The useful question is not “Was AI involved?” It is “Who did what, who decided, and who was responsible?”

The argument around AI-generated work keeps reaching for a binary.

Human or AI.

Original or generated.

Authentic or synthetic.

Made or prompted.

The appeal is obvious. Binaries are easy to label, easy to regulate and easy to turn into badges.

They are also becoming increasingly bad descriptions of how work is actually made.

A writer can spend six months researching a book, use an AI system to locate sources, reject three generated outlines, write every sentence herself, ask the model to identify structural weaknesses, rewrite two chapters and then approve the final manuscript.

Did AI make the book?

No.

Was AI involved?

Obviously.

A designer can generate twenty rough visual studies, select one composition, rebuild it manually, replace every generated element, direct the typography, correct the anatomy, control the colour system and deliver the final campaign.

Is the work human-made?

Mostly, perhaps.

AI-assisted?

Certainly.

But neither label tells us very much.

The useful question is not:

Was AI involved?

It is:

What was the division of agency?

That is where provenance becomes more interesting than disclosure.

Provenance is usually treated as a history of objects

When people hear provenance, they often think of origin.

Where did this photograph come from?

Who created this file?

When was it modified?

Which source image was used?

Which software touched it?

Was the record altered?

Those questions matter.

A provenance system should be able to establish relationships across time:

source
  ↓
revision
  ↓
transformation
  ↓
derived artifact
  ↓
publication

But once generative systems enter the process, recording only the artifact history becomes insufficient.

Suppose the record tells us:

draft-01.md
   ↓
GPT-5.6
   ↓
draft-02.md

What happened?

Did the model write Draft 2?

Did it correct punctuation?

Did it propose changes that the author manually considered?

Did it retrieve research?

Did the writer paste in three paragraphs and ask whether the argument made sense?

Did the machine rewrite everything while the human clicked Accept?

The tool relationship alone cannot tell us.

The provenance of the artifact is present.

The provenance of agency is missing.

Agency has several dimensions

We tend to talk about authorship as though it were one action.

It is not.

Making something can involve several kinds of agency.

Someone may:

• originate the purpose;
• define the problem;
• select the sources;
• generate alternatives;
• perform an operation;
• interpret the result;
• reject possibilities;
• revise;
• approve;
• publish;
• bear responsibility.

Sometimes one person performs all of these.

Sometimes a team divides them.

Sometimes software performs some of them.

AI makes that division easier to obscure because one conversational interface can participate in many stages.

A provenance system therefore needs a richer vocabulary.

Not:

human = 60%
AI = 40%

That looks precise while telling us almost nothing.

Something closer to:

ORIGIN
Farah

RESEARCH DIRECTION
Farah

SOURCE RETRIEVAL
Farah + AI research tools

OUTLINE GENERATION
AI proposed alternatives

OUTLINE SELECTION
Farah

DRAFTING
Farah

STRUCTURAL CRITIQUE
AI

REVISION
Farah

FINAL APPROVAL
Farah

PUBLICATION AUTHORITY
Farah

Now we can see the collaboration.

That is an agency map.

Execution is not authority

This distinction becomes especially important as AI systems become more capable of acting.

A machine may perform most of the mechanical work without becoming the principal agent behind the project.

Consider software development.

A human founder decides that a tool should exist.

She specifies the requirements.

An AI coding system writes large portions of the implementation.

The human tests it.

Rejects one architecture.

Approves another.

Changes the privacy model.

Refuses a feature.

Chooses the release conditions.

Signs off on deployment.

If we measure authorship only by keystrokes, the machine may appear to have done most of the work.

If we measure agency, the picture changes.

HUMAN
purpose
constraints
judgment
risk acceptance
final authority

AI
implementation
suggestions
refactoring
test generation
documentation

The distinction does not diminish machine contribution.

It places it correctly.

The machine executed substantial work.

The human governed the work.

That is what human-led AI should mean.

Not that humans must manually perform every operation.

That humans remain meaningfully responsible for what the system is doing and why.

Conversely, human labour does not always mean human agency

The reverse is also possible.

Imagine a worker receives AI-generated text and is told to manually type it into another system.

A human touched every character.

A human was physically present.

The typing rhythm may be unmistakably biological.

Yet the human exercised almost no creative authority.

Now compare that with an editor who never types an original sentence but makes hundreds of decisions about what survives.

Who had more agency?

The answer is not obvious from activity logs.

This is why Proof of Life, process telemetry and authorship provenance need to remain separate.

Human presence can support an agency claim.

It cannot define one.

Agency lives in decisions.

And decisions are more difficult to record than clicks.

OriON should record claims, not pretend to read minds

This is where our work on OriON becomes more specific.

OriON should not attempt to determine metaphysical authorship.

It should preserve evidence about what participants claim to have done and what the recorded process can substantiate.

That distinction protects the system from becoming absurd.

OriON cannot know:

Farah truly invented this idea independently at 2:43 a.m.

But it may be able to establish:

02:43 — private research note created
02:44 — note hash committed
03:07 — first implementation sketch created
09:12 — AI consultation begins

That is useful.

It establishes that the recorded idea existed before the recorded AI consultation.

It does not claim omniscience.

Likewise:

AI suggestion generated
        ↓
human rejection recorded
        ↓
alternative generated
        ↓
human edits manually
        ↓
final revision signed

OriON can preserve the chain.

It does not need to infer the soul behind it.

Provenance should preserve rejection

There is a subtle problem in current AI workflows.

We tend to record what survives.

But human agency is often most visible in what does not survive.

The model proposes a title.

Rejected.

It proposes a paragraph.

Rejected.

It generates ten images.

Nine disappear.

It suggests an architecture.

The engineer refuses it because the privacy model is wrong.

It recommends a source.

The researcher rejects it as unreliable.

Those negative decisions matter.

A finished artifact may contain only 10% of what the machine produced.

If provenance records only the accepted output, the process can look more machine-directed than it really was.

So an agency-aware provenance system might preserve at least coarse evidence of rejection:

12 candidate outputs generated
3 inspected
1 used as reference
0 inserted directly

or:

AI architecture proposal A
STATUS: rejected by human reviewer

Reason category:
privacy / governance

Not every rejected sentence needs to be archived forever.

That would become surveillance again.

But rejection itself is a form of agency.

A provenance system interested in human leadership should know that humans were not merely approving whatever appeared.

Selection is creative work

This becomes even more important in visual generation.

Suppose an artist generates 150 images.

Someone says:

The AI made the art.

But selection itself can contain meaningful artistic judgment.

Which composition is useful?

Which lighting communicates the intended emotion?

Which anatomy is wrong?

Which visual motifs belong to the character?

What must be regenerated?

What must be painted manually?

Which reference has been over-weighted?

Which image violates continuity?

The generator performs synthesis.

The artist may perform direction, evaluation and authorship through selection and transformation.

That does not mean every act of clicking Generate deserves the title artist.

It means the binary is insufficient.

A useful provenance record might say:

VISUAL CONCEPT
human

REFERENCE SELECTION
human

GENERATION
image model

CURATION
human

CHARACTER CONTINUITY CORRECTION
human-directed regeneration

FINAL COMPOSITION
human

FINAL RETOUCH
human + conventional software

Now the reader can decide how they feel about the work with substantially better information.

Provenance does not have to settle the cultural argument.

It can make the argument more honest.

Agency maps also protect machine contribution from being erased

There is another side to this.

Human-led AI should not become a rhetorical trick for pretending the machine contributed nothing.

If an AI system generated an important element, provenance should say so.

If a model performed substantial analysis, say so.

If generated code constitutes most of an implementation, say so.

If an image contains generated components, say so.

If an AI system found the critical source that changed the argument, that contribution may deserve recording too.

Human authority does not require machine invisibility.

In fact, hiding machine contribution weakens the case for human-led systems because it makes the entire relationship look dishonest.

The cleaner model is:

HUMAN AUTHORITY
      ≠
HUMAN EXCLUSIVITY

and:

AI CONTRIBUTION
      ≠
AI OWNERSHIP OF PURPOSE

Those distinctions allow both sides of the process to remain legible.

Agency also includes sources

There is another participant we routinely erase:

the people whose work entered the process before either human or AI touched it.

A provenance map should not begin with:

Farah + AI

if the actual process began with:

research paper by A
photograph by B
dataset by C
open-source library by D
interview with E
      ↓
Farah + AI
      ↓
new artifact

This matters enormously with generative systems because their apparent fluency can make intellectual lineage disappear.

OriON should therefore distinguish:

CREATOR
TOOL
SOURCE
REFERENCE
DATASET
COLLABORATOR
DERIVED WORK

Rather than throwing everything into a generic category called input.

If a human writes an essay based on somebody else’s research, that research deserves provenance.

If AI retrieves it, the AI is not suddenly the origin of the knowledge.

If a generated image uses a photographer’s work as an explicit reference, that relationship should be recordable.

This is where provenance becomes an ethical practice rather than merely technical metadata.

It keeps the lineage visible.

Final authority deserves its own field

Perhaps the most important field in an agency map is also the simplest:

Who approved this?

Not who touched it most.

Not who produced the most tokens.

Not who consumed the most compute.

Who was authorised to say:

This is the thing we are releasing.

For low-stakes creative work, that may simply be the author.

For a company:

a product owner.

For a scientific paper:

the named researchers.

For a medical system:

perhaps several qualified professionals.

For an Amānah-style care system:

the answer should almost certainly involve governed human authority appropriate to the decision.

For an autonomous-looking AI workflow, the final authority field prevents a familiar rhetorical escape:

The AI decided.

Machines do not get to become convenient places for institutions to hide responsibility.

If an organisation deploys a system that makes consequential decisions, provenance should make it possible to ask:

Who gave this system the authority to act?

That person or institution remains part of the agency map whether or not they personally clicked the button.

Agency maps can expose fake autonomy

This is one reason I am increasingly suspicious of the language of fully autonomous AI.

Many systems advertised as autonomous are surrounded by invisible human decisions.

A human selected the model.

A human granted the credentials.

A human defined the objective.

A human decided what tools the system can use.

A human established the stopping conditions.

A human approved deployment.

A company set the economic incentives.

An engineer wrote the evaluation criteria.

Then the system acts for thirty minutes and the headline becomes:

AI independently accomplishes task.

No.

The system may have operated autonomously inside a human-authored envelope.

That distinction matters.

An agency map would show something like:

GOAL
human-defined

MODEL
human-selected

TOOLS
human-authorised

EXECUTION
machine-autonomous within scope

ESCALATION RULES
human-defined

FINAL ACCEPTANCE
human

ACCOUNTABILITY
organisation

Suddenly autonomous stops meaning uncaused.

Which is healthier for everybody.

Provenance is not a morality score

There is another temptation to avoid.

A rich provenance record does not automatically make a work ethical.

A perfectly documented exploit is still an exploit.

A fully traceable propaganda image can still be propaganda.

A human-authored lie remains a lie.

An AI-generated medical summary can be excellent.

Provenance tells us something about origin, process and agency.

It does not tell us whether the final thing is wise, beautiful, lawful, fair or true.

That judgment still belongs elsewhere.

This boundary matters because otherwise provenance systems will overpromise.

OriON should not become:

VERIFIED GOOD CONTENT™.

It should become:

Here is the best available account of how this artifact came to exist.

That is already enormously valuable.

The point is legibility

Much of the anxiety around generative AI comes from losing the ability to see the human.

A polished paragraph appears instantly.

A beautiful image appears instantly.

A working program appears instantly.

The causal chain collapses into a button.

And when the process becomes invisible, people begin filling the gap with mythology.

The machine understood.

The machine imagined.

The machine decided.

The human did nothing.

Or, from the opposite camp:

The human did everything because the human typed the prompt.

Both stories are usually too simple.

An agency map restores the missing middle.

PURPOSE
   ↓
SOURCES
   ↓
HUMAN DIRECTION
   ↓
MACHINE ACTION
   ↓
HUMAN EVALUATION
   ↓
REVISION
   ↓
MORE MACHINE ACTION
   ↓
MORE HUMAN JUDGMENT
   ↓
FINAL APPROVAL

That is what collaboration actually looks like.

Not purity.

Not replacement.

A sequence of delegated actions and retained authorities.

OriON should make the causal story inspectable

This gives us a clearer direction for provenance systems like OriON.

Do not merely ask:

What software touched this file?

Ask:

What role did each actor play?

Do not merely record:

AI used.

Record:

AI generated alternatives; human selected one; human rewrote it; human approved publication.

Do not merely prove that a human was present.

Record where human decision-making actually entered the process.

Do not erase sources.

Do not erase machine contribution.

Do not erase rejection.

Do not erase responsibility.

And perhaps most importantly:

do not compress a complicated creative relationship into a badge designed to win an argument.

The purpose of provenance is not to declare a winner between human and machine.

It is to make the relationship legible enough that people can decide what the contribution means.

That is why provenance is really an agency map.

It answers a more useful question than human or AI?

It asks:

Who originated?
Who contributed?
Who transformed?
Who chose?
Who authorised?
And who stands behind the result?

Once those answers are visible, we can have a much more serious conversation about authorship.

---

OriON design note

This essay suggests a provenance vocabulary worth prototyping separately from the article itself:

ACTOR
  human | model | software | organisation | source

ROLE
  originator
  source
  researcher
  generator
  executor
  editor
  reviewer
  curator
  approver
  publisher

ACTION
  created
  retrieved
  generated
  transformed
  rejected
  selected
  revised
  verified
  approved

AUTHORITY
  advisory
  delegated
  bounded-autonomous
  approval-required
  final

EVIDENCE
  declaration
  event log
  artifact hash
  signature
  timestamp
  attestation

That gives us the beginnings of an OriON agency graph rather than just a traditional linear provenance chain.