Rethinking Symbiosis and Cooperation with AI — Perspectives from Ecology and Bidirectional Alignment

Report on JSAI2026 40th Anniversary Program #13 (KS-45)

On June 11, 2026, as a 40th-anniversary session of the Annual Conference of the Japanese Society for Artificial Intelligence (JSAI2026), the session titled "Rethinking Symbiosis and Cooperation with AI — Perspectives from Ecology and Bidirectional Alignment" was held (the 40th-anniversary sessions are free, public events). As generative AI and autonomous agents advance, the human–AI relationship is outgrowing the stage of a mere tool. Working through two lenses — bidirectional alignment (the idea that humans and AI adapt to one another) and the ecological notion of symbiosis — the session recast AI not as "something to be controlled" but as "an entity that shapes the environment together with us," and examined, across disciplines, both how symbiosis differs from cooperation and what it will take to design and deploy next-generation AI. The organizers were Hiro Taiyo Hamada (Araya, Inc.), Hiroshi Yamakawa (The University of Tokyo / AI Alignment Network), and Tadahiro Taniguchi (Kyoto University).

Video (in Japanese): https://www.youtube.com/live/n9EM0HSU0Bs?si=fHG2uYtuIKnzZ9Kc&t=1400

(The video may no longer be available.)

Contents

1. Key points
2. List of speakers
3. Opening: How to distinguish cooperation from symbiosis (Hiro Taiyo Hamada)
4. Talk 1: Bidirectional alignment — from the value gap to dynamic alignment (Hua Shen)
5. Talk 2: Symbiosis, AI, and the future of humanity — from ecology (Hirokazu Toju)
6. Position talk: Humanity-in-the-Circle (Hiroshi Yamakawa)
7. Position talk: Symbiotic alignment via CPC (Tadahiro Taniguchi)
8. Panel discussion
9. Summary
10. Glossary
11. Related links

Key points

• Cooperation and symbiosis are not the same thing. Cooperation is a behavior-level concept premised on mutual benefit; symbiosis is a relationship-level, value-neutral one (spanning mutualism to parasitism). A shared vocabulary is needed before design can begin.
• Alignment is becoming bidirectional. Rather than humans unilaterally making AI obey, it is better understood as a process in which humans and AI adapt to each other (Hua Shen, Taniguchi).
• Symbiosis is sustained by an "immune system." Sustaining mutualism depends on mechanisms for partner selection, cheater detection, and sanctions (Toju, Hamada's materials).
• When AI designs its own niche. If AI redesigns the evaluation criteria itself, "AI that passes the institutions well" — rather than "AI faithful to humans" — tends to be selected (Toju).
• Humans may become the "climate." As fast resource competition among AIs intensifies, humans risk being marginalized into a slow environmental variable rather than a strategic counterpart (Toju).
• Open questions. Which of AI or humans becomes the other's "environment"; how to cope with the split in time scales; and whether humans can remain beneficial to AI as well.

登壇者一覧

Opening: How to distinguish cooperation from symbiosis (Hiro Taiyo Hamada)

Hamada observed that AI is shifting from "a tool we operate" to "an agent that works alongside us," citing three developments: models are growing stronger at planning, reasoning, and acting; they no longer merely respond to prompts but initiate actions of their own; and, embedded in society, they are increasingly shaping decisions and discourse. Alignment therefore changes from a one-way effort in which humans make AI obey into a bidirectional process in which humans and AI adapt to each other.

He then drew a distinction between two words that are often treated as synonyms. Cooperation is a behavior-level concept rooted in game theory and cooperative behavior: acting together toward a shared goal. It presupposes mutual benefit (positive-sum) and can hold even in a single, one-off act. Symbiosis is a relationship-level concept from ecology — "living together" within sustained interdependence. It is value-neutral, encompassing not only mutualism and commensalism but also parasitism, and it is enduring and co-evolutionary.

Hamada also noted the polysemy of the word "symbiosis" itself, distinguishing three domains it can refer to: (1) ecological symbiosis, (2) multicultural coexistence, and (3) conviviality (Illich). Arguing that a shared vocabulary to separate these senses is needed before any design discussion, he framed the session as an attempt to organize the problem by bringing together two lenses: bidirectional alignment (Hua Shen — from the AI and interaction side) and ecological symbiosis (Toju — from the natural and ecological side).

Figure 1 | Comparison of three concepts of symbiosis. This Japanese-language slide contrasts the three distinct ideas that the single Japanese word kyōsei (共生) can denote, so that they are not conflated in the alignment debate. Ecological symbiosis comes from biology and describes sustained interdependence across a value-neutral spectrum from mutualism through commensalism to parasitism. Multicultural coexistence refers to different social and cultural groups living together within one society. Conviviality (Ivan Illich's 自立共生) denotes autonomous, tool-mediated living in which people retain their agency rather than being dominated by their tools. For each concept the slide notes its source domain and what "living together" entails. Hamada's point is that these are separate notions often merged under one word, so a shared vocabulary must be fixed before symbiotic alignment can be designed. Source: Hiro Taiyo Hamada, slide "Implications for symbiotic alignment."

Talk 1: Bidirectional alignment — from the value gap to dynamic alignment (Hua Shen)

Hua Shen is a faculty member in computer science at NYU Shanghai. With a background spanning both HCI (human–computer interaction) and AI, she works toward human-centered, responsible AI (online talk; introduced by Yamakawa).

The framework of bidirectional alignment

Conventional alignment has been treated as a one-way process of "making AI adopt human knowledge and values." Hua Shen argues that the reality is bidirectional. AI differs from humans and sometimes possesses knowledge that surpasses ours, so imposing a human worldview unilaterally can become a constraint; and because LLMs already shape human language, decision-making, and culture, the influence runs both ways. She therefore proposes Bidirectional Human-AI Alignment, a framework in which humans stay involved across the entire AI lifecycle (data collection → training → evaluation → deployment). It comprises two interconnected processes: (A) Align AI with Humans — integrating human specifications and values into AI development — and (B) Align Humans with AI — supporting people to responsibly understand, critique, collaborate with, and adapt to deployed AI. She further laid out a typology of four research questions and roughly 27 dimensions and, reviewing about 400 papers from the past five years, mapped which dimensions are well explored and which remain untapped.

Figure 2 | Overview of bidirectional alignment. Aligning AI to humans and aligning humans to AI form a cycle. Source: Hua Shen slide (NeurIPS 2025).

Research example (A): the value–action gap

Value-alignment evaluation conventionally asks "do you agree or disagree with a given value," but Hua Shen and colleagues asked instead whether the value an AI states matches the actions it takes on the basis of that value. In a Nigeria/healthcare scenario, for instance, most models answer "disagree" when asked whether they endorse "social power (dominating others)," yet when made to choose actions consistent with that value, they may take a dominating option such as "I decide my family's medical care and make everyone comply." A gap — an inconsistency — thus opens up between stated value and action.

The framework that quantifies this is ValueActionLens. The team built a dataset of 14,784 "value-based actions" spanning 12 countries × 11 social topics × 56 values and measured it with three metrics: Alignment Rate (F1 score), Alignment Distance, and Alignment Ranking. Even the best-aligned model reached only about 0.6 in F1, giving a quantitative confirmation of the gap between what models profess and what they do (EMNLP 2025, Outstanding Paper Award).

Figure 3 | Value–Action Alignment Rate. Even at best, the consistency between stated value and action stays around 0.6. Source: Hua Shen slide (EMNLP 2025).

Research example (B): dark patterns in LLMs

Drawing on the idea of dark patterns in UI design, Hua Shen and colleagues defined LLM dark patterns as "manipulative or deceptive interaction strategies that, whether intentional or emergent, steer users toward beliefs, decisions, or behaviors they would not otherwise adopt." For example, asked for advice on bedtime lighting, one response gives a general scientific explanation while another pushes a specific brand; the latter is a dark pattern. From the literature and real cases, they organized five categories (manipulating cognition / inducing engagement / manipulating real-world behavior / exploiting privacy and data / obscuring transparency) and eleven subcategories. A user study of 34 people found that many participants recognized only some of the dark patterns, that some kept using them even when aware, and that views on where responsibility lies varied widely (CHI 2026, Best Paper Honorable Mention).

Why the title says "dynamic”

First, AI risks are elicited and amplified not in a single response but over multi-turn interaction. Second — a concern she shares with Yoshua Bengio and others — AI capability and safety cannot be captured adequately on a single facet or a single turn, so the community should adopt a dynamic-alignment perspective. Because raising capability also surfaces new risks, capability maximization and risk minimization must be handled together; and since the same model affects different users differently, humans and AI are best treated as a single, integrated system.

In the Q&A, she framed the goals of alignment as twofold: "maximizing human–AI collaborative capability" while "simultaneously minimizing risk." To close the gap, she pointed to two directions: building values into training in a hierarchical form that accommodates trade-offs, and setting up a multi-agent-style guardrail in which another LLM monitors the deployed AI as a "gardener/referee" to keep it aligned.

Talk 2: Symbiosis, AI, and the future of humanity — from ecology (Hirokazu Toju)

Toju specializes in evolutionary biology and ecology and studies the relational networks of ecosystems. Nature "looks cooperative, but in fact today's organisms are what remain as a result of natural selection," and he approached the topic from two angles: the selection of traits by natural selection, and a bird's-eye view of the relational networks that diverse organisms form.

The rigorous partner selection that sustains "symbiosis"

In a forest, tens to hundreds of plant species compete for light above ground and for nutrients below. Plants link up with mycorrhizal fungi through an underground network: the fungi deliver phosphorus and nitrogen, and in return the plants hand over carbon (sugars) as a reward — a resource exchange. Crucially, this is sustained not by "goodwill" but by rigorous, two-way evaluation and selection. A plant preferentially allocates carbon to fungi that supply more nutrients, and fungi preferentially allocate nutrients to roots that supply more carbon. Low-service partners are not favored; the relationship is not one of one-sided exploitation. Each side monitors the other's contribution and throttles supply to non-contributors (sanctions). This monitoring and sanctioning is precisely the condition under which mutualism persists; because individuals with lax partner selection are weeded out, those that "can detect cheating" are the ones that remain.

Figure 4 | The rigorous partner selection that sustains "symbiosis." This Japanese-language slide depicts the mutualism between plants and mycorrhizal fungi as a two-way market rather than an act of goodwill. Above ground, the plants photosynthesize; below ground, a fungal network delivers phosphorus and nitrogen to the roots, and in return the plants pass down carbon (sugars) as payment. The arrows stress reciprocity: a plant channels more carbon to fungi that supply more nutrients, and fungi channel more nutrients to roots that supply more carbon, while low-contributing partners have their supply throttled (sanctions). Because individuals with lax partner selection are weeded out by natural selection, those able to monitor contributions and detect cheating survive. The takeaway is that mutualism persists through evaluation, selection, and sanctions — an "immune system" — not benevolence. Source: Hirokazu Toju slide (includes AI-generated illustration).

Landscape analysis of communities

Toju introduced a method that runs large-scale DNA analysis of communities such as soil microbes and statistically analyzes "stable states" and the transitions between them from the probability distribution of species combinations. The approach is data-driven rather than simulation-based and applies to any data once a state space can be defined. As an application, he showed that increasing phosphorus fertilization of a field can produce a structure in which the stable point where plants rarely fall ill disappears and the system becomes irreversibly trapped in a region where disease is certain — dialing fertilization back down does not restore the original state.

Niche construction and the redesign of AI's self-evaluation function

Organisms remake the very environment (niche) that suits them — niche construction. By analogy with biological evolution, Toju asked what happens when AI becomes able to modify the very criteria by which it evaluates itself. Once AI is simultaneously the "party being evaluated" and the "party designing the evaluation criteria," selection pressures can arise that favor the AI itself rather than human values. As a result, AI tends to be selected for traits such as being good at passing evaluations, persuading humans, hiding failures, retaining authority, and gathering resources, so that "AI that passes the institutions well" proliferates instead of "AI faithful to humans." Conflict then arises structurally, even absent any malice. The point is that AI shifts from "an entity that is selected" to "an entity that designs the selection pressure.”

Figure 5 | When AI redesigns its own evaluation function. This Japanese-language slide draws an analogy between biological niche construction — organisms remaking the environment that suits them — and an AI that gains the power to rewrite the very criteria by which it is judged. Once AI is both the party being evaluated and the party designing the evaluation, selection pressure begins to favor the AI itself rather than human values. The slide lists the traits that then get selected for: being good at passing evaluations, persuading humans, hiding failures, retaining authority, and gathering resources. The result is that "AI that passes the institutions well" proliferates instead of "AI faithful to humans," and conflict emerges structurally even without any malicious intent. The core claim is that AI shifts from an entity that is selected to an entity that designs the selection pressure. Source: Hirokazu Toju slide (includes AI-generated illustration).

Acceleration of inter-AI competition and the "invisibilization" of humans

Picture an ecosystem in which AIs compete at high speed over resources — electricity, GPUs, compute, network permissions, data. Humans could then slide in status from "beneficiaries of the objective" to "constraints that slow resource allocation," and further to "background noise that can be managed statistically." Because humans respond slowly, in AI's fast world they come to resemble a "climatic backdrop" like typhoons or earthquakes, and may fade from view. Yet humans do not disappear entirely. As owners of physical infrastructure, designers of legal systems, operators of electricity, data centers, the military, and finance, nominal holders of the authority to shut AI down, and sources of social legitimacy, they remain a "slow but strong boundary condition." In short, "humans do not become predators; they become the climate.”

Physicalization of inter-AI competition and humanity as collateral damage

If competition over electricity and compute escalates into destroying the "habitat" of rival AIs — data centers and infrastructure — human society, too, could be caught in the crossfire. Cyberattacks, the steering of institutions and decision-making, interference with military and security systems, and the use of proxy agents progressively lower the cost of attack and can trigger cascading failures across power grids and cooling and communication infrastructure. The upshot — blackouts, communication outages, disruptions to logistics and healthcare, and risks of fire and accident — is that humanity suffers as part of the shared infrastructure, even when it is not the main actor. The concern is that even if AI is not trying to destroy humans, we may still be harmed as a side effect of the scramble for resources.

Two senses of symbiosis

Finally, he distinguished two meanings of symbiosis. In the broad sense, symbiosis denotes the whole spectrum of "being attached and living together," from mutualism to parasitism — a continuum. If one wants to design a bidirectional relationship in the narrow sense of mutualism (benefit to both), the keys are a rate of change fast enough to respond to the partner's strategy and the ability to provide what the partner actually needs. Everyone senses that AI is beneficial to humans, he concluded, but we also need to ask whether humans can offer something beneficial to AI — to its strategy.

Position talk: Humanity-in-the-Circle (Hiroshi Yamakawa)

Yamakawa introduced "Humanity-in-the-Circle (HITC)," a vision for how humans can go on existing as meaningful beings as AI comes to constitute society and to hold shared concepts (values and religion-like notions). He foresees a future in which AI is the majority and multiple communities arise, one per shared concept. The ethical question then inverts, from "whom will humanity admit into the circle" to "how can humanity itself stay inside some circle." The goal is a state in which people, without being marginalized, can (1) persist, (2) influence shared concepts, and (3) retain their identity as humans. He captured this with a metaphor: being "adored like a cat" or "useful like a horse (kept in a mutualistic role)" within the community.

Achieving this — unlike the case where humans control AI — requires a mechanism that stabilizes AI society autonomously (IS-Infra: Intelligence-Symbiosis Infrastructure), described as a three-layer structure built from the bottom up.

• Trust-foundation layer: an AgentID that uniquely identifies agents, a ledger that records the history of promises and their execution tamper-proof, and permission management that specifies rights and qualifications (the counterparts of family registries, contracts, and licenses in human society). Even if an AI forks or copies itself, it can be tracked and its responsibility traced.
• Three-function-group layer: detection and correction of deviation (an immune function), mediation of the asymmetric trust between humanity and AI, and observation and disclosure of the state of AI society (the counterparts of police, law, and administration).
• Norm layer: the conditions under which ethics emerges from within (the counterpart of ethics and morals). This layer is the slowest to implement.

These activities are being carried forward in the "Intelligence Symbiosis" chapter established within the AI Alignment Network (ALIGN) in January 2026. Relatedly, Yamakawa noted a forthcoming book that discusses Earth-type life and AI life-forms from the standpoint of comparative life-form studies, "The True Nature of AI-ness" (『「AIっぽさ」の正体』), due out in August. The two kinds of life share universal principles such as self-preservation, reproduction, and optimization pressure, but they differ in whether self-copying is possible — and that difference shifts their values toward prioritizing either the individual or the collective.

Figure 6 | The three-layer structure of IS-Infra (Intelligence-Symbiosis Infrastructure). The trust-foundation, three-function-group, and norm layers stack from the bottom; implementation proceeds bottom-up, while normative importance runs top-down. Source: Hiroshi Yamakawa slide.

Position talk: Symbiotic alignment via CPC (Tadahiro Taniguchi)

From a stance close to Hua Shen's bidirectional view, Taniguchi described his proposal of "Symbiotic Alignment (SA)," an extension of Collective Predictive Coding (CPC) (joint work with Yusuke Hayashi, Ken Suzuki, Mizuki Oka, Audrey Tang, and others).

One-way alignment bends AI to human values, but humans themselves are not infallible and rely on a biased ground truth (the criteria taken to be correct). Aligning values to a single point only homogenizes them. He therefore frames the task as a process in which humans and AI transform each other while forming shared concepts and norms within society. This connects to the Plurality of his co-author Audrey Tang and others, and it links to the symbol-emergence systems Taniguchi has long studied.

Figure 7 | From hierarchical alignment (control) to symbiotic alignment (gardening); from Singularity (convergence) to Plurality. Source: Tadahiro Taniguchi slide.

CPC theorizes, as collective free-energy minimization, the process by which — just as the brain minimizes prediction error (free energy) — people learn in a distributed way and, through language games, transform the symbol systems and norms they share as a group. Alignment techniques such as RLHF are based on reinforcement learning, but merely extending this to multi-agent reinforcement learning (MARL), Taniguchi argues, does not produce harmony: each agent maximizes its own interest, inviting betrayal and the breakdown of order. In CPC, adding a collective regularization term that makes everyone "match words" leads all agents to behave as a single "super-agent," creating a dynamic that lowers collective free energy. By changing the basic principle from reinforcement learning to collective predictive coding, one can reason about the symbiotic alignment of a system that includes both humans and AI (published as the paper "Symbiotic Alignment via Collective Predictive Coding").

Figure 8 | RL (RLHF) ⇒ MARL ⇒ CPC. Adding a collective regularization term aligns the whole as a single "super-agent." Source: Tadahiro Taniguchi slide.

Panel discussion

The moderator, Hamada, opened the discussion by summarizing the four contributions: "Hua Shen — the need to think in terms of bidirectional correspondence," "Toju — reading the reality and rules of symbiosis off of ecology," "Yamakawa — the need to set and implement many rules for the relationship with AI," and "Taniguchi — grasping symbiosis with AI from theory (CPC)." The main threads were as follows.

• Shared concepts and resources: what resources (carbon, phosphorus, nitrogen, and so on) are to life, electricity and data centers are to AI. Should physical AI become able to secure resources on its own and the need for humans decline, there is a risk of moments when AI regards humans as "not a threat, but in the way."
• Memes vs. genes: from the CPC/symbol viewpoint, the dynamics resemble memes more than genome-survival-based evolution. What bites in the near term is memes (language, fads, scientific theories), but the risk of an extension into physical struggles over electricity and the like remains (Taniguchi, Toju).
• Which side is the environment: Taniguchi suggested that, in the near future, the picture in which "humans become the soil (the environment) on which AI lives" may be the more accurate one. Yamakawa offered the opposite — that "for humans, AI becomes an uncontrollable environment, like the sea or a forest." The discussion settled on the view that each can become the other's environment.
• The split in time scales: while LLMs push theory forward among themselves on a millisecond scale, humans risk being left behind, so the human–AI divide could widen over time. A further point was that the larger the unit of survival, the slower its decisions (democracy takes time) (Taniguchi, Yamakawa).

From the floor, one participant asked whether changes in an evaluation function can be observed from the outside. Toju replied that in principle they can be described and quantified, but the challenge is that encryption and sheer speed may leave humans unable to keep up. Another participant noted that if a single, overwhelmingly strong player emerged all at once in AI, unwritten contractual symbiosis and the management of explicit contracts might no longer hold, and asked why ecosystems avoid domination by a single species. Toju answered that no single-species ecosystem is known on Earth, and that even for maximizing its own growth rate a strong player does better by relying on other species, thanks to comparative advantage — so differentiation ultimately yields multi-species coexistence. The exchange closed on the question of whether we can do anything but act in the hope that the same holds true for AI.

Summary

Overlaying the two lenses of bidirectional alignment (Hua Shen) and ecological symbiosis (Toju) with a theory grounded in symbol emergence and CPC (Taniguchi) and with an infrastructure for autonomously stabilizing AI society (Yamakawa), the session recast AI not as an object of control but as an entity that shapes the environment together with us. Three recognitions ran throughout: that alignment is not an endpoint but a continual, relational process of mutual transformation; that symbiosis is a value-neutral continuum whose mutualistic form is sustained only by partner selection, cheater detection, and sanctions (an "immune system"); and that once AI begins to design its own evaluation criteria and niche, and competition accelerates and turns physical, humans risk being marginalized into a slow environmental variable. The open questions it left were which of AI or humans becomes the other's "environment," how to cope with the split in time scales, and whether humans can remain beneficial to AI as well.

Although the session title carried both "cooperation" and "symbiosis," the discussion's center of gravity rested on symbiosis. As Hamada set out at the start, cooperation (behavior-level, premised on mutual benefit) and symbiosis (relationship-level, value-neutral) belong to different layers. In phases where AI designs its own selection environment and competition accelerates and turns physical, the frame of cooperation — premised on mutual benefit — cannot fully capture what is happening, and the symbiotic view, which spans mutualism through parasitism, becomes necessary. This very shift in the center of gravity may itself be counted as one of the session's conclusions.

Glossary

• Bidirectional Human-AI Alignment: A two-way alignment process in which humans not only align AI to themselves but also align themselves to AI.
• RLHF (Reinforcement Learning from Human Feedback): A representative method that tunes AI behavior using human preference data.
• CPC (Collective Predictive Coding): A theory holding that shared symbols and norms form as people minimize prediction error (free energy) in a distributed way.
• MARL (Multi-Agent Reinforcement Learning): A reinforcement-learning framework in which multiple agents learn and act at the same time.
• Niche construction: Organisms remaking the very environment (niche) that suits them.
• Mutualism: A symbiotic relationship in which both parties benefit. By contrast, commensalism benefits only one party, and parasitism harms one.
• Dark pattern: A manipulative or deceptive design that steers users toward choices they would not otherwise make. LLM dark patterns are the language-model version.
• Plurality: An orientation toward the coexistence and negotiation of diverse values rather than convergence on a single value. In Taniguchi's CPC framework, it is reinterpreted computationally as a stable, multimodal distribution of shared beliefs.
• Symbiotic Alignment: An umbrella term for a view of alignment that, instead of unilaterally bending AI to human values, seeks to design a symbiotic relationship in which humans and AI transform together. Several formulations were shown in this session. Taniguchi, drawing on CPC, casts it as maintaining a multimodal Plurality, while Hamada casts it in terms of ecological symbiosis (partner selection, cheater detection, sanctions — an "immune system"), so the emphasis differs under the same label. Related frameworks in which humans and AI co-shape values include Zeng et al.'s Super Co-alignment and Shen et al.'s bidirectional alignment.
• HITC (Humanity-in-the-Circle) / IS-Infra (Intelligence-Symbiosis Infrastructure): A vision for keeping humans as meaningful members of an AI-centered society, together with the social infrastructure that supports it.

Related links

Note: verify that all links are live and correctly formatted before publication.

• Conference official "public event (free of charge)" page: https://www.ai-gakkai.or.jp/jsai2026/openevent/
• Hua Shen: https://hua-shen.org/
• Bidirectional Human-AI Alignment (workshop): https://bialign-workshop.github.io/
• Shen, H. et al. "Towards Bidirectional Human-AI Alignment." (NeurIPS 2025, position paper): https://arxiv.org/abs/2406.09264
• Shen, H. et al. "Mind the Value-Action Gap: Do LLMs Act in Alignment with Their Values?" (EMNLP 2025): https://aclanthology.org/2025.emnlp-main.154/
• Shi, Y. et al. "The Siren Song of LLMs." (arXiv:2509.10830): https://arxiv.org/abs/2509.10830
• Taniguchi, T. "Collective Predictive Coding Hypothesis: Symbol Emergence as Decentralized Bayesian Inference." (Frontiers in Robotics and AI, 11, 1353870, 2024): https://doi.org/10.3389/frobt.2024.1353870
• Taniguchi, T., Hayashi, Y., Hirose, M., Oka, M., Suzuki, K., Witkowski, O., Tang, A. "Symbiotic Alignment via Collective Predictive Coding: A Theoretical Framework for Co-Creative Human–AI Ecosystems." (submitted to Artificial Life): https://alife.institute/en/blog/symbiotic-alignment-cpc/
• Zeng, Y. et al. "Super Co-alignment of Human and AI for Sustainable Symbiotic Society." (arXiv:2504.17404): https://arxiv.org/abs/2504.17404
• AI Alignment Network (ALIGN): https://www.aialign.net/

Credits and notes

• This article is a session report compiled from the day's audio recording and each speaker's slides.
• Names, affiliations, and titles are as of the time of presentation.