People and AI in one undertaking: verification, responsibility and authorship

Epigraph

making and acting are different

Aristotle, Nicomachean Ethics, VI.4, an explanatory clause within a sentence translated by W. D. Ross, Internet Classics Archive. The distinction concerns a philosophical account of production and action; it does not allocate wisdom and production between people and a contemporary AI system.

What this article is about

In shared work, AI may prepare text while people select material, examine it, and present a result. But checked remains empty if nobody knows what was compared or who accepted the latest version. In a hypothetical story, an urban walking club makes an exhibition booklet from approved street notes and participants' recollections. Irina uses AI to organise a draft, Sergei checks defined information, and the group selects the completed text. The article considers material transfer, verification's boundaries, error correction, and honest description of the work's origin. Philosophical distinctions between judgement and reproduction are compared with bounded official NIST contexts, preserving their different foundations. The practical aim is intelligible participation without promises of infallibility, provider comparisons, or deriving legal authorship rights merely from someone pressing a button, writing a prompt, or adding their name.

I. Shared work begins with a defined result

The booklet should present selected streets and recollections rather than become a complete city guide. This boundary determines which information is needed and which claims should not be added for completeness. Irina can ask AI to group approved notes and propose transitions. Sergei can check dates and names against available grounds. This does not yet accept every new description appearing in a coherent draft. The result needs sufficiently concrete definition: a text of defined scope for the exhibition, based on accepted material and with recollections clearly identified. The system's participation then concerns an operation within shared work rather than an unknown promise to prepare everything necessary.

People also perform different actions. Collecting notes, selecting content, checking, and deciding to present a result do not constitute one default author role. One person can combine them, but this should be understood. If Sergei agreed to check dates, his name does not mean he examined artistic transitions, obtained participants' consent to use recollections, or accepted every future question from visitors. The limit needs naming while material is still being transferred. The shared task then does not lack an owner at the point where a separately checked part becomes a whole presented to others. Preparing text and accepting this particular version remain different actions.

The initial boundary may change. The group decides to add a historical section absent from the notes. This is a new task even if the booklet keeps its name. Different grounds and suitable examination are needed, or perhaps a smaller scope. Sergei's earlier agreement cannot confirm every added fact. Nor must Irina personally fill every new gap because she can use AI. Quickly obtaining a coherent version does not create actual research readiness. Shared work starts with an available object and the possibility of accepting it; expansion requires reconsidering those conditions within the changed scope.

II. Accessible material is not permitted for every use

The club holds recorded recollections, photographs, and working notes. Defined fragments are approved for the exhibition, while a participant asked that some details remain within discussion. Sending the whole archive to a system is more convenient than preparing a limited selection, but convenience establishes no basis for transfer. Irina needs to understand which materials are available for this operation and under what conditions. The article does not supply a particular service's retention rules: they need current examination for the chosen use. The editorial boundary concerns possession of a file and consent to defined use. A shared project does not cancel the origin and substantial limits of information people provided within a different scope.

Material sufficient for the assignment can be prepared without excessive personal details. An approved recollection with its status identified might be supplied instead of a complete conversation containing family history. Yet shortening itself needs precision: a condition cannot be removed if it made a phrase permissible only as recollection or related it to another place. If data are anonymised, the actual scope of that action must be understood without promising that context cannot identify a person. The article provides no universal protective technique. It proposes naming what is prepared and what is withheld, and considering a chosen system's properties through actual conditions rather than the general word safe.

Existing permission should also be recognised. Where someone approved a particular text for this booklet, every permissible move between internal sections need not trigger a repeated request. But revision can exceed the earlier object: a recollection becomes a promotional testimonial for the club, or a photograph becomes another project's cover. Its basis then needs reconsideration. A clear boundary supports independent action within permitted scope and noticing expansion. Examining material includes factual truth and correspondence with accepted use. Attractive wording may preserve a date yet change the role of a person's statement supplied to the group with a different meaning.

III. A coherent draft does not establish a fact

AI connects two notes and adds that participants gathered in one courtyard every spring. The initial information contained only two separate recollections. New regularity makes the story smoother but remains an unconfirmed claim. NIST AI 600-1, section 2.2 considers confidently false content, including explanations and citations. This is a bounded official context about generative AI risks rather than a measurement of the chosen system's quality. The booklet's editorial connection is simple: substantial assertions need comparison with actual material. A transition's persuasiveness does not ground treating an added event as part of known history.

Asking the system again whether the account is correct may identify something to consider, but a renewed assurance is not independent confirmation. Irina can ask it to mark added assumptions, then examine them against the original notes. Sergei may see that a source exists but confirms another year or place. A citation's existence does not establish correspondence between claim and source. This text requires examining the particular relationship: what a statement rests on and which qualifications were preserved. Sometimes the available solution is removing an addition or marking it as an editorial question rather than obtaining confident alternatives until one sounds familiar.

Not every sentence requires identical examination. A transition may express editorial thought while a date is a verifiable assertion. Distinguishing them avoids spending effort on ritual while leaving a significant fact ungrounded. If the transition itself says an event happened, smoothness does not exempt it from consideration. Scope depends on the task and possible consequences of error. The article proposes no universal verification level for every AI application. It concerns ordinary exhibition work; medical, legal, technical, or other specialised content needs suitable competence and its own grounds, neither supplied by this everyday example.

IV. Verification should leave an intelligible record

Sergei writes that he looked through everything, but Irina does not know whether he compared dates with notes, corrected spelling, or merely read the sequence. A short record can name the actual object: these names and dates were checked against the identified material version; two connections remain unestablished; recollections' provenance was preserved. This is no mandatory large report for every paragraph. The record should be proportionate to what the next person relies on. To accept the booklet, that person needs to distinguish confirmed parts from an open condition. A general mark does not silently expand to examination of something the checker never accepted as a task.

The current version also needs naming. After verification, Irina changes a transition and introduces a new assertion about a joint walk. Sergei's earlier record concerns the preceding text. It may remain applicable to unchanged dates without confirming the addition. The actual change should be examined and checked within a suitable scope while earlier history is preserved. This need not mean researching every unchanged part again. Yet the old mark cannot keep being presented as acceptance of a new whole. A version connects checking with what was actually considered rather than making it an eternal property of a similarly named file.

A verification record should not be manufactured instead of verification. An automatic table can show matching fields or existing links without establishing that a recollection retains its meaning. Calling this result mechanical is useful. Human reading has limits too: it does not confirm an unknown fact merely because the reader noticed no contradiction. Both kinds of evidence and their scope can be preserved. Appendix C of NIST AI RMF 1.0 distinguishes roles and forms of human participation, including oversight's limitations. It appoints no identical examination for every operation; an approach is selected for a defined task, and its occurrence cannot be replaced by the checker's name.

V. Creating an option and choosing an action are different questions

In Nicomachean Ethics, VI.4–5, Aristotle distinguishes making and acting while considering art and practical wisdom within his own moral and philosophical account. This is neither a historical AI theory nor an assertion that a system can make things while people automatically possess wisdom. A limited connection helps reveal two questions: can a persuasive text be prepared, and should it be presented within this scope? A well-written booklet may still contain an unsupported promise, an inappropriate relabelling of recollection, or material whose use was unaccepted. Formal quality does not authorise the next act by itself.

Irina chooses to retain a courtyard remark because it animates the account. If the event's grounds are unknown, that is her editorial decision rather than technological inevitability. It can be named and reconsidered without making a human participant guilty of every future response from any system. Responsibility here means examining accepted actions and available conditions in this undertaking; legal allocation is not established. This meaning preserves a practical object. Who selected material, what was known, which limit was seen, and which option was accepted? These questions are more precise than a general assertion that a person must guarantee everything AI ever creates nearby.

A reasonable limit can be accepted instead of full readiness. The group removes a disputed paragraph while preserving the remaining booklet, or postpones release pending clarification. This establishes no technical incapacity of the system and does not make created parts useless. It connects decision with available grounds. Personal writing does not bypass the problem either: an author can mistake the same date without AI assistance. The claim and action need examining beyond a sentence's origin. The philosophical connection preserves human judgement's question; official technical contexts identify tool limitations without replacing one question with the other or promising an infallible combined solution.

VI. Understanding differs from convenient reproduction

Irina receives a fluent explanation of a street's history and can repeat it almost verbatim. But she does not know where fact, recollection, and editorial transition differ. In Seneca's Letter 33.7–11, independent judgement differs from repeating others' formulas while learning from predecessors remains possible. This is a Stoic moral context, not a prediction of machine writing. A limited connection concerns grounds accessible for one's own statement. If Irina presents text as accepted, explaining what it says and why a substantial part was included is useful. Repeating attractive wording creates no such understanding, regardless of who first typed it.

Understanding does not require knowing everything about every street personally. Someone with suitable knowledge can be consulted, an assertion limited, or uncertainty honestly named. A reference to a specialist should not become a new unsupported promise that the entire booklet was examined. If Sergei checked only two places, that is the available scope. Different competences can be connected with their boundaries preserved. Personal judgement also appears in choosing whom to ask and where to withhold acceptance for now. It is not isolation from help and does not demand repeating research already actually performed and still applicable to unchanged material.

The opposite mistake treats assistance as removing authorial participation altogether. Irina can accept AI's proposed structure, change it, and explain her choice. Sergei can use automatic name comparison, then consider an identified difference. These actions do not become empty because a tool performed an operation. Yet participation should be described factually. Structural assistance is not source verification, and personal editing establishes neither every image's originality nor a legal right to every material. The article preserves understanding and choice without reducing them to the number of personally typed words or declaring every delegated operation a refusal to think.

VII. The work's origin should be described without an invented legend

Human roles and AI use can be named within the scope that actually occurred: approved materials, draft organisation, editorial choices, defined checks. A suitable form depends on the task and actual conditions of presentation. The article establishes no universal legal disclosure formula or authorship rights. It proposes honest provenance. If a text is called a participant's personal recollection, generated experience cannot silently be attributed to that person. An editorial transition should not look like an eyewitness document. Naming a tool does not itself correct this substitution; the status of the particular statement must be preserved.

Nor should AI be credited with independent examination that never happened. The system helped identify dates, and Sergei compared them with accepted material. These roles differ. Saying that a person and AI checked the booklet without naming scope can conceal that both used the same initial draft and nobody consulted factual grounds. Sometimes a short editorial explanation suffices while detailed working history remains available internally. Brevity does not require a false impression of complete independence. Substantial actions can be shown so the next person does not mistake preparation assistance for confirmation of the whole result.

Human contributions should not disappear into a general credit either. One person gathered recollections, another agreed use, and a third determined structure and checked selected facts. Running the system does not make Irina sole creator of everything. But a shared participant list does not make everyone author of every decision. Roles can be named more accurately while accepted conditions are preserved. This editorial practice promises no resolution of legal rights disputes. It prevents another substitution: participation history becomes a legend about a magical machine or one person to whom others merely supplied material, although they actually created significant parts and grounds of the shared text.

VIII. Handover includes open limitations

Sergei finishes checking and gives Irina a list of corrections. The group needs to know which corrections she actually implemented and which version it is now considering. A comments file does not mean changes were completed. Irina can accept some, decline others on named grounds, and leave a question open. The next role receives an actual state rather than an assumption that all remarks were accepted automatically. If somebody else then lays out the booklet, preserving the distinction between recollections and editorial text, current credits, and unchecked boundaries matters. Handover is not complete through sending material while its substantial conditions remain unavailable for the next action.

AI can help prepare a change summary, but it needs relating to the actual version. A system may report a correction absent from the file or overlook an addition. Depending on the object, mechanical comparison and targeted reading of the altered place may be available. A summary should not be treated as evidence of occurrence merely because it exists. Nor must all earlier material be read again if the basis for limited renewed consideration is clear and earlier examination remains applicable. Scope follows the actual change and dependency. Shared work then neither loses history in every new draft nor extends old confirmation over an unexamined part.

The recipient must also accept a defined role. A booklet's layout designer may have no opportunity to check historical information. Text with an intelligible state is passed on rather than a silent obligation to reconstruct every inquiry. If layout changes meaning or removes a qualification, a new verification object arises. A readiness decision includes this current relationship. An earlier author's presence does not relieve others of their own accepted actions; a tool's presence does not expand these into unlimited support. Handover becomes clear through result and participation limits: what a person received, what was accepted as their task, and what must return for consideration if changed.

IX. Errors need correction with accurate history

After release, a date proves inconsistent with accepted grounds. The original note, prepared version, human checking, and subsequent choice need distinguishing. Perhaps the system added a year, perhaps it was already in a note, or perhaps a correction never entered the latest version. These are different causes and available actions. Every defect cannot be explained in advance by technology or one person's inattention. If a cause is unestablished, that should be named. A confirmed part can be corrected while the error's origin remains a separate question. Concrete examination requires neither psychological diagnosis, a promise of complete reliability, nor a legal determination of every possible consequence.

History is needed proportionately. Preserving versions, grounds, and significant messages about the date may suffice; an entire archive of participants' private conversations does not become necessary merely because an error occurred. An available clarification of text already received and a future edition need determining. A corrected master does not automatically change printed copies, and an updated file does not prove earlier recipients learned of a substantial correction. Scope depends on subject and conditions. The article appoints no procedure for every release. It distinguishes current correction, communicating clarification, and future arrangements so one successful step does not become invented closure of the whole history.

An error can ground changing a defined transition: for example, accepting only a version connected with the latest verification record. This proposal needs actual application rather than merely a new form. Nor does an error make every other part useless or erase the checker's labour. If Sergei checked another place correctly, that result remains within its scope. If Irina corrected the date, she has not thereby independently established the streets' entire history. Precise recognition prevents responsibility being exchanged for universal blame or credit. Shared work stays open to reconsideration when history connects with actual decisions rather than a convenient image of a good person and a bad machine.

X. Judgement remains an action when some work is automated

An AI-assisted booklet can preserve intelligible human participation. This requires neither declaring a system a full moral participant nor treating every generated result as alien to a human author. The selected operation, actual material, checked scope, and decision about a particular version need seeing. Here, responsibility concerns people's accepted actions under available conditions; legal questions need separate consideration. This frame does not promise a person can prevent every error. Neither does it authorise presenting every output as though preparation made choice inevitable and removed grounds for reconsidering it.

Official NIST documents provide bounded technical context: different forms of human participation and confidently false content require attention to task arrangement. AI RMF 1.0 is a voluntary 2023 edition; the current official page reports its revision, providing no basis to announce a completed successor. The generative AI profile dates from 2024. These documents confirm neither the chosen product's quality nor this practice's success. Aristotle and Seneca consider their own philosophical questions of action and judgement. Comparison does not turn them into one AI doctrine. Different references help hold different questions without presenting a citation to authority as actual verification of the booklet.

In the hypothetical story, work ends with a defined result: Irina prepared an option from permitted material, Sergei named what was actually examined, the group accepted the current version, and the origin of recollections and editorial text stayed clear. Open questions were neither concealed nor presented as confirmed. This proves no future infallibility. It is a history of available actions that can be continued or corrected. People and AI enter one undertaking without dissolving human choice into a tool or making a person a magical guarantee of everything. Cooperation gains substance through what was actually prepared, considered, accepted, and passed to the next participant.

What this means in practice

Name the shared result and the operation assigned to AI. Prepare material within permitted scope, preserving substantial qualifications and human statements' status. Separately establish the chosen service's current conditions where transfer's permissibility depends on them. Do not claim these are known through a general article. Distinguish a coherent option from factual grounds. For substantial assertions, identify available comparison with material or a source; specialised objects need suitable competence. Do not use repeated system assurance as independent checking. If grounds are unavailable, limit the assertion, postpone it, or honestly leave the question open.

When checking, name the object, version, and open conditions. Mechanical matching, human reading, and actual acceptance have different scopes. If text changes, consider the actual change and update applicable confirmation while preserving earlier history. Do not demand renewed examination of unchanged parts without a reason, or present an old record as confirmation of a new assertion. Describe human roles and tool use accurately enough that recollection's origin does not merge with an editorial transition. Naming AI does not correct a statement's false status; a human signature does not prove every part checked and every right established.

In handover, show what was actually corrected, what was accepted, and what the next person must return for consideration if changed. After an error, preserve necessary history, distinguish established facts from an assumed cause, and name correction's actual scope. This practice promises neither elimination of risks nor legal authorship resolution. It supports intelligible participation: a tool performs a defined operation, people retain available judgement and accepted actions, and a result acquires an accurate state instead of a general checked label. Automation can thereby become part of work in which choice's grounds and readiness's limits remain accessible to participants themselves.

Key quotation

But it is one thing to remember, another to know.

Seneca, Moral letters to Lucilius, 33.8, complete concluding sentence of the paragraph translated by Richard M. Gummere, Wikisource. The context of 33.7–11 distinguishes independent judgement from reproduction while retaining the possibility of using predecessors' road; this is not an AI prediction.

Connections between traditions

Question	Historical or technical reference	Editorial connection	Boundary
Creating or acting	Aristotle, VI.4–5	An option's quality does not authorise the next act	Not allocation of wisdom between people and AI
Understanding or repeating	Seneca, 33.7–11	One's statement needs accessible grounds	Help and learning are not rejected
How human participation is arranged	NIST AI RMF, Appendix C	Roles and verification limits need naming	A voluntary framework guarantees no oversight
What confirms a claim	NIST AI 600-1, 2.2	Coherence and citation do not establish a fact	Not assessment of a particular product's quality
What to pass onward	Different questions of action and verification	Version, provenance, and open conditions are preserved	No single philosophical and technical doctrine

The map preserves differences between philosophical teachings and contemporary official technical contexts. The exhibition's practical arrangement is an editorial proposal without an empirical promise of reliability.

Primary sources

Aristotle. Nicomachean Ethics, VI.4–5. Translated by W. D. Ross, Internet Classics Archive; complete selected sections about making, acting, and practical wisdom were read, and metadata checked. The exact print edition and worldwide electronic rights were not independently established. The source does not consider contemporary AI.
Seneca. Moral letters to Lucilius, 33.7–11. Translated by Richard M. Gummere, Wikisource; the complete selected context was read afresh, with collection metadata reused from actual reading in this batch. The collection displays Heinemann/Putnam MCMXXV; the letter's exact print volume was not separately collated. Worldwide rights of the particular edition were not established; site terms are not treated as an unconditional licence for all material.
NIST AI RMF 1.0, Appendix C. NIST AI 100-1, 2023, printed pages 40–41. The identical complete Appendix C context actually read for A39 in this batch was reused, together with the official framework page reporting revision of 1.0. Reading the entire framework is not claimed; no completed successor edition is assumed.
NIST Generative Artificial Intelligence Profile, section 2.2. NIST AI 600-1, 2024. Actual A39 reading of the complete main section 2.2 on confabulation, printed page 6, and selected recommendations for checking grounds and sources was reused. The whole profile is not claimed read. Documents are voluntary and confirm no product promises, legal authorship, or effectiveness of the proposed practice. No direct NIST quotations are reproduced; worldwide rights are not assumed.

About the authors

Andrey Zarinsh is the author and founder of United Wisdom. He explores inner transformation, the unity of traditions, intercultural dialogue, and contemporary spiritual discipline. Instagram: @andrey.zarinsh.

Maria Evilkut is a co-author of United Wisdom. Alongside her research, she seeks to test these ideas in practice and observe how principles of unity and mutual understanding work in real life between people and cultures. Instagram: @ilkutki.