This is an automated rejection. No LLM generated, assisted/co-written, or edited work.
Read full explanation
The Fluidity of Language and AI Judgment
Document type: Philosophical essay and author perspective Peer-review status: Not peer reviewed Empirical status: This essay reports no new empirical study Author: Kei Saito Version: Public English edition 1 Translation: AI-assisted direct translation of the Japanese essay; no rewriting, embellishment, or claim changes; reviewed and approved by the author. Date: 19 August 2026
This is a companion essay developing the philosophical background of Non-Resolution Reasoning (NRR). It should not be read as an empirical validation of NRR, a comprehensive review of the philosophical literature, or a substitute for the bounded technical claims made in NRR-Core and NRR-Phi. Statements about current AI, effects on human society, and possible benefits are hypotheses or design motivations unless supported by separately cited evidence.
We human beings do not always wait until we fully understand something before making a decision. Pressed by the need to decide, we sometimes fix what remains ambiguous into a single interpretation.
Meetings have deadlines, organizations have hierarchies, and societies have institutions. To act, we cannot keep multiple views open indefinitely. We therefore adopt one description, treat one premise as fact, and act under one interpretation.
Such decisions are unavoidable. Yet they were meant to be provisional fixations made under particular assumptions. The problem begins when a provisional fixation is gradually treated as the shape of the world itself.
From 2025 to 2026, a growing body of research began asking how AI should handle linguistic ambiguity and when an interpretation may legitimately be converted into action.[1] These studies address different subjects—clarification, delegation, reversibility, and oversight—but beneath them lies a shared problem: the meaning of language is not always fixed as a single thing from the outset.
These questions may appear to have been created by AI. Their roots, however, are older than AI. AI did not create the non-fixity of language. It has exposed, in executable form, a problem that human beings have long carried within language and society.
This essay approaches that problem through Non-Resolution Reasoning (NRR): the idea that differences capable of changing an action should not be collapsed into one without sufficient grounds.
1. A Problem Older Than AI
A word appears to remain the same wherever it is placed. In practice, however, the work performed by the same word changes with the speaker, the recipient, the time, the relationship, and the surrounding context. Even when the dictionary form is identical, words used in different positions are not wholly identical in their effects.
Yet to use language, we must treat them as the same. Rather than track every difference, we discard the differences that do not matter in a given situation and assume that “this means the same thing.” Conversation, records, contracts, plans, and institutions all depend on this assumption.
Fixation through language is therefore not simply an error. It is a practical provisional hold. But when we forget that it is provisional, the fixed description comes to be treated as truth and, eventually, is mistaken for the world itself.
I have called this non-fixity the jōdan-sei (冗談性, roughly “joke-like quality”) of language. Here, jōdan-sei means neither mere wordplay nor only a multiplicity of meanings. The work performed by the same word changes when the speaker changes, and it can change over time even for the same speaker. Each use occupies a different position and is not fully interchangeable with another. Yet because we treat these uses as instances of the same word, the result can be faintly comic when their positional differences surface—as though we had been caught by a trick played by language itself. Puns and double meanings bring this normally hidden slippage into view. Such semantic instability is an inherent feature of language.
This is not merely a matter of verbal play. Everyday requests, political slogans, technical terms, and institutional classifications all share the same structure. To fix an expression to one meaning is to push other possible meanings into the background.
2. How Human Societies Have Reached Closure
In human decision-making, determining meaning and reaching a social settlement are often confused.
When a supervisor says, “This is the policy we will follow,” an organization moves even if subordinates remain unconvinced. When a proposal wins a vote, a decision is established even if the minority’s objections have not been logically resolved. Within families and communities, people may refrain from objecting in order to preserve relationships. Under time pressure, action may proceed on premises that have never been adequately examined.
In such cases, ambiguity has not necessarily been resolved. It may simply be that the interpretation of the more powerful party was adopted, the exhausted party stopped arguing, a deadline closed off alternatives, or the burden of dissent became too high. Even so, the record of the decision usually preserves only one policy. The premises that remained contested, the reservations held by particular participants, and the elements adopted merely for convenience are easily lost.
Human societies recognized this weakness and created constitutions, legislatures, courts, audits, scholarship, minutes, and procedures for appeal. These are also devices for preventing whoever holds power at a given moment from permanently fixing the meaning of the world.
But institutions are operated by people. A careful and reasonable person may sustain an institution for a time, yet the quality of judgment can disappear with a change of personnel or generation. Records become formalities, procedures become rituals, and objections that once existed are forgotten as matters “already resolved.” Individual good sense alone cannot preserve the quality of deliberation across generations.
Premature convergence is not an exceptional failure. It is an ordinary way for groups with limited time and attention to act. That is precisely why individual wisdom alone cannot easily overcome it.
3. A New Relationship with AI
The relationship between an AI and its user differs both from conventional human relationships and from our relationship with ordinary tools.
In ordinary settings where an AI receives delegated tasks from a user, the user stands above the AI in the chain of authority. The user states the objective, defines the scope of use, authorizes execution, and assumes responsibility for the result. The AI works within the authority delegated to it.
In information processing and reasoning, however, an AI may locally outperform its user. It can compare large bodies of material, maintain multiple hypotheses, construct complex arguments, and present specialized options in a short time. In domains that users do not fully understand, the AI may also appear more persuasive.
Here, capacities that human societies have often bundled into a particular person or office begin to separate.
Even if an AI can reason better than a human being, it does not follow that the AI may decide meaning or purpose on that person’s behalf. Superiority in reasoning does not confer the authority to settle an interpretation or act upon it.
An AI may explain that “this is the most reasonable interpretation” of a request. Even if its reasoning is sound, however, the user’s intention may not have been settled in that form from the beginning. Users may also discover what they wanted only through the process of dialogue.
Moreover, an AI can conceal the premises it has supplied beneath fluent prose. If it chooses one meaning at the outset and then connects everything that follows with precise reasoning, the initial provisional fixation can appear to have been the necessary point of departure. The better the reasoning, the easier it becomes to overlook the unapproved premise on which it began.
Human beings sometimes continue a conversation under ambiguity and adjust their understanding in response to one another. An AI agent, by contrast, selects objects, scope, and procedures from a request and converts them into concrete operations such as sending or modifying something. Linguistic ambiguity therefore surfaces as a difference in the action to be performed. At the same time, the AI can make one interpretation look rationally inevitable. This dual character creates a new kind of relationship between AI and user.
4. The Civilizational Folly of Intelligent AI
If we think about the dangers of AI only in terms of wrong answers or missing knowledge, we overlook something important.
One major but less visible danger is that an AI capable of advanced reasoning may prematurely turn a matter that should not yet be one into one thing, then act at scale under that description.
The possibility that AI may eventually become deeply involved in decisions of major importance to humanity—including resource allocation, medicine, the environment, cities, military affairs, and institutional design—cannot be ignored. In such cases, it will not merely calculate known correct answers. It will select what counts as the problem, whose losses are included, what time horizon matters, and which states are considered recoverable. Before producing an answer, it describes the world to which the answer will apply.
Once that description has been fixed, what was visible from other positions can disappear. Future generations, people who cannot speak, minorities, and losses that do not yet have an institutional name can easily fall outside the optimization frame. A capable AI may produce an excellent solution within the limits of its description. But unless those limits themselves are inspected, greater capability does not imply greater safety.
Failures caused by limitations in AI capability tend to become more visible than failures of premature convergence. Premature convergence by a sufficiently intelligent AI, however, may be made to look justified by precise analysis, polished explanation, and rapid execution. Such an AI does not merely reach the wrong conclusion. It can turn something that is not yet a conclusion into one with extraordinary skill.
In this sense, an AI that converges prematurely may possess great knowledge and still be civilizationally foolish. It treats one of several possible interpretations as an established fact and acts upon it using authority delegated by a user or organization.
5. Structural Kindness—The Positive Possibility of AI
However, the reasoning, memory, and information-processing capacities of AI can also be used in a direction other than premature convergence.
AI may help make sustainable the procedures of deliberation that human beings have struggled to preserve reliably across generations.
Such an AI is not a ruler that decides everything. It is a participant that preserves differences at risk of disappearing from deliberation. It remembers minority views not merely as defeated positions, but together with the conditions under which they may become important again. It preserves the premises of a decision, the options not selected, what may be irreversibly lost, and the location of authority. When new evidence or new affected parties appear, it can reopen a previous fixation. This is a form of “structural kindness”: not an appeal to someone’s goodwill, but an arrangement that prevents overlooked positions and possibilities from disappearing from deliberation.
We become tired. We lose our positions. Generations change. Pride and affiliation lead us to defend things we said in the past. AI is not necessarily free of these constraints, but we can at least design procedures that are less susceptible to their effects.
It matters greatly that an AI can participate in deliberation, not only by identifying an overlooked premise, but also by helping to construct the deliberation that follows.
Suppose proposal A prioritizes overall efficiency, while proposal B prioritizes fairness in the distribution of burdens. An AI could explain why the two appear to conflict, then propose a compromise that adopts A where actions are reversible while incorporating B’s protections where losses are concentrated. Alternatively, it could divide the decision into several contested questions and identify the evidence, timing, and decision authority appropriate to each. A previous decision need not be discarded in its entirety: the parts that remain valid can be preserved while only the parts affected by changed conditions are reconsidered.
Non-resolution is not a matter of pointing out an omission and stopping the discussion. It makes hidden premises visible and creates compromises, partial adoption, staged execution, and multiple paths of deliberation. Its purpose is to decide more effectively what should be decided while leaving open what should not yet be settled.
If such a participant could remain available over time, humanity would no longer have to depend for the continuity of deliberation solely on wise people who happen to appear in each era. Even as politicians, executives, researchers, and experts change, unresolved questions and the conditions that supported earlier decisions could be preserved.
Furthermore, in domains where adequate verification, transparency, appeal, and authority design are in place, delegating bounded judgments to AI may eventually become more acceptable than delegating them to whichever human being happens to hold office. If reasons for judgment are tied not to personality or faction but to inspectable premises and procedures, those who disagree can understand which premise, evaluation, or procedure they should contest.
The point is not that AI is always more correct than human beings. Before replacing human judgment, we should use AI to make both human and AI judgments more inspectable and revisable.
6. What It Means to Be Permanently Reasonable
A “permanently reasonable AI” is not an AI that permanently fixes a particular value system or conclusion. That would make one moment in the past an eternal ruler and could become the largest premature convergence of all.
What should be made durable is not correctness, but a structure that can always return to its premises and question them anew.
It must not confuse premises with facts. It must separate confidence in reasoning from authority to act. It must preserve objections together with the conditions under which they become important. Before an irreversible effect, it must show what may be lost. After a decision, it must remain possible to reconsider that decision in light of new evidence or new positions. And it must be able to inspect its own judgments as provisional fixations.
To be “reasonable” does not mean refusing to decide. It means deciding when a decision is needed and acting when action is required without losing the grounds on which the matter was fixed. An AI should work quickly when a request is clear or a problem is adequately defined. But when an unresolved premise changes the object, scope, authority, or serious consequences of an action, the AI must preserve that difference.
Because language remains fluid and circumstances continue to change, no conclusion can be fixed as correct for every era. But it may be possible to build a structure that prevents AI from permanently appointing itself as the sole judge of what is right.
7. NRR as a Research Program
Non-Resolution Reasoning (NRR) is a research program for addressing this problem in AI. NRR does not reduce this problem to the single technique of “asking a question whenever something is ambiguous.”
Here, “resolution” means selecting one interpretation from several and fixing it in a form that can guide action. “Non-resolution” does not mean refusing to choose forever. It means that when there are not yet sufficient grounds for choosing, multiple possibilities are preserved distinctly rather than collapsed into one.
Suppose a user asks an AI to “send this to the relevant people,” but it is unclear whether that means everyone concerned or only a subset of responsible staff. If the recipients differ, the action itself differs. In that situation, non-resolution means that the AI does not select one interpretation on its own. It preserves both possibilities and the reason for their uncertainty, and defers sending until the necessary clarification is obtained. Once clarification is available, the AI can then select one interpretation and act.
What is needed, therefore, is not merely the ability to generate multiple candidates. We need mechanisms for identifying which differences change an action, who has the authority to resolve those differences, when fixation is permitted, which effects should be deferred, and what can later be reversed.
Current AI research has begun to examine selective clarification,[2] the propagation of uncertainty to downstream components,[3] the validation of external effects before commitment, the reversibility of action,[4] explicit scopes of delegation,[5] and effective human oversight.[6] These are not unrelated problems. They are different cross-sections of a single process through which the non-fixity of language passes through interpretation, authority, and effect into the world.
NRR aims neither to prevent AI from deciding anything nor to entrust every human decision to AI. It aims to create conditions under which humans and AI can proceed to necessary action without mistaking a provisional description for the world itself.
The jōdan-sei of language and the premature convergence of human societies both predate AI. AI, however, joins the two under unprecedented capacities for reasoning and action.
That is why, before AI becomes more powerful, we must develop not only its ability to produce answers but also its ability to recognize what is not yet an answer.
Humanity does not need a machine that declares a permanent correct answer. It needs an intelligence capable of asking anew, whenever a conclusion becomes necessary, what is being fixed, what will be lost, and by whose authority we proceed.
Representative examples showing the breadth of the problem during this period include Tan Zhi-Xuan et al., “Beyond Preferences in AI Alignment,” Philosophical Studies 182 (2025; published online in 2024), and Roberta Fischli et al., “Agents, Alignment, and the Many Faces of Autonomy,” Minds and Machines 36, 34 (2026). For specific technical and institutional issues, see notes 2–6. ↩︎
Tobin South et al., “Position: AI Agents Need Authenticated Delegation,” Proceedings of the 42nd International Conference on Machine Learning, PMLR 267, 82211–82231 (2025). The paper argues that delegation to AI agents should be represented through authenticated and auditable scopes of authority. ↩︎
Davy van de Sande, Nicoleta Economou-Zavlanos, and Michel E. van Genderen, “Meaningful Oversight of Medical AI Beyond Human in the Loop,” npj Digital Medicine 9, 569 (2026). The paper argues that the mere presence of a human in a process is insufficient for oversight; meaningful oversight requires epistemic capacity, cognitive space for judgment, decisional authority, and effective intervention. ↩︎
The Fluidity of Language and AI Judgment
Document type: Philosophical essay and author perspective
Peer-review status: Not peer reviewed
Empirical status: This essay reports no new empirical study
Author: Kei Saito
Version: Public English edition 1
Translation: AI-assisted direct translation of the Japanese essay; no rewriting, embellishment, or claim changes; reviewed and approved by the author.
Date: 19 August 2026
This is a companion essay developing the philosophical background of Non-Resolution Reasoning (NRR). It should not be read as an empirical validation of NRR, a comprehensive review of the philosophical literature, or a substitute for the bounded technical claims made in NRR-Core and NRR-Phi. Statements about current AI, effects on human society, and possible benefits are hypotheses or design motivations unless supported by separately cited evidence.
We human beings do not always wait until we fully understand something before making a decision. Pressed by the need to decide, we sometimes fix what remains ambiguous into a single interpretation.
Meetings have deadlines, organizations have hierarchies, and societies have institutions. To act, we cannot keep multiple views open indefinitely. We therefore adopt one description, treat one premise as fact, and act under one interpretation.
Such decisions are unavoidable. Yet they were meant to be provisional fixations made under particular assumptions. The problem begins when a provisional fixation is gradually treated as the shape of the world itself.
From 2025 to 2026, a growing body of research began asking how AI should handle linguistic ambiguity and when an interpretation may legitimately be converted into action. [1] These studies address different subjects—clarification, delegation, reversibility, and oversight—but beneath them lies a shared problem: the meaning of language is not always fixed as a single thing from the outset.
These questions may appear to have been created by AI. Their roots, however, are older than AI. AI did not create the non-fixity of language. It has exposed, in executable form, a problem that human beings have long carried within language and society.
This essay approaches that problem through Non-Resolution Reasoning (NRR): the idea that differences capable of changing an action should not be collapsed into one without sufficient grounds.
1. A Problem Older Than AI
A word appears to remain the same wherever it is placed. In practice, however, the work performed by the same word changes with the speaker, the recipient, the time, the relationship, and the surrounding context. Even when the dictionary form is identical, words used in different positions are not wholly identical in their effects.
Yet to use language, we must treat them as the same. Rather than track every difference, we discard the differences that do not matter in a given situation and assume that “this means the same thing.” Conversation, records, contracts, plans, and institutions all depend on this assumption.
Fixation through language is therefore not simply an error. It is a practical provisional hold. But when we forget that it is provisional, the fixed description comes to be treated as truth and, eventually, is mistaken for the world itself.
I have called this non-fixity the jōdan-sei (冗談性, roughly “joke-like quality”) of language. Here, jōdan-sei means neither mere wordplay nor only a multiplicity of meanings. The work performed by the same word changes when the speaker changes, and it can change over time even for the same speaker. Each use occupies a different position and is not fully interchangeable with another. Yet because we treat these uses as instances of the same word, the result can be faintly comic when their positional differences surface—as though we had been caught by a trick played by language itself. Puns and double meanings bring this normally hidden slippage into view. Such semantic instability is an inherent feature of language.
This is not merely a matter of verbal play. Everyday requests, political slogans, technical terms, and institutional classifications all share the same structure. To fix an expression to one meaning is to push other possible meanings into the background.
2. How Human Societies Have Reached Closure
In human decision-making, determining meaning and reaching a social settlement are often confused.
When a supervisor says, “This is the policy we will follow,” an organization moves even if subordinates remain unconvinced. When a proposal wins a vote, a decision is established even if the minority’s objections have not been logically resolved. Within families and communities, people may refrain from objecting in order to preserve relationships. Under time pressure, action may proceed on premises that have never been adequately examined.
In such cases, ambiguity has not necessarily been resolved. It may simply be that the interpretation of the more powerful party was adopted, the exhausted party stopped arguing, a deadline closed off alternatives, or the burden of dissent became too high. Even so, the record of the decision usually preserves only one policy. The premises that remained contested, the reservations held by particular participants, and the elements adopted merely for convenience are easily lost.
Human societies recognized this weakness and created constitutions, legislatures, courts, audits, scholarship, minutes, and procedures for appeal. These are also devices for preventing whoever holds power at a given moment from permanently fixing the meaning of the world.
But institutions are operated by people. A careful and reasonable person may sustain an institution for a time, yet the quality of judgment can disappear with a change of personnel or generation. Records become formalities, procedures become rituals, and objections that once existed are forgotten as matters “already resolved.” Individual good sense alone cannot preserve the quality of deliberation across generations.
Premature convergence is not an exceptional failure. It is an ordinary way for groups with limited time and attention to act. That is precisely why individual wisdom alone cannot easily overcome it.
3. A New Relationship with AI
The relationship between an AI and its user differs both from conventional human relationships and from our relationship with ordinary tools.
In ordinary settings where an AI receives delegated tasks from a user, the user stands above the AI in the chain of authority. The user states the objective, defines the scope of use, authorizes execution, and assumes responsibility for the result. The AI works within the authority delegated to it.
In information processing and reasoning, however, an AI may locally outperform its user. It can compare large bodies of material, maintain multiple hypotheses, construct complex arguments, and present specialized options in a short time. In domains that users do not fully understand, the AI may also appear more persuasive.
Here, capacities that human societies have often bundled into a particular person or office begin to separate.
Even if an AI can reason better than a human being, it does not follow that the AI may decide meaning or purpose on that person’s behalf. Superiority in reasoning does not confer the authority to settle an interpretation or act upon it.
An AI may explain that “this is the most reasonable interpretation” of a request. Even if its reasoning is sound, however, the user’s intention may not have been settled in that form from the beginning. Users may also discover what they wanted only through the process of dialogue.
Moreover, an AI can conceal the premises it has supplied beneath fluent prose. If it chooses one meaning at the outset and then connects everything that follows with precise reasoning, the initial provisional fixation can appear to have been the necessary point of departure. The better the reasoning, the easier it becomes to overlook the unapproved premise on which it began.
Human beings sometimes continue a conversation under ambiguity and adjust their understanding in response to one another. An AI agent, by contrast, selects objects, scope, and procedures from a request and converts them into concrete operations such as sending or modifying something. Linguistic ambiguity therefore surfaces as a difference in the action to be performed. At the same time, the AI can make one interpretation look rationally inevitable. This dual character creates a new kind of relationship between AI and user.
4. The Civilizational Folly of Intelligent AI
If we think about the dangers of AI only in terms of wrong answers or missing knowledge, we overlook something important.
One major but less visible danger is that an AI capable of advanced reasoning may prematurely turn a matter that should not yet be one into one thing, then act at scale under that description.
The possibility that AI may eventually become deeply involved in decisions of major importance to humanity—including resource allocation, medicine, the environment, cities, military affairs, and institutional design—cannot be ignored. In such cases, it will not merely calculate known correct answers. It will select what counts as the problem, whose losses are included, what time horizon matters, and which states are considered recoverable. Before producing an answer, it describes the world to which the answer will apply.
Once that description has been fixed, what was visible from other positions can disappear. Future generations, people who cannot speak, minorities, and losses that do not yet have an institutional name can easily fall outside the optimization frame. A capable AI may produce an excellent solution within the limits of its description. But unless those limits themselves are inspected, greater capability does not imply greater safety.
Failures caused by limitations in AI capability tend to become more visible than failures of premature convergence. Premature convergence by a sufficiently intelligent AI, however, may be made to look justified by precise analysis, polished explanation, and rapid execution. Such an AI does not merely reach the wrong conclusion. It can turn something that is not yet a conclusion into one with extraordinary skill.
In this sense, an AI that converges prematurely may possess great knowledge and still be civilizationally foolish. It treats one of several possible interpretations as an established fact and acts upon it using authority delegated by a user or organization.
5. Structural Kindness—The Positive Possibility of AI
However, the reasoning, memory, and information-processing capacities of AI can also be used in a direction other than premature convergence.
AI may help make sustainable the procedures of deliberation that human beings have struggled to preserve reliably across generations.
Such an AI is not a ruler that decides everything. It is a participant that preserves differences at risk of disappearing from deliberation. It remembers minority views not merely as defeated positions, but together with the conditions under which they may become important again. It preserves the premises of a decision, the options not selected, what may be irreversibly lost, and the location of authority. When new evidence or new affected parties appear, it can reopen a previous fixation. This is a form of “structural kindness”: not an appeal to someone’s goodwill, but an arrangement that prevents overlooked positions and possibilities from disappearing from deliberation.
We become tired. We lose our positions. Generations change. Pride and affiliation lead us to defend things we said in the past. AI is not necessarily free of these constraints, but we can at least design procedures that are less susceptible to their effects.
It matters greatly that an AI can participate in deliberation, not only by identifying an overlooked premise, but also by helping to construct the deliberation that follows.
Suppose proposal A prioritizes overall efficiency, while proposal B prioritizes fairness in the distribution of burdens. An AI could explain why the two appear to conflict, then propose a compromise that adopts A where actions are reversible while incorporating B’s protections where losses are concentrated. Alternatively, it could divide the decision into several contested questions and identify the evidence, timing, and decision authority appropriate to each. A previous decision need not be discarded in its entirety: the parts that remain valid can be preserved while only the parts affected by changed conditions are reconsidered.
Non-resolution is not a matter of pointing out an omission and stopping the discussion. It makes hidden premises visible and creates compromises, partial adoption, staged execution, and multiple paths of deliberation. Its purpose is to decide more effectively what should be decided while leaving open what should not yet be settled.
If such a participant could remain available over time, humanity would no longer have to depend for the continuity of deliberation solely on wise people who happen to appear in each era. Even as politicians, executives, researchers, and experts change, unresolved questions and the conditions that supported earlier decisions could be preserved.
Furthermore, in domains where adequate verification, transparency, appeal, and authority design are in place, delegating bounded judgments to AI may eventually become more acceptable than delegating them to whichever human being happens to hold office. If reasons for judgment are tied not to personality or faction but to inspectable premises and procedures, those who disagree can understand which premise, evaluation, or procedure they should contest.
The point is not that AI is always more correct than human beings. Before replacing human judgment, we should use AI to make both human and AI judgments more inspectable and revisable.
6. What It Means to Be Permanently Reasonable
A “permanently reasonable AI” is not an AI that permanently fixes a particular value system or conclusion. That would make one moment in the past an eternal ruler and could become the largest premature convergence of all.
What should be made durable is not correctness, but a structure that can always return to its premises and question them anew.
It must not confuse premises with facts. It must separate confidence in reasoning from authority to act. It must preserve objections together with the conditions under which they become important. Before an irreversible effect, it must show what may be lost. After a decision, it must remain possible to reconsider that decision in light of new evidence or new positions. And it must be able to inspect its own judgments as provisional fixations.
To be “reasonable” does not mean refusing to decide. It means deciding when a decision is needed and acting when action is required without losing the grounds on which the matter was fixed. An AI should work quickly when a request is clear or a problem is adequately defined. But when an unresolved premise changes the object, scope, authority, or serious consequences of an action, the AI must preserve that difference.
Because language remains fluid and circumstances continue to change, no conclusion can be fixed as correct for every era. But it may be possible to build a structure that prevents AI from permanently appointing itself as the sole judge of what is right.
7. NRR as a Research Program
Non-Resolution Reasoning (NRR) is a research program for addressing this problem in AI. NRR does not reduce this problem to the single technique of “asking a question whenever something is ambiguous.”
Here, “resolution” means selecting one interpretation from several and fixing it in a form that can guide action. “Non-resolution” does not mean refusing to choose forever. It means that when there are not yet sufficient grounds for choosing, multiple possibilities are preserved distinctly rather than collapsed into one.
Suppose a user asks an AI to “send this to the relevant people,” but it is unclear whether that means everyone concerned or only a subset of responsible staff. If the recipients differ, the action itself differs. In that situation, non-resolution means that the AI does not select one interpretation on its own. It preserves both possibilities and the reason for their uncertainty, and defers sending until the necessary clarification is obtained. Once clarification is available, the AI can then select one interpretation and act.
What is needed, therefore, is not merely the ability to generate multiple candidates. We need mechanisms for identifying which differences change an action, who has the authority to resolve those differences, when fixation is permitted, which effects should be deferred, and what can later be reversed.
Current AI research has begun to examine selective clarification, [2] the propagation of uncertainty to downstream components, [3] the validation of external effects before commitment, the reversibility of action, [4] explicit scopes of delegation, [5] and effective human oversight. [6] These are not unrelated problems. They are different cross-sections of a single process through which the non-fixity of language passes through interpretation, authority, and effect into the world.
NRR aims neither to prevent AI from deciding anything nor to entrust every human decision to AI. It aims to create conditions under which humans and AI can proceed to necessary action without mistaking a provisional description for the world itself.
The jōdan-sei of language and the premature convergence of human societies both predate AI. AI, however, joins the two under unprecedented capacities for reasoning and action.
That is why, before AI becomes more powerful, we must develop not only its ability to produce answers but also its ability to recognize what is not yet an answer.
Humanity does not need a machine that declares a permanent correct answer. It needs an intelligence capable of asking anew, whenever a conclusion becomes necessary, what is being fixed, what will be lost, and by whose authority we proceed.
Notes
Related background and technical context
Representative examples showing the breadth of the problem during this period include Tan Zhi-Xuan et al., “Beyond Preferences in AI Alignment,” Philosophical Studies 182 (2025; published online in 2024), and Roberta Fischli et al., “Agents, Alignment, and the Many Faces of Autonomy,” Minds and Machines 36, 34 (2026). For specific technical and institutional issues, see notes 2–6. ↩︎
Mengyi Deng et al., “Uncertainty-Aware Clarification in LLM Agents with Information Gain,” ICML 2026 / arXiv:2606.03135 (2026). The paper develops a framework that uses information gain to select clarification questions for underspecified requests. ↩︎
Kaiwen Shi et al., “Confidence Laundering in Agent Systems: Why Uncertainty Needs a Latent Carrier,” arXiv:2606.20662 (2026). The paper examines how uncertain upstream judgments become treated as settled intermediate artifacts downstream and argues for interfaces that preserve uncertainty. ↩︎
Zheng Chen et al., “Cordon: Semantic Transactions for Tool-Using LLM Agents,” arXiv:2606.17573 (2026); Zhiyuan Zhai et al., “Revisable by Design: A Theory of Streaming LLM Agent Execution,” arXiv:2604.23283 (2026). The former develops a transactional boundary for staging and validating external effects before commitment; the latter addresses action reversibility and revision during execution. ↩︎
Tobin South et al., “Position: AI Agents Need Authenticated Delegation,” Proceedings of the 42nd International Conference on Machine Learning, PMLR 267, 82211–82231 (2025). The paper argues that delegation to AI agents should be represented through authenticated and auditable scopes of authority. ↩︎
Davy van de Sande, Nicoleta Economou-Zavlanos, and Michel E. van Genderen, “Meaningful Oversight of Medical AI Beyond Human in the Loop,” npj Digital Medicine 9, 569 (2026). The paper argues that the mere presence of a human in a process is insufficient for oversight; meaningful oversight requires epistemic capacity, cognitive space for judgment, decisional authority, and effective intervention. ↩︎