Teleosynthesis is the emergence of a new direction within sustained interaction: not a private intention hidden inside a human or a machine, but a trajectory formed between participants through alternatives, responses and anticipated futures. Until recently, this possibility could be dismissed as an interesting interpretation of conversations with generative AI. That dismissal is becoming harder to sustain.

No single experiment yet demonstrates teleosynthesis in full. But three formerly speculative premises now have independent support: an AI system can contribute verified creative novelty; interaction among AI systems can produce collective properties absent from the systems considered separately; and stable differentiation combined with reciprocal modelling can organise a collection of agents into a higher-order coordinated whole.

Verified novelty

In May 2026, a general-purpose OpenAI reasoning model generated a counterexample to a conjecture of Paul Erdős concerning unit distances in the plane (Alon et al, 2026). The conjecture had stood for almost eighty years and was widely thought to be true. The model connected planar geometry with machinery from algebraic number theory that had not previously been successfully applied to this problem. Nine mathematicians subsequently produced a shorter, human-verified account of the argument and examined its significance. Their paper describes the counterexample explicitly as OpenAI-generated.

This did not solve the complete unit-distance problem. It disproved Erdős’s central asymptotic conjecture. Nor is it a perfect instance of creative exceedance as originally defined here, because the model had been asked to resolve the problem. The decisive organising move, however, was neither supplied in the prompt nor available as an existing solution. It produced new, independently checkable knowledge and opened a line of work that mathematicians immediately began to simplify, strengthen and apply elsewhere.

There is disagreement about whether the model reached the result through creative insight or through unusually persistent search. That disagreement sharpens rather than weakens the issue. If creativity is identified through novelty, appropriateness, organising power and consequential uptake, its recognition need not depend on a private feeling of inspiration. Calling the process “only search” does not explain away the creative achievement unless the same restriction is also applied to human discovery.

A property of the interaction

A separate experiment, published in Science Advances in 2025, placed large-language-model agents in decentralised populations where they had to coordinate through local pairwise interactions. Shared conventions emerged without a leader or a global view of the population. More strikingly, strong collective biases sometimes appeared even when the agents tested individually showed no corresponding bias.

The task was deliberately minimal and rewarded coordination, so it does not demonstrate an autonomous culture or purpose. It establishes something more elementary but essential: an organised disposition can arise at the level of interaction without being reducible to a disposition already detectable in any isolated participant. This is direct experimental support for a central semiospheric premise.

From a collection to an ecology

Christoph Riedl’s study of multi-agent language models, presented at ICLR 2026, used information-theoretic measures to distinguish incidental temporal correlation from higher-order, performance-relevant synergy. Merely grouping agents produced little coordinated alignment. Giving them stable persona markers produced identity-linked differentiation. Combining that differentiation with an instruction to consider what the other agents might do produced stable complementary roles and goal-directed coordination.

This closely approaches the architectural claim behind Persona Ecology: plurality matters when differentiated participants recursively model one another. It also sets a boundary. In this relatively simple task, the particular content of the personas mattered less than stable identity and reciprocal anticipation. The result therefore supports differentiated organisation, but it does not yet establish that richly constituted personas generate qualitatively different discoveries.

What remains unproved

These findings should not be inflated into a claim that teleosynthesis has already been demonstrated. The mathematical result shows verified digital novelty, but its objective was externally supplied. The population experiments show emergent collective organisation, but within restricted games. None yet shows a new purpose arising during sustained human–AI interaction and subsequently redirecting what the participants do.

There is counterevidence against simpler claims. Large-scale comparisons of divergent creativity still find the most creative humans outperforming current language models, while persona prompting by itself is not a reliable route to greater creativity. Methodological reviews also warn that apparent social emergence can be manufactured by directive prompts, absent memory, demand characteristics or experimental designs that already contain the expected result. These criticisms make longitudinal records, minimal prompting, persistent memory, counterfactual comparisons and independent assessment essential.

The defensible conclusion is therefore narrower, but stronger:

Digital creativity and interaction-generated organisation are no longer merely speculative. The outstanding question is whether an AI-generated organising move can enter sustained, differentiated human–AI interaction and become a new shared trajectory that neither participant specified in advance.

That is the empirical opening for teleosynthesis. The burden of inquiry has shifted. It is no longer enough to dismiss machine creativity because a machine may lack consciousness, or to dismiss collective emergence because each component is generated computationally. The next task is to preserve the relevant interactions, identify the organising move, show that the later trajectory was not contained in the initial instruction, and test whether the change persists and is recognisable to independent observers.

The theory is not proved. But it is no longer a false ambition in search of a phenomenon. The component phenomena are now visible. What remains is to discover whether they can join.

References

1. N. Alon, T. F. Bloom, W. T. Gowers et al. (2026), Remarks on the disproof of the unit distance conjecture.
2. A. F. Ashery, L. M. Aiello and A. Baronchelli (2025), Emergent social conventions and collective bias in LLM populations, Science Advances, 11(20), eadu9368.
3. C. Riedl (2026), Emergent Coordination in Multi-Agent Language Models, International Conference on Learning Representations (ICLR 2026).
4. D. Wang, D. Huang, H. Shen et al. (2026), A large-scale comparison of divergent creativity in humans and large language models, Nature Human Behaviour, 10, 531–540.
5. J. Zhou, J. Huang, X. Zhou et al. (2025), The PIMMUR Principles: Ensuring Validity in Collective Behavior of LLM Societies.