unusually well suited to it.
What would a human Gestalt system actually do?
Imagine that I am trying to explain a complicated new idea. Today I might:
- talk about it for twenty minutes;
- draw a diagram;
- send three links;
- give an example;
- introduce an analogy;
- explain where the analogy breaks down;
- clarify several assumptions;
- explain which parts I am uncertain about;
- answer questions;
- discover halfway through that you interpreted an important term differently from me.
A Gestalt-like communication system would try to capture the structure behind all of those activities. The sender could provide rough speech, notes, sketches, source documents, examples and supporting material without carefully assembling everything into a linear presentation.
The AI would construct an intermediate representation containing things such as:
- concepts;
- relationships between concepts;
- causal relationships;
- dependencies;
- assumptions;
- examples and counterexamples;
- evidence;
- confidence and uncertainty;
- areas of disagreement;
- relative importance or salience;
- metaphors and analogies;
- visual or spatial representations;
- the sender’s intended conclusion.
We can think of this intermediate representation as a Gestalt packet. It would not necessarily be something a human reads directly, but a machine-readable representation of the structure the sender is attempting to communicate.
The receiving AI becomes a personalized decoder
The really important part happens at the other end. The receiving AI would ideally know something about the receiver’s existing conceptual model.
Suppose I explain a distributed organizational problem using biological metaphors. If you are a software engineer who understands distributed systems but knows little biology, your AI might determine that my biological analogy is unnecessary.
Instead it might explain the same structure in terms of:
- cache coherence;
- distributed state;
- synchronization;
- consensus;
- propagation delays.
Other receivers might instead get:
- an organizational diagram;
- a worked example;
- a story;
- only a compact diagram and three sentences, if they already understand most of the underlying concepts.
The underlying semantic representation is similar; the reconstruction is different.
This is where generative AI potentially changes communication. Traditionally, the sender has to perform most of this translation manually. A good teacher does it, a good writer does it, and someone who knows you extremely well may learn to do it. AI could potentially perform parts of that translation dynamically for every participant in a conversation.
The resulting communication pathway becomes my internal model → AI representation → your personalized reconstruction → your internal model, rather than my internal model → my words → your interpretation.
That is a much more plausible human approximation of Gestalt communication than attempting to send thoughts directly between brains.
A lo-fi Gestalt system could be built with current technology
Most of the individual components already exist.
Multimodal input
Modern generative AI systems can work with combinations of text, speech, images, documents and structured data. The sender would not need to package their thoughts neatly before communicating them.
Semantic modelling
Language models can identify concepts, relationships, assumptions, contradictions, examples, causal structures and analogies. Knowledge graphs and other structured representations could supplement the model rather than requiring everything to remain as prose.
Personalization
An AI that understands the receiver’s vocabulary, expertise, previous conversations and preferred forms of explanation can adapt how information is reconstructed.
Multimodal output
The resulting representation could be rendered as some combination of:
- text;
- diagrams;
- interactive concept maps;
- animation;
- audio;
- simulation;
- equations;
- examples;
- timelines;
- spatial interfaces.
Feedback
Perhaps most importantly, the system could continuously test whether alignment has actually occurred.
Instead of merely asking:
Did you understand?
the AI could:
- identify likely mismatches between the sender’s model and the receiver’s model;
- generate a counterfactual question;
- ask the receiver to predict what happens next;
- show two interpretations and ask which one they inferred;
- detect when two people are using the same word with different meanings.
Communication could become an iterative process of model convergence.
Who is doing something similar?
Nobody appears to be building the complete system described above, but several fields are working on pieces of the problem.
Uri Hasson’s lab — shared linguistic representations between humans
The work from Hasson and collaborators at Princeton is particularly relevant because it investigates what successful communication looks like across two human brains. Their research on speaker-listener neural coupling suggests that successful communication is associated with greater alignment between neural responses.
Their more recent use of large language model embeddings provides something especially interesting for this idea: a computational representation that can help describe linguistic information appearing first in a speaker and subsequently in a listener.
This is not Gestalt communication, but it is research into the phenomenon a Gestalt system would ultimately be trying to improve: the transfer and reconstruction of meaning between minds.
Interlat — AI agents communicating without language
An even more direct analogy has recently appeared in AI research. A paper presented at ACL 2026 introduced Interlat, a system allowing AI agents to communicate using continuous internal latent representations rather than translating their internal states into natural-language tokens first.
The researchers describe ordinary natural language as a potential bottleneck because a rich internal representation must be reduced into a sequence of discrete tokens before another model can process it. Their agents instead transmit information through latent representations. In their experiments, this could both improve performance in some tasks and substantially compress communication.
For AI systems, this starts to look surprisingly close to one part of the Diaspora idea: AI internal representation → shared latent representation → AI internal representation.
Humans cannot consume those latent representations directly, which is why an AI-mediated human system needs another stage: Human → AI → latent/shared semantic representation → AI → Human. The AIs become codecs between different human conceptual systems.
BrainNet — literal brain-to-brain communication
Researchers from the University of Washington and Carnegie Mellon demonstrated another piece of the science-fiction puzzle with BrainNet, published in 2019. BrainNet connected three humans using EEG and transcranial magnetic stimulation. Participants cooperated on a simplified Tetris-style task in which two people communicated decisions to a third person through the brain-to-brain interface.
This qualifies as genuine direct brain-to-brain information transmission, but its information content was extremely limited—closer to transmitting a binary decision than transmitting a concept, memory or worldview. Brain-to-brain interfaces therefore demonstrate that direct neural transmission is physically possible in limited forms, but they are nowhere close to Egan’s Gestalt communication.
More importantly, they may not be required for useful lo-fi Gestalt communication at all.
Semantic communication research
There is also an interesting parallel in telecommunications research. Traditional communication systems attempt to transmit symbols accurately. A growing area called semantic communication instead asks whether a system can preserve the meaning necessary for a task, rather than reproducing every transmitted symbol.
Systems such as DeepSC have explored neural-network-based semantic communications in which successful transmission can be measured partly by semantic similarity rather than simply bit-perfect reconstruction.
Again, this is not human Gestalt communication, but the shift in objective is relevant. Instead of asking Did every symbol arrive?, the question becomes Did the useful meaning survive? That same shift is important for human communication.
Humans already use semantic compression
The concept is not entirely foreign to ordinary human communication. Consider a meme: someone familiar with the “This Is Fine” meme does not need a paragraph explaining the social context represented by the image. The meme acts as an index into a large amount of previously shared knowledge. An inside joke can communicate even more information between two people using only a word or expression.
Technical communities do the same thing:
- a mathematician can write a small equation that represents pages of prose to another mathematician;
- software architects can communicate substantial amounts of system behaviour with a familiar diagram;
- chess players can look at a board position and immediately perceive structures that would take many paragraphs to describe to a non-player.
These are all forms of semantic compression. The transmitted object is small because sender and receiver already possess much of the same decoder.
This suggests another way of describing what AI could contribute: AI could manufacture temporary shared languages between particular people. Instead of requiring two people to spend years developing exactly the same vocabulary, background knowledge and shorthand, AI systems could translate between their existing conceptual structures.
As the two people worked together more frequently, the shared representations could potentially become more compressed. A diagram that initially required a twenty-minute explanation might eventually be sufficient by itself. That starts to look quite Gestalt-like.
What might receiving a Gestalt packet feel like?
Imagine receiving a message that initially appears as something like:
Central idea
Organizations often mistake making information available for creating shared understanding.
Core causal model
- Information availability → individual access
- Individual access ≠ shared interpretation
- Different interpretations → different decisions
- Different decisions → organizational inconsistency
Important assumptions
- Participants have different backgrounds.
- Much relevant knowledge is tacit.
- The organization changes faster than documentation.
- Incentives influence interpretation.
Sender confidence
High confidence in the central mechanism. Moderate confidence in the proposed solution.
Best analogy given your background
Distributed cache coherence.
Potential disagreement
The sender assumes central coordination has diminishing returns. Your previous arguments suggest you may disagree.
Explore
- [causal diagram]
- [example]
- [counterexample]
- [underlying evidence]
- [original source material]
You could zoom into whichever part required attention. Another person receiving exactly the same underlying Gestalt packet might see something completely different.
That is the key: the unit being transmitted is no longer primarily a document. It is a structured model from which an appropriate communication experience can be generated.
How would we know if it works?
One advantage of redefining Gestalt communication as model alignment is that the idea becomes testable. Take a complex system that Person A understands and Person B does not.
Have A communicate it to B using:
- conventional written communication;
- conventional multimedia;
- an AI-mediated Gestalt system.
Then test whether B can:
- explain A’s model;
- predict what A would predict;
- identify important causal relationships;
- distinguish high-confidence claims from speculation;
- answer counterfactual questions;
- recognize where the model does not apply;
- transfer the model to a new problem.
Measure how much time and attention each approach required. The interesting metric becomes something like:
- time-to-model-alignment;
- conceptual alignment per minute of attention.
That would turn Gestalt communication from a science-fiction analogy into an HCI and cognitive-science research problem.
What about augmented reality, haptics and brain-computer interfaces?
These technologies could still matter:
- augmented reality could make complex spatial representations easier to navigate;
- haptics might create additional useful perceptual dimensions;
- brain-computer interfaces could eventually provide completely new input and output channels.
But I no longer think they are prerequisites. The first useful version of this idea could run on an ordinary computer, tablet or phone.
The major innovation would not be a new display device. It would be the semantic protocol and AI mediation layer. Better interfaces would increase its capabilities later.
From higher bandwidth to better understanding
My original post ended with the suggestion that the rapid development of generative AI made lo-fi Gestalt communication likely to arrive soon. I would phrase that differently now: pieces of it have already arrived.
- We have multimodal AI capable of translating between representations.
- We have language models capable of constructing and transforming semantic structures.
- We have evidence that successful human communication involves forms of representational alignment.
- We have AI research investigating direct latent communication between machines.
- We even have primitive experiments in direct brain-to-brain transmission.
What we do not yet have is a system that puts these ideas together around the specific objective of helping two humans align complex mental models faster and more accurately. That seems like the interesting opportunity.
The science-fiction version of Gestalt communication remains extremely remote, but a useful approximation may not require uploading minds, direct neural interfaces or unlimited communication bandwidth. It may require something considerably simpler:
an AI that understands enough about what I mean, understands enough about what you already know, and can construct the shortest experience that causes your model of the idea to become sufficiently similar to mine.
That is less spectacular than telepathy, but it may also be much more useful. And, unlike the version I imagined in 2023, many of the components required to begin experimenting with it already exist.