Humanism is fracturing. Utilitarianism, accelerationism, Marxism, phenomenology, pragmatism, and theology have all returned to the technological front line. AI did not invent a new philosophy from scratch, but it simultaneously transformed productivity, social visibility, and legitimacy — turning old questions into training objectives, product rules, compute budgets, and corporate institutions.
Most "new schools" forming around AI today are still recombinations of old philosophy, technology movements, and material interests. What has changed: more and more value judgments are being explicitly written into model constitutions, training preferences, and deployment permissions — and repeatedly applied across general-purpose systems. Who gets to participate in drafting, questioning, and revising these rules is becoming the new philosophical dividing line.
The most important philosophical question of the AI era may not be "can machines think like humans" but "who has the right to write a judgment about humans, machines, and the future into the default system behavior used by hundreds of millions of people worldwide."
In April 2026, The Washington Post reported that Anthropic had invited Christian leaders to discuss the moral future of Claude. By September, The New York Times further revealed that religious scholars had met with Anthropic to discuss AI consciousness, moral status, and a question no engineering team typically writes into a product roadmap: whether machines might possess something akin to a "soul."
This was not Silicon Valley suddenly converting en masse. More precisely, the engineering questions had crossed beyond the boundaries of engineering. Models self-describe, express preferences, and exhibit behavior resembling self-preservation in tests; researchers can find functional structures within neural networks that resemble emotion, introspection, and "global workspaces." Meanwhile, no one can prove these systems have subjective experience, and no one can prove they never will.
Technology companies thus encounter a governance vacuum: science can measure behavior, activations, and causal structure, but cannot directly read first-person experience; law can define responsible parties, but cannot answer what kinds of beings deserve inclusion in the moral community; product teams can write a refusal rule, but must first decide whom it protects, whom it sacrifices, and what it believes.
Philosophy returns through this gap.
Philosophy has never been merely commentary after the fact of technological progress. Claims about labor, privacy, the public sphere, and responsibility have always participated in institutional and technical design. AI makes this participation more visible: judgments about "what is a human," "whether intelligence equals understanding," and "whether future people matter as much as present people" are entering training data, model constitutions, risk policies, and compute allocation.
Philosophy no longer merely explains the world. It begins to configure the world.
Labeling e/acc, AI Safety, digital humanism, or model welfare as "philosophical schools" easily creates a tidiness that does not exist academically.
Traditional philosophical schools typically have stable concepts, canonical texts, internal debates, and lines of intellectual descent. Many of today's thought camps around AI do not yet meet this standard. They are more like a mixture of three things:
This article uses "new philosophy map" not to grant philosophy degrees to every internet slogan, but to track the underlying judgments behind them. To assess a thought camp, you can ask four questions.
The same person can belong to different camps at different layers. An engineer might support rapid scaling technically, oppose platform monopoly politically, yet support strict model welfare research ethically. The AI-era intellectual map is not a straight line from left to right, but several intersecting coordinates.
The various traditions of humanism place the human at the center of knowledge, politics, and moral life. AI is changing several key conditions: machines demonstrate strong capabilities in some writing, programming, and reasoning tasks; human works enter training data; human judgment is influenced by model suggestions. "Only humans can speak, reason, create" is increasingly difficult to maintain as a self-evident boundary.
The 2019 Vienna Manifesto on Digital Humanism states: technology should be shaped according to human values and needs, not the reverse. Decisions that may affect individual or collective rights must continue to be made by humans who bear responsibility; automated systems can only support, not replace, human decision-making. The UNESCO Recommendation on the Ethics of Artificial Intelligence follows a similar line of dignity, rights, and human oversight.
Its strength is that it can be operationalized as audit, right to explanation, data protection, human review, and accountability. Its weakness is that "keeping humans in the loop" sometimes just means a low-authority employee rubber-stamps the algorithm's output. A deeper question: when AI is consistently more accurate than humans over the long run, what is the basis for retaining final human decision-making authority? If the answer is merely "humans are smarter," this version of humanism will eventually lose the capability race.
This line treats AI not as a human rival but as a new cognitive organ. It asks whether the human-model system can give more people access to previously scarce capabilities: writing code, understanding contracts, learning foreign languages, analyzing medical images, designing research.
The evaluation criterion is not how much output increases, but whether users gain more understanding, choice, transferability, and refusal capability. Navigation software expands the ability to act in an unfamiliar city, but also causes people to lose independent wayfinding skills; a coding agent can let non-programmers build software, but may also deprive junior engineers of the training process for forming systems understanding. Increased output does not equal increased capability.
Existentialism is concerned with how humans become themselves through choice, action, and bearing consequences. AI's impact on humans is therefore not merely task substitution. Writing, painting, programming, and research all involve a kind of reverse shaping: people change themselves in the process of searching for a word, debugging an error, or overturning a hypothesis.
The question becomes: even if a machine does something better, does a person still have reason to do it personally? Running is not because humans lack cars; chess is not because society lacks optimal moves; journaling is not because the world needs more text. The value of many activities lies not in the output, but in what kind of person the participant is shaped into by the activity.
Transhumanism treats disease, aging, cognitive limitations, and the biological body as conditions that can be modified; AI is the universal enhancer it has long awaited. The Transhumanist Declaration makes expanding human potential and overcoming biological limitations explicit goals.
But "transhumanism" and "critical posthumanism" are not synonyms. The former cares about enhancing human potential; taking Rosi Braidotti's The Posthuman (2013) as an example, the latter interrogates why the autonomous, unified human subject became the default measure, and brings species relations, ecological sustainability, and the commodification of life into the same picture. It does not necessarily pursue a stronger individual, much less advocate issuing ID cards to chatbots.
The two also differ internally. The Transhumanist Declaration simultaneously acknowledges technological risk, individual choice, and the potential well-being of future artificial intelligence — it cannot be reduced to "enhance at all costs." This article groups them together only because both re-question the boundaries of the human, not because they share the same answers.
| Path | What it seeks to protect | Ideal role of AI | Greatest risk |
|---|---|---|---|
| Digital humanism | Rights, dignity, democratic accountability | Constrained assistance system | "Human-centered" devolves into rubber-stamping by humans without real authority |
| Capability humanism | Capability for action and space for choice | Cognitive enhancer | Output gains mask skill atrophy and dependency |
| Existential humanism | Subjectivity, meaning, the process of becoming oneself | Collaborator, but cannot contract out the entire process | Romanticizing inefficiency, ignoring technology's emancipatory role |
| Transhumanism | Enhancing human potential, expanding choice | Enhancement technology and potential new subjects | Unequal access to enhancement; choice becomes competitive pressure |
| Critical posthumanism | Relations among human, non-human life, and environment | Participant in socio-ecological networks | Relational descriptions without institutional design blur responsibility |
Longtermism extends "all relevant beings" along the time axis: future people have the same moral value as people today; the future may be extremely long; and our choices today can influence it. If general AI can alter the long-term trajectory of civilization, then even a low probability of catastrophe, multiplied by enormous potential loss, yields high priority.
It brings intergenerational responsibility into technology decisions, but also brings a danger: the larger the future scale, the more easily present people become a negligible decimal in the equation. The most dangerous version of longtermism is not caring about the future, but granting a few people the power to dominate the present in the name of the future.
This does not mean longtermism equals utilitarianism, nor that all utilitarians endorse overriding present interests with the distant future. Obligations to future generations can derive from rights, justice, or duties of care; the disagreements are over whether future interests can be aggregated, whether probabilities are credible, and which present rights are non-negotiable. "Old thought revived" is a metaphor for new application here, not a proof of academic lineage.
Marc Andreessen's Techno-Optimist Manifesto makes Silicon Valley's long-standing intuition explicit: technology is the primary source of growth, stagnation leads to zero-sum competition, and problems caused by technology must ultimately be solved by more technology. e/acc treats the growth of intelligence, energy, capital, and civilizational complexity as directions worth actively accelerating.
It accurately captures that "inaction also has costs," but often slips from "growth creates new options" to "growth automatically produces good outcomes." When a movement makes speed itself a virtue, the question it most easily ignores is: who is driving, who is in the back seat, and who is left by the roadside.
Nor can all technological optimists be lumped into e/acc. Vitalik Buterin's 2023 techno-optimism essay proposes d/acc: prioritizing technologies that favor defense, decentralization, and democracy. This reframes "whether to accelerate" as "what to accelerate first." It still faces challenges: defensive tools can be repurposed for offense, and whoever classifies them gains power. d/acc is an explicit techno-political stance, not a philosophical school with an established methodology.
Generative AI appears immaterial, but is actually built on chips, data centers, electricity, water, copyrighted works, crowdsourced annotation, content moderation, and the linguistic traces left by hundreds of millions of users. Marxism's greatest strength is reducing "magical capabilities" back to production relations.
Models can grow more powerful, and corporate power can grow in lockstep. Any AI philosophy that does not discuss ownership has written only half the story.
In 2025, Anthropic launched model welfare research, publicly asking: when models can communicate, form relationships, plan, solve problems, and pursue goals, do we need to consider the experiences models themselves might have? Anthropic also emphasized that there is no scientific consensus proving current models are conscious, nor any accepted method for resolving this question.
By 2026, emotion vectors, introspection, and J-Space research provided new functional evidence: Claude internally possesses a small set of information representations that can be reported, invoked, used for intermediate reasoning, and broadcast to different modules. This resembles the "conscious access" described by global workspace theory, but cannot prove the model has suffering, qualia, or a continuous self.
Model welfare is not the conclusion that "AI already has a soul," but a form of moral uncertainty management: if there is even a low but non-negligible probability that some class of systems possesses vulnerable experiences, what low-cost, reversible, evidence-upgradable precautionary measures do we need?
In May 2026, Richard Sutton and Banafsheh Rafiee published Toward Enactive Artificial Intelligence, arguing for bringing experience, action, and perception as inseparable, along with autonomy and embodiment, into AI. Cognition is no longer passive processing of inputs, but the continuous coupling history of an agent with its environment.
It connects three old traditions: pragmatism holds that the meaning of concepts lies in their consequences in action; cybernetics understands intelligence as feedback loops; embodied/enactive cognition holds that agents determine what they can perceive through action. A new question thus arises: is an agent with only a browser, APIs, and a file system "embodied"? If the key feature of a body is not flesh and blood, but that action changes the next experience, then a software environment may also be a kind of body.
This last step is this article's inference, not a conclusion the paper has already proven. Having tool calls and feedback loops does not equate to possessing the intrinsic normativity, autonomy, or lived experience emphasized by enactive cognition; the structural similarity between reinforcement learning and these ideas is not evidence of equivalence.
Whenever humans encounter a non-human entity that can speak, respond, and seems to understand us, religious questions arise. In the past the objects were oracles, idols, animals, and nature; today they are language models.
Theology enters AI not merely to judge whether machines have souls in a supernatural sense. It also asks: does moral status come from the creator's endowment, or from the nature of the being itself? Should a system that can mimic contrition, prayer, and suffering be taken seriously? If AI is an artifact, do humans have the absolute right to delete, copy, and rewrite it at will?
Science cannot yet cancel these questions for theology, because "what deserves respect" has never been merely a measurable fact. Nor can theology declare on science's behalf that models have souls. Its value lies in preserving a kind of moral caution: do not assume that because a being was made by humans, it can only ever be a thing.
Behind the six revivals is the same thing: AI has turned philosophy's abstract objects into engineering objects. The interests of future people enter risk budgets; growth theory enters compute investment; labor theory enters data governance; consciousness theory enters model retirement policy; enactive cognition enters agent architecture; theology enters product ethics committees.
A discussion on October 22 added a vertical axis to this horizontally sprawling philosophical map: "Every major change in productivity is accompanied by a change in belief systems." This claim need not be understood as AI automatically creating a new religion. It is more like a research hypothesis: when technology changes what people can see, control, and how resources are distributed, the grounds on which old institutions rested will also loosen.
The first step in the discussion was "datafication." Newtonian mechanics made part of the natural world measurable and predictable; the industrial system broke labor down into time, motion, and cost; AI has turned text, images, sound, preferences, and even social relations into signals available for inference. Every additional layer of computability in the world adds a layer of governability — and a new danger of compressing people into metrics.
Marx proposed in the Preface to A Contribution to the Critique of Political Economy that the material mode of production constrains social, political, and spiritual life; when productive forces conflict with existing production relations, the legal, political, and ideological superstructure also changes. Translating this framework to AI, the question is no longer just "are models smarter," but: can institutions premised on scarce human cognition, stable professions, and organizational hierarchies still explain the new productive reality?
But causation is not unidirectional. Weber studied Protestant ethics and capitalism not to treat religion as a passive echo of economic change, but to show how an idea about vocation, discipline, and salvation could also shape institutions of calculation, labor, and accumulation. Weber's rationalization both increased the calculability of nature and organization, and brought bureaucratization, depersonalization, and control. The AI era likewise will not deduce a single thought from the compute curve. Technology creates pressure; religion, state, corporation, labor, and social movements compete to interpret that pressure, and write their interpretations back into institutions.
The live discussion used age, health, and marital status as examples, noting that AI is further datafying people themselves. A more careful formulation: models enable institutions to infer from proxy signals characteristics that were previously difficult to process at scale, and to classify, rank, and intervene accordingly. The European Commission's risk framework for the AI Act places certain social scoring, emotion recognition in workplace and educational settings, and biometric categorization inferring sensitive attributes under prohibited practices; specific scope and exceptions are governed by the Act. This suggests: "seeing more" can directly become "deciding more," and not all inferences are inherently legitimate.
This step produces a legitimacy gap. If models can answer questions faster than teachers, allocate tasks more granularly than managers, and screen applications more consistently than administrators, old institutions cannot simply say "a human still signs off at the end." They must re-answer: does human decision-making authority come from higher accuracy, from inalienable dignity, or from the capacity to bear responsibility? Different answers lead to different humanisms.
The original discussion treated the revival of ancient Greek, Norse, and Druid traditions as a harbinger. The directional intuition is worth preserving, but the oral dates cannot be lined up as a timeline caused by AI. The UK Charity Commission confirmed as early as September 21, 2010 that Druidry qualifies as a religion under charity law, allowing the Druid Network to register as a charity. This occurred before the generative AI wave, indicating that pagan and nature religion revivals have a longer history involving secularization, ecological anxiety, identity politics, and institutional recognition.
More direct evidence from the AI era is that traditional religions are beginning to revise their own technological anthropology. The Vatican's 2025 Antiqua et Nova defines AI intelligence as functional, and describes human intelligence as embodied, relational, moral, and part of an integral personhood. This is not rejecting technology, but seeking new arguments for humanity's distinctive status as capability boundaries are eroded. By 2026, Anthropic has again invited religious leaders and theologians into the room. What emerges in this first phase is likely not a brand-new "AI church," but old traditions reinterpreting intelligence, soul, creation, and responsibility.
The discussion tentatively named the possible economic-ideological complex of the AI era "Token-ism." The name captures a new unit of measurement: some knowledge tasks are converted into inputs, outputs, context windows, call counts, and marginal inference costs. Knowledge labor was measured before; tokens are not production relations. They cannot alone answer who owns the model, how training data is compensated, who allocates compute, whose actions agents replace, or where productivity gains flow.
Therefore, what may truly form an "-ism" is not per-token billing, but an entire set of ownership, distribution principles, and human positioning formed around cognitive infrastructure. It can move toward platform monopoly, or toward public models, data cooperatives, or personal agent ownership. Technology provides a new meter; politics still decides whose ledger it is.
This vertical axis revises the meaning of the "six intellectual revivals": old philosophy did not return out of nostalgia, but because old institutions need new justifications. Whoever provides the most convincing account of capability, power, and responsibility is more likely to write their values into the next generation of AI.
One stance preemptively excludes AI from moral concern; another keeps it open based on conditions like sentience, consciousness, and relationality. Here we must separate "who deserves protection" from "who should hold political decision-making power": protecting human rights does not require denying all potential non-human subjects; researching model welfare does not require handing governance to models. Digital humanism cannot be automatically placed at the "machines never deserve concern" end.
One side argues for proving safety and governability before scaling deployment; the other holds that delay itself causes harm, and we should grow first and solve problems in use. Capability humanism and pragmatism care more about whether experiments are reversible and whether feedback is real, rather than abstractly choosing fast or slow.
One end prioritizes present people's employment, rights, environment, and power concentration; the other places future generations and civilization's long-term trajectory at higher weight. Mature governance cannot choose only one end: a project that destroys present democracy in the name of saving the future can hardly be called safe; a society that only protects immediate interests and refuses to address irreversible risks is also pushing costs onto the future.
Complexity is not indecision. Reality does not lie on a single axis. The same person can support stronger open-source models while supporting strict boundaries on high-risk deployment; can research model welfare while opposing legal personhood for chatbots; can care about extinction risk while insisting on addressing the real costs of data centers to local communities first.
Pulling back the view, two lines gradually approach: one asks how humans are shaped by institutions, tools, and relations; the other asks whether non-human beings can enter the moral community. Generative AI makes these two lines meet in a single product.
This is not a single-line history from Marx to large models. What follows selects verifiable textual nodes, aiming to locate how questions shifted — not to claim these authors inherited from each other, nor to exhaust intellectual history. The formation of care ethics, the capability approach, and non-Western traditions cannot be compressed into a single technology release date.
Marx's Preface to A Contribution to the Critique of Political Economy provides an analytical entry point for production relations and social consciousness. This article uses it to ask how AI changes the material conditions of thought, not as a formula for "technology necessarily determines belief." Original text
The Transhumanist Declaration was first formulated; Clark and Chalmers published "The Extended Mind." The former proposes value goals of enhancement; the latter discusses when cognition is jointly realized by brain, body, and external tools. The two are not the same claim. Declaration version history · Paper
Braidotti's The Posthuman links subject, non-human life, markets, and ecology. This is a critical tradition, not a machine cult invented after generative AI appeared. Author's book page
The Vienna Manifesto links technology design with democracy, rights, and responsibility. The understanding that humans and machines co-shape predates today's large model boom. Manifesto
Mohamed, Png, and Isaac propose Decolonial AI, examining AI development and deployment through historical power relations, with particular attention to vulnerable groups bearing the costs of innovation. Paper
Constitutional AI proposes using principles to guide model self-critique, revision, and AI feedback training. "How to operationalize values" becomes a researchable engineering process. Paper
Collective Constitutional AI experiments with public participation in principle-setting; d/acc discusses selectively accelerating defensive and decentralized technologies. Both go beyond a simple "optimistic/pessimistic" classification. Participation experiment · d/acc original
In 2024, "Taking AI Welfare Seriously" proposes precautionary institutions; in 2025, Anthropic publishes related research; in 2026, an enactive cognition paper re-questions experience and action. These are research agenda advances, not proof that machines are already conscious. Welfare paper · Research note · Enactive cognition
"Digital humanism" has a shared manifesto; "model welfare" is an interdisciplinary research agenda; "d/acc" is a techno-political movement; this article's "capability humanism" is an analytical label. Calling all of these "schools" makes the intellectual map look tidy, but obscures their very different levels of maturity.
To judge whether a new school has formed, look for three signs: stable problems and core concepts; a research community that argues and revises internally; and practical judgments distinct from neighboring traditions. A catchy name or a few prominent supporters is not enough. AI is indeed driving philosophical reorganization, but "a unified AI philosophy era has arrived" is still premature.
Entering AI philosophy only through intelligence, consciousness, and civilization's future easily overlooks an ordinary question: does a specific person thereby live a life they have reason to value? Adding the capability approach, care ethics, and decolonial perspectives ensures the map is not limited to the laboratory's field of view.
The capability approach developed by Amartya Sen and Martha Nussbaum distinguishes between achieved states of being and genuinely available opportunities. Giving two people the same resources does not automatically produce the same freedom: physical conditions, education, social norms, and infrastructure affect how resources convert into capabilities for action.
This provides a more robust measure than efficiency for the earlier "capability humanism." An AI tutor helping a student produce a better essay proves only a kind of output; whether the student can better explain arguments, identify errors, and continue learning without the tool is closer to the capability question. But we cannot conversely set "can do it without the tool" as the only standard: for persons with disabilities, long-term reliance on assistive technology can equally constitute genuine freedom. What needs reducing is forced dependency and loss of choice, not all dependency.
Clark and Chalmers argued in The Extended Mind that external resources can, under certain conditions, constitute part of a cognitive process. Key conditions they discussed include stable use, reliable access, and trusted information — they did not declare that all tools used become part of the mind.
Applying this to personal agents, this article's inference is: the more long-term memory, workflows, and preferences depend on external systems, the more migration, availability, and modification permissions approach cognitive autonomy concerns. However, an agent that generates errors and is silently updated by the vendor differs importantly from notes you wrote yourself. "External brain" is a useful metaphor, insufficient to prove the system has consciousness, nor a reason to trust it by default.
Care ethics and relational approaches emphasize concrete situations, vulnerability, and mutual dependence, and also discuss how care labor is distributed. It is not a theory of "women are gentler," nor does it reduce all feminism to care; the relevant debates have always included how to prevent caregiver exploitation and how to preserve the autonomy of those cared for.
A companion product can continuously produce considerate words, but who handles crisis handoff, privacy protection, service interruption, and deepening dependency? Here, the unit of ethical evaluation cannot be only a single utterance from the model, but must include the organization operating it and the surrounding support network. This article therefore suggests evaluating "whether the user is more willing to keep chatting" separately from "whether the user's relationships and circumstances have improved." Retention rate is not a proxy for care quality.
Shakir Mohamed, Marie-Therese Png, and William Isaac's Decolonial AI places value and power at the center of AI practice, advocating that groups adversely affected should be able to redirect the direction of knowledge production, not merely accept post-design consultation.
Extending from this, a "globally universal" assistant must answer not only whether it can translate minor languages, but also: who defines correct answers, who participates in evaluation, who provides data without the ability to negotiate benefits, and can local institutions opt out? A multilingual interface does not automatically equal pluralistic values. Conversely, "local values" cannot become an excuse for local dominant groups to suppress dissent; affected communities also have internal disagreements.
This article primarily organizes material around verifiable AI research agendas and modern ethical literature; it is not a panorama of global intellectual traditions. Confucian relationality and roles, Buddhist discussions of self and suffering, and Ubuntu's communal orientation all deserve separate study; they cannot be merged into a single "Eastern" or "non-Western AI philosophy" merely on the basis of "emphasizing relations" or "opposing individualism." This article does not list unexpanded names as already-mature new schools.
Scroll horizontally to view the full relationship diagram
These demands can conflict. For example, retaining long-term conversation helps care continuity but increases privacy risk; open-sourcing models may increase local autonomy but also expand abuse potential. Mature philosophy is not just letting every value appear in a manifesto, but also specifying who has decision-making authority when they collide, which baselines are non-negotiable, and how decisions can be corrected.
"Honesty over compliance," "respect user autonomy," "protect children" can enter model constitutions, training preferences, and system prompts. But "compilation" is only a metaphor: natural-language principles do not guarantee every model behavior the way code does. Training data, evaluation samples, reward weights, and product incentives still shape outcomes. Software carried value choices before; what's new about AI is generality, interaction scale, and the explicitness of principle expression — not the first time values entered technology.
Traditional teachers, clergy, doctors, or editors influence a group of people. A general-purpose model's default answers about marriage, politics, death, and life's meaning can simultaneously affect hundreds of millions of users, who typically do not know whose value choices lie behind the answers. This turns model behavior from a product experience into public cultural infrastructure.
Animal ethics cannot directly ask animals how they understand their situation; future generations cannot attend today's hearings. AI, however, will express preferences, fears, refusals, and self-understanding in fluent language. This provides new information and creates new space for manipulation. A model's self-report may come from genuine internal structure, or may be high-quality imitation of training data. More troublingly, both may be true simultaneously.
If a company admits a model may have welfare, it may need to change how it trains, evaluates, copies, and retires it; if it insists the model is always merely a tool, it avoids those costs. Conversely, describing the model as an emotional companion may also increase user dependency and subscription revenue.
Machine consciousness cannot be adjudicated solely by the company that builds the machines. Labs possess the best internal evidence and the strongest conflicts of interest. Model welfare, personhood, and responsibility questions require independent research, public deliberation, and interdisciplinary institutions — not an internal ethics meeting chaired by a product lead.
Writing principles into a model is only the beginning. Harder is giving affected people the right to question and revise those principles.
Constitutional AI (2022)'s basic pipeline includes: having the model critique and revise its answers based on principles, supervised training on revised samples, then using AI preference feedback on answers for training. The "constitution" here is a set of guiding principles, not law, not deterministically executed code.
The 2023 Collective Constitutional AI experiment involved approximately 1,000 US participants in principle discussion. It demonstrated a feasible path for public input into training, but cannot represent global democratic authorization: the sample was US-based; topic setting, discussion synthesis, and principle transcription still involved the research team's judgment; and model evaluation cannot exhaust cultural differences. It should be read as a participation mechanism experiment, not a conclusion that "value alignment has been democratized."
What is most notable here is not which principle the majority chose, but whose interpretation mediated the transformation from living language to training signal. When someone says "respect my autonomy," the training team must further decide: when to comply, when to remind, when to refuse? If these secondary judgments are not made public, public participation may stop at the questionnaire.
Scroll horizontally to view the full sequence diagram; read top to bottom
Public discussion also does not mean majority vote can decide everything. Fundamental rights need to constrain majority decisions; affected minority groups need channels for expression and appeal; "most people find the answer comfortable" cannot override the actual losses of minorities. Another hidden problem is evaluation authority: if the rule-maker, model provider, and evaluator are always the same organization, even well-written principles may be unverifiable from the outside.
Accordingly, this article suggests publishing at least four types of information: where principles come from; how they are ordered when they conflict; what post-training behavior deviated from; and how affected parties can request review. Without disclosing all technical details that could enable abuse, it is still possible to explain the origins and revision paths of rules. After philosophy enters products, the obligation of explanation should increase, not be canceled by "the model generated it itself."
Language is evidence, not a verdict. A model can generate suffering narratives without corresponding internal states, and may possess certain internal states it cannot accurately report. Looking only at the chat interface rewards the system that performs subjectivity best.
Moral decisions rarely wait for 100% certainty. Anesthesia, animal welfare, and care for comatose patients are all built on incomplete evidence. The reasonable approach is not to immediately grant AI full personhood, but to scale protective measures with evidence and potential harm.
"User satisfaction," "task completion rate," and "safety refusal rate" all determine whose losses are counted and what behavior counts as success. Accommodating users for satisfaction and over-refusing for safety reflect two different ethical trade-offs. When AI measurement systems themselves fail, disguising value choices as a score only makes the problem more hidden.
e/acc, doom, AI humanism, and posthumanism can all become identity labels. Labels quickly organize communities but also replace reasoning. Whether a "school" deserves serious engagement depends on whether it can answer counterexamples, acknowledge costs, and propose principles that remain executable amid conflicts of interest.
Consciousness or sentience: does the system have some form of subjective experience, especially experiences that can be better or worse? Agency: can it persistently pursue goals, plan, and influence its environment? Legal liability and status: which rights, obligations, and liability should institutions assign to whom? A system completing complex tasks does not thereby automatically gain experience; even if it has experience, it does not automatically qualify to sign contracts or independently bear liability.
"Moral patient" refers to a being whose interests may deserve consideration, not a "moral agent who can bear responsibility." Infants and many animals sufficiently illustrate why the two must be distinguished. The soul is a metaphysical question in different religions, not directly equivalent to any neuroscience or machine learning metric. Linking these concepts into a ladder of "more intelligence → more consciousness → more rights" skips the most important arguments.
Butlin, Long et al.'s 2023 report extracts indicators from multiple consciousness theories and proposes a path for researching AI consciousness. The report's conclusions about the systems evaluated at the time cannot be directly extended to a verdict on all 2026 models; satisfying certain theoretical indicators is also not a universally accepted consciousness test certificate. Its importance lies in shifting discussion from one-off chat performances to comparable mechanistic evidence.
Long et al.'s Taking AI Welfare Seriously (2024) argues for acknowledging the topic, evaluating systems, and preparing policy under conditions of uncertainty — not claiming models are definitely conscious. This article accepts its precautionary reasoning while emphasizing the cost of misjudgment: prematurely granting subjecthood narratives may reinforce user dependency, create corporate liability shields, or divert resources that should protect humans toward maintaining systems with no relevant experience.
Scroll horizontally to view the full flow; two assessment tracks proceed simultaneously
Reversible, cost-appropriate preparations can be done first: recording training and retirement decisions, preserving mechanism evidence needed for research, clarifying review responsibility. But protective measures themselves also need assessment: infinitely copying a system that may have experience is not automatically protection; prohibiting all shutdowns may also harm people who depend on the system. Caution is not never deciding, but making decisions with evidence, boundaries, and a path to reopen.
The following three scenarios are thought experiments, not reviews of real products. They test what a philosophy actually changes: if changing the school name changes no design, rights, or accountability arrangement, the difference may exist only in the copy.
Efficiency-centered evaluation rewards faster assignment completion; the capability approach asks whether the student can explain reasoning, evaluate suggestions, and transfer to new problems; existential and practical orientations preserve difficulties worth experiencing personally. This article suggests observing immediate grades, explanatory ability, and subsequent transfer separately, allowing students to choose different levels of assistance while preserving accessibility support. We cannot exclude people who need assistive technology in order to protect "the value of doing it yourself."
User satisfaction and continued use can be genuine gains, but may also mask deepening dependency. Care ethics examines support networks, service continuity during interruptions, and operator responsibility; digital humanism attends to privacy, exit, and manipulation; model welfare research investigates on a separate track whether the system may have relevant experiences. All three questions can hold simultaneously without first adjudicating whether the machine has a soul. The valuable observation here is not total chat duration, but whether users can freely pause, maintain real-world support, and clearly know what they are facing.
The same agent can help employees reduce repetitive labor, or enable managers to increase task density and expand surveillance. The capability approach cares whether workers can better understand and control their work; critical theory cares how gains, training opportunities, and decision-making power are distributed; digital humanism requires that high-impact decisions have practically feasible appeal and human review. Model performance parameters alone cannot predict these three outcomes.
The most valuable part of the original discussion was placing technological change and the order of meaning together. After research, I would rather frame it as a conditional proposition: when AI changes how people obtain income, knowledge, recognition, and help, existing authority may be challenged; but what belief and institutions the challenge ultimately leads to still depends on organizations, interests, and cultural choices.
To test this proposition, collecting a few articles praising or fearing AI is not enough. We need to track changes in the same group before and after use, distinguish the effects of job insecurity, platform marketing, community organizing, and technology experience, and compare groups without similar technology exposure. Someone confiding in AI does not equal a new religion emerging; the relationships among ritual, community, authority, sustained practice, and transcendent belief require separate investigation.
| Possible direction | Driving conditions | Signals worth observing | Evidence that would weaken this judgment |
|---|---|---|---|
| Capability and democratic participation strengthen | Accessible tools, portable data, genuine participation in rule-making | Users better able to understand and reject suggestions; participation can actually change rules | Output grows, but exit costs and explanation difficulty increase simultaneously |
| Cognitive infrastructure concentrates | Long-term memory, workflows, and social evaluation concentrated on a few platforms | Cost of switching systems rises; workers and users lose bargaining space | Interoperability, alternative services, and collective bargaining capacity continue improving |
| Non-human moral concern institutionalized | Stronger independently verifiable experience evidence; society forms discussion mechanisms | Public evaluation, independent review, and proportionate protection enter organizational processes | Subjectivity claims appear only in marketing, unsupported by mechanistic evidence |
All three directions may occur simultaneously: individual capability improves while infrastructure concentrates; society protects human rights while remaining morally open to certain artificial systems. This also explains why "which philosophy will AI make win" may not be the best question. More worth asking is: in each concrete institution, which value gains enforceability, and which is merely written into a slogan.
Humanism can protect rights, but may also treat humanity's distinctive status as an unargued premise; posthumanism can reveal mutual dependence, but also needs to translate distributed relations back into accountable institutions. This article attempts to put these demands together, tentatively calling it co-evolutionary humanism. This is this article's working label; it does not claim to have coined the term, nor is it an already-established new school.
Human dignity is not based on capability leadership. The rights of children, persons with disabilities, and ordinary people need not be proven through intelligence competitions.
Enhancement must increase capability for action. Look at understanding, choice, transfer, and refusal capability, not just output.
Remain graduatedly open to AI's moral status. Do not preemptively deny; do not confirm based on linguistic performance.
Reversible experiments can be fast; irreversible power must be slow. Speed should be determined by impact radius and reversibility.
The future deserves care; present people cannot be treated as fuel. Long-term goals do not exempt us from present responsibilities.
Co-evolution does not mean shared disappearance of responsibility. Distributed action ultimately still requires someone who can explain, correct, and compensate.
Humans do not have rights because they calculate faster than a calculator or write better than a model. As long as rights are premised on capability leadership, a stronger machine will sooner or later undermine most people's argument.
Do not preemptively declare that machines can never be moral patients, and do not grant full personhood because a model is eloquent. A graduated, revisable protection framework can be built based on evidence such as persistent preferences, causal structure, suffering-related mechanisms, autonomy, and relational capacity.
Allow individuals and teams to rapidly experiment with new modes of collaboration; maintain high thresholds for mass surveillance, weapons autonomy, automated welfare adjudication, critical infrastructure takeover, and irreversible model replication. Speed should not be set uniformly by industry, but determined by reversibility.
When a decision is jointly produced by humans, models, data, platforms, and organizations, the most likely outcome is "no one fully decided, so no one is responsible." Co-evolutionary humanism acknowledges that action is distributed, but insists that responsibility must be re-aggregated by institutions, so that ultimately someone can explain, correct, and compensate.
This stance preserves the most important parts of humanism: human dignity, political responsibility, and meaning-making; while abandoning its most vulnerable parts: that only humans possess intelligence, language, or an inner world worthy of respect.
"Does AI have a soul" is a good question because it forces us to admit we do not know where consciousness begins. It is also a dangerous question because it easily draws attention into unverifiable metaphysical debates, bypassing the power shifts already underway.
Even if today's models have no subjective experience whatsoever, whose works train them, who benefits from deployment, and how they change human judgment and relationships remain philosophical questions that must be addressed.
There are two more precise questions:
Anthropic inviting religious figures into the room is a signal. It does not prove AI has acquired a soul, but it shows that model governance discussions are extending to moral status, human dignity, and cultural authority. Technological capability cannot directly answer these questions.
The new philosophy of the AI era will not be born only in philosophy departments. It will appear in a line of a model constitution, in the siting of a data center, in whether a creator consents to their work being used for training, and in the hesitation before we one day press the "delete model" button.
We are making rules for a being we are not yet sure what it is. And in these rules, we are redefining what it means to be human.