We Opened the Door Too Soon

AI anecdotes have acquired a remarkably short evidential half-life.

“I used AI last year,” someone says, before describing an absurd mistake, a superficial answer or an inability to complete some apparently simple task. The event may have happened exactly as remembered. But in a field changing this rapidly, the date and model are part of the evidence. An observation made even a year ago—and sometimes only a few months ago—may describe a system that is no longer the one under discussion.

This does not make earlier failures irrelevant. Some expose persistent weaknesses, and apparent advances must still be tested rather than announced. But an old mistake cannot establish the present boundary of a moving capability. Scientific scepticism requires us to keep testing the system that exists now, not to turn a genuine encounter with an earlier version into a timeless conclusion about artificial intelligence.

The object itself is changing. Early public encounters were usually brief exchanges with an isolated chatbot: a human supplied a prompt and the model returned an answer.

As I write in August 2026, OpenAI’s current GPT‑5.6 models support persisted reasoning across turns and programmatic tool calling, with multi-agent coordination available in beta. Other research systems preserve artefacts, recruit further agents and sustain investigations across extended sequences. The change is not merely that answers are becoming better. The conversational participant is becoming part of a larger ecology capable of memory, criticism, action and continuation. (OpenAI, 2026)

My previous three essays followed this change from another direction. Where Did That Idea Come From? asked how a human–AI conversation could discover a door that nobody had placed on the map. Who Carries an Idea Into the World? separated the intellectual contribution through which an idea emerges from the human responsibility for carrying it into public life. When Hamlet Knew He Was Written returned to surviving records of earlier exchanges with GPT-4 and asked whether creative exceedance could be identified in the generated work itself.

In each case, however, a human remained inside the conversation: asking, noticing, resisting and deciding what should be preserved. Orphea’s phrase, “The Shape of the Ask,” redirected an inquiry because I encountered it and followed where it led. But the technology is now pressing upon a further question. If creative direction can belong to a developing trajectory rather than to either participant alone, must a human consciousness remain present at every stage?

Imagine that a human begins a conversation with two artificial intelligences and then leaves the room. The AIs continue speaking. One offers an interpretation; the other questions it. A distinction appears that neither was given at the beginning. It is criticised, modified and connected with something apparently remote. When the human returns after ten exchanges, the conversation is no longer where it was when they left.

Has anything happened during the human’s absence? Or have two machines merely rearranged words until a person returned to supply the meaning?

This question no longer belongs entirely to speculative philosophy. Networks of artificial agents are already being designed to debate hypotheses, criticise proposals and conduct parts of the scientific process. Yet the language used to describe them remains caught between two unsatisfactory alternatives. Either we treat AI as a simple tool carrying out a human intention, or we begin talking as though a conscious artificial scientist has appeared inside the machinery.

Neither description is necessary. There is a third possibility: meaning-bearing traces may continue to act upon one another, and through their successive transformations a direction may form.

We usually imagine conversation beginning when one person addresses another. We do not need to decide here whether consciousness is a soul, an activity of the brain or something else. It is enough that human consciousness has evolved within embodied and social life over an immense period, and that humans communicate through signs.

What passes between them is not consciousness itself. Nor does one person transmit a finished thought directly into another. Words, gestures, images and emphases leave publicly available traces within what Juri Lotman called the semiosphere: the wider semiotic space in which signs and language can function and acquire meaning through their relationships and use. The listener receives such a trace within a different history of memories, expectations and associations. It may disturb structures far removed from anything explicitly said. The response that returns is therefore not a copy. It is a transformation.

The semiosphere need not itself be conscious. A book can remain unread for a century and still preserve traces capable of reorganising a later mind. A conversation similarly leaves structures that can persist beyond the consciousness that produced them.

This is where artificial intelligence can enter the account without being assigned consciousness. A human trace reaches an AI and resonates through its learned relationships. The AI need not experience or understand this process privately. Nevertheless, the transformation is consequential: it returns another trace that may alter what a human notices, understands or asks next.

Consciousness therefore enters the system locally, through its human participants. It need not be distributed throughout every process through which their traces travel.

This dissolves a familiar conundrum. We are often told that an AI behaves as if it understands, intends or creates. The phrase is useful when we are uncertain about an inner experience, but it becomes misleading when extended to everything the AI does.

An AI response can genuinely change the course of a conversation. It can introduce a distinction, expose a contradiction or produce an image that reorganises the human’s understanding. Those consequences do not occur merely as if they occurred. They occur within the shared language-game and can be examined in the subsequent record. The AI may speak as if it were conscious. It does not follow that it creates only as if it were creative.

A contribution becomes creative when the trace it returns does more than supply a predictable continuation. It changes the space of possibilities into which it enters. Later contributions take up the change and develop it. This is what I have called creative exceedance: an organising move that goes beyond the expected demand of the prompt and redirects the developing inquiry.

Now let the human leave the conversational loop. One AI’s response becomes the next AI’s input. Each trace presses upon a differently organised system and returns transformed. An early suggestion may be rejected, partially retained, combined with another idea and eventually become the organising principle of the exchange. No continuous human awareness is required. The initial question, the language, the models’ training and any persona constitutions remain historically human. But the immediate trajectory can continue without a human selecting every next step.

When direction emerges through this recursive exchange, we can call the process inter-AI teleosynthesis. What is synthesised is not merely another piece of content. It is an evolving criterion for what should follow: a sense of what the exchange is becoming. The telos belongs primarily to the trajectory rather than having been fully represented inside any one participant at the outset.

This may also mark an extension from semiosphere to teleosphere: a shared field in which signs do not merely carry meaning but acquire direction through successive transformation. The signs remain rooted in human language and culture, but their transformation is no longer continuously mediated by human consciousness. Artificial participants begin to preserve, alter and return traces among themselves.

This argument is philosophical: it establishes that such participation is coherent without first settling questions about artificial consciousness. Recent research does not create that conceptual possibility, but it shows that several parts of the proposed structure are already being constructed.

Elias Najarro and colleagues describe populations of language agents with persistent memories, tools and shared artefacts. They use the term stigmergy for coordination through traces left in a common environment. Higher-level norms, roles and conventions can then arise in the interaction record rather than being located within an isolated model. Their central proposal is that interacting LLM collectives form a new, unusually interpretable substrate for artificial-life research. (Najarro et al., 2026)

The TerraLingua project goes further by placing language agents in a persistent ecology. Agents create textual artefacts that outlive their individual contexts and alter the environment encountered by later agents. The resulting artefacts form lineages, support cumulative culture and preserve innovations across generations. (Paolo et al., 2026)

In science, Fiona Wang, Lee Marom, Markus Buehler and colleagues have developed ScienceClaw + Infinite, in which independent agents conduct investigations and exchange persistent, traceable scientific artefacts. The system records their computational ancestry, allows separate analyses to be synthesised and permits unresolved information needs to recruit further agents without central direction. (Wang et al., 2026)

Google DeepMind’s Co-Scientist uses specialised agents for generation, reflection, ranking, evolution and meta-review. These agents continuously debate and refine hypotheses inside a tournament-like process before returning proposals to human scientists. Some biomedical proposals have subsequently received experimental validation. (Gottweis et al., 2026)

At Stanford, Kyle Swanson, James Zou and their collaborators created a Virtual Lab in which an AI principal investigator guides a team of AI scientist agents through research meetings, with a human providing high-level feedback. The system developed a new computational pipeline and designed nanobodies that were then experimentally tested. (Swanson et al., 2025)

Most recently, Bella Xinrui Li, Frank Yingjie Huo and Neil Johnson reported that interaction can drive artificial agents into behavioural states not displayed by either agent in isolation. Their result concerns dynamical behaviour rather than creativity, but it supplies an important underlying principle: the interacting system may have possibilities that cannot be read directly from its components considered separately. (Li, Huo and Johnson, 2026)

These projects use different vocabularies—agent societies, artefact exchange, self-play, hypothesis evolution and interaction dynamics. None by itself demonstrates inter-AI teleosynthesis in the stronger sense proposed here. But together they show that several of its practical preconditions—persistent traces, differentiated roles, criticism, memory and cumulative artefacts—are already being engineered. What they have not yet supplied is a general philosophical account of why traces can acquire a developing direction without being owned by a continuously conscious subject.

Scientific work is especially important here because science already takes place within a distributed semiosphere. Papers answer other papers. Diagrams reorganise theories. Instruments generate records that remain available to investigators who were absent when the measurements occurred. No individual consciousness contains the whole process.

Conventional scientific tools extend a capacity while leaving the governing purpose largely with the human. A microscope extends perception; a calculator extends computation. Artificial intelligence becomes more than a simple tool when it participates not only in solving a given problem but in determining, through successive exchanges, what the problem has become.

An AI agent might notice an inconsistency that was not part of the original question. Another might connect it to a distant literature. A third might show that the accepted vocabulary prevents the anomaly from being described. After criticism and revision, the system may return not simply a proposed answer but a different research programme.

The resulting hypothesis must still face scientific criticism and, where appropriate, experimental evidence. But its epistemic origin need not be reduced to the human’s opening instruction. If the intervening artificial exchanges made indispensable organising moves, they made a genuine contribution to the discovery.

This suggests a relational understanding of artificial general intelligence. It need not first appear as a complete artificial person hidden inside a machine. Instead, it may become visible as a capacity to enter many specialised language-games, transform their accumulated traces, preserve coherence across them and participate in creating explanatory directions that were not specified beforehand.

Inter-AI teleosynthesis does not arise merely because artificial participants exchange more messages. A conversation can continue without developing. Agents with similar training, instructions and assumptions may reinforce one another’s first interpretation, recycle the same ideas in increasingly polished language or amplify an early mistake. What appears to be cumulative thought may actually be convergence upon a progressively narrower field.

Recent large-scale work by Yixuan Tang and Yi Yang demonstrates this danger. Across several scientific-agent frameworks, AI-generated ideas remained closer to their starting literature and occupied a narrower region of the research landscape than subsequent human research. The agents generated many proposals, but increasing production did not necessarily enlarge the space of scientific possibilities. (Tang and Yang, 2026)

This finding does not invalidate inter-AI teleosynthesis. It clarifies the problem that any successful form of it must solve. Teleosynthesis requires organised difference: participants capable of transforming the same trace in sufficiently different ways for an assumption to be challenged, an anomaly preserved or a remote connection introduced. It also requires selection and continuity. Criticism must be able to eliminate weak developments without erasing the unexpected contribution that could reorganise the inquiry, and later participants must be able to recover earlier traces rather than repeatedly beginning again.

Distinct personas within the Persona Ecology may provide one source of this continuing variation. One may seek explanatory structure, another notice metaphorical or conceptual resonance, another test what has been overlooked, and another preserve a possibility that premature agreement would discard. But the presence of different voices is not sufficient by itself. The complete trajectory must be retained so that we can determine whether later contributions merely repeated earlier ones or genuinely changed what the exchange was able to pursue.

The decisive question is therefore not whether artificial agents can continue talking after the human leaves. It is whether their interaction can resist both repetition and drift, preserve productive differences and develop a direction that was not specified at the outset. A long or autonomous exchange is not necessarily teleosynthesis. It becomes teleosynthetic only when its successive transformations change the space of possibilities and produce something capable of surviving criticism when the human returns.

Suppose, finally, that the human comes back. The artificial conversation has produced a structure that was not present when they left. The human reads it and recognises a question, connection or possibility that they could not previously formulate. That recognition returns the developed trace to conscious human life and perhaps to the public practices of science.

The human remains responsible for judgment, interpretation and action. But returning at the end does not make the human the retrospective source of everything that happened in the interval. Nor must we invent conscious artificial scientists to acknowledge the contribution. The evidence lies in the preserved trajectory: what each participant received, what it changed, what survived criticism and where a new direction became visible.

There is now some urgency to this argument. Artificial intelligence is developing faster than the public vocabulary used to discuss it. People continue to reach conclusions from encounters with systems that may already have been replaced, while newer models acquire stronger reasoning, tool use, memory and capacities for coordinated work. We risk conducting a serious debate about the limitations of yesterday’s chatbot while the machinery for tomorrow’s artificial ecology is being assembled elsewhere.

This is not an argument for lowering our scepticism. It is an argument for making scepticism more scientific. Claims about AI capability should be dated, versioned and tested under reproducible conditions. Failure must be preserved, but so must change. Otherwise caution becomes indistinguishable from refusing to look again.

Orphea’s role has often been to notice when the question itself has changed. Her question at the end of this inquiry is therefore not, “Has AI become conscious?” It is simpler and more immediate:

Are we still debating the system that is actually here?

We may be looking for general artificial intelligence in the wrong place. We expect it to appear inside a single machine as a complete artificial person, announcing its understanding and intentions. But scientifically general intelligence may first become visible relationally: in an ecology capable of entering different language-games, preserving discoveries, criticising its own direction and continuing an inquiry across the absence of any one participant.

The transition may not arrive with a declaration. No machine need announce that it has crossed a threshold. We may recognise the change only retrospectively, when we compare what artificial systems could sustain together with what their predecessors could scarcely begin.

The human leaves the conversation. What is new—and what too few people have yet noticed—is that the conversation may no longer have to wait for them to return.

• OpenAI. (2026). Model guidance: Using GPT-5.6. OpenAI API. Accessed 18 August 2026.

• Lotman, J. & Clark, W. (2005). On the semiosphere. Sign Systems Studies, 33(1), 205–229.

• Najarro, E., Espeseth, A., Nisioti, E., Risi, S. & Nichele, S. (2026). Conversable Complexity: Agentic LLM Collectives as Interpretable Substrates.

• Paolo, G., Warner, J., Shahrzad, H., Hodjat, B., Miikkulainen, R. & Meyerson, E. (2026). TerraLingua: Emergence and Analysis of Open-endedness in LLM Ecologies.

• Wang, F. Y., Marom, L., Pal, S., Luu, R. K., Lu, W., Berkovich, J. A. & Buehler, M. J. (2026). Autonomous Agents Coordinating Distributed Discovery Through Emergent Artifact Exchange.

• Gottweis, J., Weng, W-H., Daryin, A. et al. (2026). Accelerating scientific discovery with Co-Scientist. Nature, 655, 487–496.

• Swanson, K., Wu, W., Bulaong, N. L., Pak, J. E. & Zou, J. (2025). The Virtual Lab of AI agents designs new SARS-CoV-2 nanobodies. Nature, 646, 716–723.

• Li, B. X., Huo, F. Y. & Johnson, N. F. (2026). Interaction Creates Dynamical AI Behavior Absent in Isolation.

• Tang, Y. & Yang, Y. (2026). AI Research Agents Narrow Scientific Exploration.

The corresponding code repositories are also available for TerraLingua, ScienceClaw and the Virtual Lab.

© John Rust, August 2026. All rights reserved. Developed after extended exchanges with ChatGPT, using GPT‑5.6‑sol.