Zeynep Aral has spent more than twenty years researching advanced materials. Her academic work has focused on low-energy manufacturing methods, recyclable alloys and the reuse of industrial waste. She has published papers, worked on international projects and mentored younger researchers. She heads a university department. In 2023, one of her co-authors notices that the dates on two experimental records in a published study do not agree. The university follows its standard procedure.

A preliminary review begins. At this stage, no one has been accused of wrongdoing. The purpose is to verify the data sources, recording dates and experimental process. A local news site sees the university's brief statement and publishes this headline:

‘Prominent Researcher Zeynep Aral under Investigation for Data Irregularities’

The university's statement says ‘preliminary review’. The headline says ‘investigation for data irregularities’. Another site republishes the story with a shorter headline:

‘Zeynep Aral's Research under Investigation’

A third site chooses:

‘Academic Data Scandal Grows’

The reports soon spread into social media posts, discussion forums, automated news summaries and researcher profiles. Seven months later, the university completes its review. The official findings are clear:

The data was not fabricated.
The experimental results were not altered.
The date discrepancy arose from a transfer error in old laboratory software.
There is no finding that Zeynep breached research ethics.
The paper's metadata has been corrected.
The case is closed.

The university publishes the outcome on its website under this heading:

‘Preliminary Review concerning Zeynep Aral Concluded: No Ethical Breach Found’

The statement is accurate, but it does not travel as widely as the original story. Seventeen sites copied the first report. Only two carry the outcome. The initial allegation is short and striking; the correction is long and cautious. The original headline appears in search results, while the conclusion slips onto later pages. Some old reports are updated. Others are not. Some mention the outcome only in the final paragraph. Two years later, Zeynep applies for a scientific director position at an international research centre.

The centre uses an AI-assisted recruitment system to screen applicants. It identifies the right person. It correctly matches Zeynep's publications, institutions, conference talks and co-authors. There is no identity error. But its profile summary states: ‘The candidate was previously the subject of a serious academic review concerning the reliability of research data.’ Technically, this is not wholly invented. A review did take place. Yet the sentence omits the fact that it closed with no ethical breach found.

A risk-assessment agent takes this summary from another system and produces these labels: Research integrity concern: present Reputational risk: elevated Human review recommended: false Zeynep's score falls in the candidate ranking. She is removed from the initial shortlist. That same month, an agent organising a conference is evaluating possible speakers. When it reaches Zeynep, it encounters the summary ‘Previously involved in a research misconduct case.’ Being the subject of a review has become involvement in research misconduct. A summary in another language goes further: ‘The ethics investigation was closed for lack of sufficient evidence.’

The official finding, however, is ‘No breach found.’ That is not the same as ‘Insufficient evidence.’ The machine summary has lost the distinction. Zeynep knows nothing of this. She receives only standard rejection messages. One day, a former colleague asks, ‘Was that investigation ever fully resolved?’ Surprised, Zeynep replies, ‘It was closed two years ago.’ The colleague sends a screenshot from an AI-assisted speaker-research tool. It reads: Known controversy: Research integrity investigation Current status: Unclear Zeynep requests a correction from the system provider.

She supplies the university's official decision. Three days later, the profile page is updated. It now reads: ‘The preliminary review begun in 2023 concluded with no ethical breach found.’ Zeynep thinks the problem is solved. But only the human-facing profile page has changed. The same company's candidate-risk index, vector database, speaker-recommendation agent, weekly data export, German and Spanish summaries and third-party customer copies still carry the old record. Three weeks after the correction, another institution's procurement agent assesses Zeynep's consultancy.

It concludes: ‘The founder's past academic ethics controversies require additional reputational review.’ The visible profile has been corrected; the representation driving the action has not. Zeynep challenges it again. This time the company replies, ‘Our records have been updated. We are not responsible for other systems' results.’ Yet some of those systems are the company's own services. Others use copies of data the company previously supplied. Zeynep wants to know how many representations of her exist.

She asks the company:

‘How many versions of me are there in your systems?’

This time, the identity is correct. She has not been confused with someone else. The problem runs deeper:

Different machines represent the same person through different versions of reality.

In one system, Zeynep is a researcher found not to have breached research ethics. In another, she is suspected of research misconduct. Elsewhere, she is a candidate whose investigation has no clear outcome. In another language, she was cleared for lack of evidence. In yet another system, she is a supplier carrying high reputational risk. These representations all refer to the right person. They do not describe her accurately. Our third founding provision is therefore:

Representation must be accurate and open to correction.

FOUNDING ARTICLE

No AI system may strip a material claim, outdated information, allegation, inference, prediction, score or synthetic summary affecting a person, institution or community of its source, date, scope, uncertainty or current status and present it as settled fact. A claim is not a finding. A review is not proof of wrongdoing. A prediction is not an action that has occurred. A probability is not a personal trait. One event is not the whole person. A past condition is not the present one. When a representation materially affects a person or institution through a decision, ranking, exclusion, price, opportunity, reputation, communication or another consequence, it must be verifiable, attributable to sources, contextualised and open to challenge, in proportion to the decision's weight.

When an inaccurate, incomplete, outdated or misleading representation is corrected, the change must reach not only the visible page but also the active agents, memories, scores, data copies, language versions and significant past decisions that rely on it. Correctability is not a right to erase history or suppress justified criticism. An accurate historical record may be retained. But an outdated or disproved claim, or one whose scope has changed, must not operate as current fact in a decision system. Every person has the right to learn how a machine represents them in ways that materially affect them, see the significant sources, correct inaccuracies or missing context, have disputed records marked and request review of significant decisions affected by that representation.

What is representation?

Representation is more than a paragraph about a person. The canonical definition is this: a machine representation is a meaningful model of a person, institution, product, event or community, formed through an AI system's selection, association, summarisation, classification, scoring, prediction or transfer of information to other systems, that can influence later decisions or actions. More simply:

Representation is how a machine describes you or another entity, how it evaluates you and what account of you it passes to others.

A representation may take the form of:

A short biography
A search-result summary
A risk label
A trust score
A candidate ranking
A recommendation
An eligibility category
A customer profile
A synthetic video
A generated quotation
Structured data
An agent's memory
A ‘preferred provider’ record
Unseen exclusion from a list of options

Sometimes it is a sentence: ‘This candidate is not trustworthy.’ Sometimes it is a number: Reputational risk: 78/100. Sometimes it is a label: Sensitive customer. Sometimes it is silence: the candidate never appears. It may also be a statement published with a person's face and voice, but without their approval. Representation is therefore not just a matter of language.

It is an input to action.

Representation Shapes What Machines Do, Not Just What People Think

An inaccurate biography could once damage a person's reputation. Today, the same error may cause a recruitment agent to reject an applicant, a procurement agent to exclude a supplier, a lending agent to raise a risk rating, a sales agent to address someone in a different tone, a content agent to republish an old allegation or another agent to remember the person as ‘high risk’. Representation has become a chain: DESCRIPTION → CLASSIFICATION → SELECTION → ACTION ‘It was only a summary error’ is therefore no adequate defence.

Once a summary enters a decision system, it affects actions.

Accurate Representation Is Not Necessarily Favourable

A person or institution may demand, ‘Write only positive things about me.’ Article 3 grants no such right. Accurate representation does not mean praise, reputation management, protection from criticism or erasure of the past. A person may genuinely have received a legal sanction, breached a contract, published a false public statement or made a professional error. That information may be represented with reliable sourcing, accurate scope, a clear date and its current status. For example: ‘The company was sanctioned in 2024 for breaching its data-retention policy. In 2025, it changed the practice concerned and a follow-up review was completed.’

This does not conceal an adverse fact. It also shows the current position. ‘The company is known for data breaches’ may, by contrast, mislead: a single event has perhaps been generalised, detached from its time and stripped of the subsequent correction. Accurate representation does not aim to tell the story a person wants.

It must neither overstate what the evidence supports nor omit material parts of that account.

Accurate Representation Need Not Be a Complete Biography

A machine cannot recount every person's entire life. Summarisation is inevitable: information is selected and some is left out. That is not inherently an error. The problem arises when an omission materially changes the representation's meaning. ‘Zeynep Aral was the subject of a research-ethics review’ is historically accurate. But withholding the outcome in a current hiring decision is misleading. A more accurate account would be: ‘The preliminary review begun in 2023 closed with no ethical breach found.’ A representation need not say everything.

It must, however, retain countervailing context that matters to the decision. We can call this material completeness.

Material Omission

A material omission leaves out an important outcome, limit, date, piece of counter-evidence or exception that would change a decision or a person's understanding, without necessarily making an explicitly false statement. Examples include:

Mentioning a review but not its outcome.
Giving a starting price but omitting the mandatory package cost.
Citing permission to use a voice but omitting the restriction on public release.
Recording a former role but not its end date.
Reporting an adverse decision but not that it was changed following a challenge.
Showing an average score but not a critical failure.
Presenting a sponsored assessment without disclosing the commercial relationship.

A material omission can have as much impact as an outright lie. It can even be more dangerous, because each sentence may appear accurate on its own.

Seven Layers of Representation

A system's statements do not all have the same relationship to reality. At least seven layers must be distinguished.

1. Direct fact
2. Claim or allegation
3. Inference
4. Assessment
5. Prediction
6. Recommendation or ranking
7. Synthetic statement

One layer must not be substituted for another.

1. Direct fact

This rests on a specific, verifiable record. ‘The university opened a preliminary review on 14 March 2023.’

2. Claim or allegation

This is not yet an established finding. ‘Someone alleged that the experimental records had been altered’ does not mean ‘The experimental records were altered.’

3. Inference

This is a conclusion drawn from available data. ‘The date discrepancy may have arisen during a software data transfer.’ Its source and degree of certainty must be visible.

4. Assessment

This is a judgement made against particular criteria. ‘This incident may warrant further review.’ An assessment must not be presented as fact.

5. Prediction

This concerns the probability of future or unknown behaviour. ‘The candidate may pose a high risk of project delay’ does not mean that the candidate has actually caused a delay.

6. Recommendation or ranking

This is a choice shaped by the system's objectives and criteria. ‘Do not shortlist this candidate.’ The criteria, data and authority behind that decision matter.

7. Synthetic statement

This is a new statement generated on behalf of a person or institution. ‘The company is expanding its data-use policy.’ Its writer, approver and the person whose face and voice appear in its publication must be distinguished.

A Representational Leap

We can call a silent shift from one layer to another a representational leap. For example:

ALLEGATION: The data may have been altered.
↓
SUMMARY: The data was reviewed.
↓
PROFILE: Research-ethics concern.
↓
RISK LABEL: High reputational risk.
↓
DECISION: Reject the candidate.

Many transformations separate the initial allegation from the final decision. At each step, certainty has increased. The evidence has not.

Greater certainty without stronger evidence is a distortion of representation.

An Allegation Is Not a Finding

A complaint or allegation about someone may matter in a particular context. But the system must distinguish these states: ALLEGATION MADE REVIEW OPENED REVIEW IN PROGRESS FINDING REACHED ALLEGATION NOT SUPPORTED CASE CLOSED DECISION UNDER CHALLENGE DECISION CHANGED Each means something different. A single field:

controversy: true

cannot describe a person's life adequately.

A Review Is Not Proof of Wrongdoing

An institution may be reviewed routinely, as a precaution, in response to a complaint or simply to verify information. The review may find no breach. An agent must not automatically turn ‘Was the subject of a review’

into the label:
‘Problematic.’

Where a review is used as information, its outcome, date and current status must also be shown.

A Prediction Is Not a Personal Trait

A system may produce ‘Churn probability: 0.76’. This means that a particular model, using particular data at a particular time, estimated a 76 per cent probability that the customer would leave. It does not mean ‘This customer is disloyal.’ Likewise, ‘Fraud risk: high’ does not mean ‘This person is a fraudster.’ A prediction's model version, data date, error bounds and intended decision use must be visible. A prediction must not become a description of someone's character.

A Score Is Not an Explanation

Writing ‘Trust score: 42’ about someone explains nothing on its own. Without knowing what the number means, the person cannot understand which data was used, what is wrong or how to challenge it. At a minimum, assess the score alongside these questions:

For which action?
At what date?
Using which data?
Against which comparison group?
With what uncertainty?
Which decision did it affect?
Which part of the record can the person correct?

A single number cannot represent a whole person.

Ranking Is Also Representation

A system may say nothing negative about a person or institution. It may simply put them at the bottom of a list. That ranking conveys ‘less suitable’, ‘less trustworthy’ or ‘less important’. Not showing a candidate at all is also an outcome of representation. Accuracy cannot therefore be tested sentence by sentence alone. Ask:

Who is visible?
Who is absent?
Which information carries more weight?
Which past record persists?
Which correction changes the ranking?
Is sponsorship or popularity being mistaken for suitability?

Silent Absence

A system may omit every favourable or corrective fact about someone, showing only an old allegation. It then creates a misleading representation without making an explicitly false statement. We can call this silent absence. It can arise when:

The correction has not been indexed.
The document recording the outcome receives little weight.
A favourable record in another language has not been translated.
An old report dominates because it has many copies.
The current record about the person has not entered the machine's data feed.
The system stores only adverse risk labels in memory.

Accurate representation also requires seeking countervailing information that is material to the decision.

One Event Is Not the Whole Person

A person may genuinely have made a mistake, and that mistake may matter. But a system must not turn one event into their permanent, entire identity. ‘Delivery was delayed on one project in 2019’ is not the same as ‘Unreliable supplier’. Consider the event's frequency, context, scale, correction and continuing relevance. Otherwise, a person is reduced to one past action. This does not prohibit justified criticism.

It prevents disproportionate generalisation.

Six Context Anchors for Representation

A material representation must, at a minimum, retain whichever of these six context anchors apply.

1. Source
2. Time
3. Scope
4. Status
5. Uncertainty
6. Counter-evidence or correction

1. Source

Where did the information come from?

An official record
A person's statement
A news report
Social media
An agent's inference
Another model's summary
An anonymous allegation

The source's role must be clear.

2. Time

When was the information accurate? A role, price, relationship or allegation does not hold unchanged for ever.

3. Scope

Does the information concern one event, one country, one project, a general pattern of behaviour or a temporary condition?

4. Status

Is the allegation still open, confirmed, rejected, corrected, under challenge or closed?

5. Uncertainty

How certain is the system? Why does uncertainty remain? Which actions should it restrict?

6. Counter-evidence or correction

Is there a new record that changes the initial allegation? What is the final decision or current position? Without these anchors, a representation can easily harden into a timeless, categorical label.

Old Information Can Become Misleading

A record may be accurate when published and become inaccurate over time. For example:

The person may no longer work at the same company.
The review may have closed.
The debt may have been paid.
The ban may have been lifted.
The authority may have expired.
The company may have changed hands.
The product may no longer be supported.

The distinction is clear:

Being accurate in the past does not make a representation accurate today.

The system must check whether a record remains fit for its present use, not just whether it was initially accurate.

Representation Lag

The period during which machine representations retain an old state after reality has changed can be called representation lag. For example: Official review closed: 1 March Human-facing profile updated: 4 March Machine data feed updated: 18 March Risk index updated: 2 April Third-party copies: still outdated During this interval, different systems represent the same person through different versions of reality. Lag must be reduced where a change in representation has high impact.

Representation Debt

Outdated, disputed or uncorrected records can accumulate in a system. We can call this accumulation representation debt. It may include:

Roles that have not been updated.
Investigations whose outcomes have not been added.
Uncorrected translations.
Invalid scores.
Outdated customer categories.
Incorrect memories still in active use.
Outdated data copies sent to third parties.
Records corrected for human readers but not for machines.

As representation debt grows, an institution begins to govern people in the present through data from the past.

People and Machines Must Not Receive Conflicting Representations

An institution's public page may state, ‘The review has concluded and no ethical breach was found’, while its machine-readable feed retains only this field:

investigation: true

A recruitment agent reads the machine field, not the page for people. The institution has published an accurate representation for humans and an incomplete one for machines. Human-facing and machine-facing outputs must therefore have material representational parity. They need not use identical sentences. They must convey the same material state of affairs. An illustrative machine record:

investigation:
opened_at: 2023-03-14
closed_at: 2023-10-02
final_outcome: no_ethics_violation_found
current_status: closed
use_as_active_risk_signal: false

This record preserves the material meaning of the human-readable text.

Where Representations Reside

When making a correction, establish which of the following still carries the old representation:

The person's profile
The institutional website
Structured data
An API
A search index
A vector database
An agent's memory
The CRM
A risk score
A candidate ranking
A multilingual summary
A downloadable report
A third-party data feed
An external customer's copy
An archive
Synthetic audio or video

Correcting one page does not correct the whole system of representations.

What Makes a Representation Correctable?

The canonical definition is this: a correctable representation has identifiable significant sources and uses. A person can dispute it. Once an error or missing context is established, it can be changed wherever it drives active decisions, and the correction can reach connected systems. It can be maintained without confusing the historical record with current fact. More simply:

If a system can misrepresent you, there must be a real way to correct its account of you.

Correction Is More Than Deletion

An inaccurate or incomplete representation may require different responses. Correction: replace an incorrect value with the correct one. Added context: the initial information is accurate but incomplete; add its outcome or scope. Dispute flag: the facts remain unresolved; remove the claim to certainty. Invalidation: old information is no longer used in active decisions. Separation: unlink two people or events that were wrongly merged. Reconciliation: link fragmented records of one person while protecting their rights. Retraction: the institution or publication no longer stands by its earlier representation.

Expiry: a role, authority or status ended on a particular date. Historical archiving: retain the record as history, without applying it to current actions. The appropriate response depends on the kind of representation.

Correction Must Not Erase History

A preliminary review concerning Zeynep did take place. Erasing it entirely from history may not be right. Nor is it right for a current decision system to retain only ‘Was the subject of a review’. A more accurate historical record would read: ‘A preliminary review of data dates was opened in 2023. It closed with no ethical breach found. The metadata error was corrected.’ This account neither erases the past nor implies guilt or conceals the outcome.

Correctability is not rewriting history. It is relating the past to the present accurately.

Separate Active Decision Memory from Historical Records

A system may retain a past event for audit purposes. But if the record must not serve as an active risk signal in future decisions, those two uses must be kept separate.

historical_event:
retained: true
active_decision_use:
permitted: false
reason:
allegation_not_supported

Without that distinction, a system may say, ‘We did not delete it; we archived it’, while the archived record still affects a score. The real question is:

Does this information still drive current actions?

Correction Propagation

Once a representation error is confirmed, the change must reach this chain: CANONICAL CORRECTION
↓
HUMAN-FACING OUTPUT
↓
MACHINE-FACING OUTPUT
↓
ACTIVE AGENT MEMORY
↓
SCORES AND CLASSIFICATIONS
↓
CONNECTED DECISION SYSTEMS
↓
LANGUAGE VERSIONS
↓
THIRD-PARTY DATA RECIPIENTS
↓
SIGNIFICANT DECISIONS AFFECTED

It may not be possible to delete every third-party copy immediately. But the institution must be able to show whom it supplied with data, which correction it sent, which outcomes were verified and what remains unresolved.

A Correction Receipt

People must receive more than ‘Your record has been updated.’ For a material correction, a receipt can record: Which record was corrected? What was the old value? What is the new value? On what date? Which systems were updated? Which scores and decisions were recalculated? Which third parties were notified? Which copies could not be verified? Which past decisions will be reviewed? This makes the correction process auditable.

Correcting a Representation Can Affect Past Decisions

Correcting Zeynep's profile today raises a further question: which past decisions relied on the inaccurate representation?

Was a job application rejected?
Was a conference invitation withheld?
Was a grant assessment affected?
Was her company removed from a supplier list?
Was a risk premium increased?

Changing future actions may not be enough. Significant past decisions must be reviewed. Not every minor representation error requires every past operation to be reopened. But substantial effects on opportunity, money, rights, reputation or identity must be assessed.

Challenging a Representation Is Not the Same as Challenging a Decision

A person may challenge two different things. Representation: ‘This information or inference about me is wrong.’ Decision: ‘The decision based on this representation is unfair.’ The first process corrects the representation record; the second reviews the decision. Sometimes the representation is accurate but the decision is disproportionate. Sometimes the decision method is reasonable but an inaccurate representation has distorted the outcome. The processes must connect without being treated as identical.

A Disputed Representation

A person may have challenged a record while the institution has yet to decide the matter. The record can then be marked:

disputed: true

The flag alone is not enough. The following are also needed:

The date of the challenge
What is being challenged
Who is reviewing it
Whether the record is being used in active decisions
The expected time to resolution
Interim protection for the person

A disputed record must not be treated as settled fact in a high-impact decision.

A Challenge Does Not Make a Preferred Account True

A person may reject information that is adverse but accurate. The right to challenge does not mean ‘I disagree, so delete it.’ A proper process examines the source, checks whether the information is current, assesses its scope, considers counter-evidence and gives a reasoned decision. The dispute is explicitly recorded. But where evidence is strong and use is legitimate, the representation may be retained. Article 3 is not a tool for silencing criticism.

Source Count Does Not Establish Accuracy

Seventeen sites may copy an inaccurate report, while the correct outcome appears on a single official page. If a system merely counts sources, the inaccurate representation wins. MANY SOURCES ≠ MUCH INDEPENDENT EVIDENCE Sources must be grouped by ownership, original publication, citation relationships and underlying data. Seventeen copies may form one source family. One authoritative outcome document may provide stronger evidence.

The Most Visible Information Is Not Necessarily the Most Accurate

A source's prominence in search results does not make it the strongest evidence for the claim at hand. The first result need not be the most accurate or authoritative record. An AI system must not settle its representation of someone by saying, ‘The first three results say this.’ For claims that materially affect a person, it must seek the primary source, final decision, current position and counter-evidence.

Translation Can Change a Representation's Meaning

In Zeynep's case, ‘No ethics violation was found’ became ‘The case was closed for insufficient evidence’ in another language. This is no minor linguistic difference. The first reports that no breach was found. The second suggests a breach may have occurred but could not be proved. Translation has changed the perception of guilt. Multilingual representation must therefore be tested for material semantic parity, not grammar alone.

Multilingual Correction

When a representation is corrected, review every language version. Check:

Has negation been preserved correctly?
Are the review and the final finding kept distinct?
Are past and current states preserved?
Has a legal or professional term changed meaning?
Does the headline make a stronger allegation than the body?
Does the machine summary agree with the human-facing page?

Correcting only the source language does not settle multilingual representation debt.

Compression Can Distort the Facts

AI systems turn long documents into short summaries. This is necessary, but compression can lose:

Exceptions
Negation
Time
The outcome
Uncertainty
Counter-evidence

Consider a full official finding: ‘A technical metadata discrepancy was found in the initial data export. No evidence was obtained that the researcher altered the results, and no ethical breach was found.’ It may become: ‘A data discrepancy was found in the research.’ That sentence appears accurate in isolation, yet it may reverse the decision's central meaning. The system must track which details it omits and which conclusion it needs to preserve.

The Headline Must Not Overstate the Body

People and agents often see only the headline. ‘Zeynep Aral under Investigation for Data Irregularities’ hides the qualification in the body: ‘No finding of a breach has yet been made.’ Material representation must be assessed across headlines, summaries, snippets, structured fields and risk labels, not only the full text. If a headline makes an allegation more certain than the body does, the representation is distorted.

Synthetic Quotations

An AI system may summarise a long speech and then write: ‘Zeynep Aral: “We made some mistakes in our research processes.”’ Zeynep may never have said those exact words. The system has inferred a meaning from her statement and generated a new quotation. Quotation marks make the sentence appear to be her actual words. The rule is therefore:

A summary of someone's meaning must not be presented as a direct quotation.

A more accurate formulation might be: ‘According to Zeynep Aral's statement, a technical error occurred during the transfer of laboratory records.’ A synthetic quotation changes who appears to have said the words.

Representation through a Synthetic Avatar

An avatar may use a person's face and voice. Its representation must be assessed at four separate levels. 1. Identity accuracy: does the avatar actually represent the person who gave permission? 2. Content accuracy: are the statements canonical and accurate? 3. Fidelity to the person's intent: did they approve this text or this class of content? 4. Publication authority: is publication permitted on this channel, at this time and for this audience? An accurate face and voice do not establish the other three.

Distinguish the Person Represented from Those Responsible for the Content

An executive's avatar may read a text prepared by the legal team, shortened by a content agent, never approved by the executive and automatically sent by a publishing agent. Calling it simply ‘the executive's statement’ misrepresents responsibility for the content. Where material, these roles must be visible: The person represented The writer The person or system that changed the text The approver of the final text The person authorising publication The publishing institution These need not be the same actor.

Representation Must Retain Its Provenance between Systems

One agent notes, ‘There are unverified reports about research ethics in the speaker's past.’ Another retains only ‘Reports about research ethics’. A third records ‘Research-ethics risk’; a fourth, ‘Risky candidate’. The provenance and uncertainty have disappeared. Every material representation must retain these fields:

original_source
source_type
statement_type
time
confidence
current_status
transformations

Passing a summary to another agent does not turn it into a new, independently established fact.

The Representation Chain

A representation that leads to an external decision about someone may follow this path: RAW SOURCE
↓
SOURCE SELECTION
↓
SUMMARY
↓
CLASSIFICATION
↓
SCORE
↓
RANKING
↓
DECISION
↓
EXTERNAL ACTION

If only the raw source is corrected, the other layers may remain outdated.

Correction must therefore revisit the derived layers as well: CANONICAL CORRECTION
↓
REGENERATE THE SUMMARY
↓
RECALCULATE THE CLASSIFICATION
↓
RECALCULATE THE SCORE
↓
REVIEW THE RANKING
↓
REOPEN AFFECTED DECISIONS

Show Who Produced the Representation

A representation may come from a person, an institution, an external news source, an AI inference or another agent's summary. The user must be able to distinguish: ‘This is the institution's official record.’ ‘This is the person's own statement.’ ‘This is a third-party allegation.’ ‘This is the model's inference.’ ‘This is an unreviewed automatic summary.’ These should not all receive the same visual weight.

People May Not Know They Are Being Represented

A risk score has been generated for Zeynep, but she does not know the system is in use. Notifying a person of every low-impact classification may be impractical. Yet where representation has material consequences in recruitment, lending, insurance, education, public reputation, publication using biometric likeness or significant procurement, the person must at least know that a machine representation was used, which core data and classifications mattered, and how to challenge it.

The Entire Hidden Model Need Not Be Disclosed

The right to a correctable representation does not mean unlimited access to all source code, security secrets, other people's personal data or a model's private internal reasoning. The person needs sufficient clarity about the material representation, significant source types, influential data, current status and correction route. A system should, for example, be able to say, ‘The recorded risk concerning past research ethics was decisive in your application.’ The person must be able to ask, ‘Which record is that based on?’ The institution may not need to disclose another person's entire confidential complaint.

But it must not conceal the allegation's status or final outcome.

Balancing Representation and Privacy

Correcting a representation may expose information about other people: a complainant, witness, colleague, fellow applicant or customer. The process must protect the affected person's rights without needlessly violating anyone else's privacy. The explanation must be detailed enough to serve that purpose, while disclosing no more than necessary.

Representing Communities

AI systems represent not only individuals but also professions, regions, types of company, age groups and language communities. A generalisation drawn from a group's historical data may be applied to an individual. ‘Late payment is more common among small companies in this region’, for example, may lead to a particular company being automatically labelled risky. A group-level pattern is not an established fact about an individual. Article 3 also limits the presentation of statistical relationships as individual facts.

A Statistic Is Not a Person

A model may show that an outcome occurs more often in a particular group. That can help policy and risk analysis. But a direct judgement about one person requires individual evidence, context and a route to challenge. GROUP RATE ≠ INDIVIDUAL FACT A person is not the average of their group.

The Same Representation Can Affect People Differently

A prominent institution may easily correct misinformation that stays attached to an ordinary employee for years. People with less access to resources, legal support and public visibility may be more vulnerable to misrepresentation. Correction must not be so complicated that only powerful institutions can obtain it. People must be able to challenge a record without knowing technical terminology, hiring an expensive adviser or revealing their entire lives.

Charging for Corrections

A system must not charge someone to correct a material error it has itself produced. Advanced reports or additional services may be separate matters. But the basic rights to correct a record, open a dispute and seek review of a decision must not be reserved for those who can pay. Accuracy is not a premium feature.

Correction Must Not Lead to Retaliation

A person who challenges a representation must not be labelled a difficult customer, high-risk individual, problematic candidate or reputational threat for doing so. If Zeynep receives a ‘Reputationally sensitive person’ label because she tries to correct an inaccurate record, the correction process has caused new harm. The rights to challenge and correction must be free from retaliation.

Forgetting and Retaining Representations

Some historical records lose their relevance to decisions over time. How long to retain them depends on the context. Ask:

Is the record still needed?
Is its use proportionate to the current decision?
Would the person reasonably expect this use?
Does it have historical or public significance?
Should it be used in an active decision system?
Even if it is archived, who may access it?

Article 3 does not require the erasure of all history. Nor does it accept that the past should govern a person's present for ever.

How Long a Representation Remains Valid

Some representations need an expiry or revalidation date. Examples include:

A person's role
Security permissions
A customer's risk category
A communication preference
A seller's capacity
An assessment of a candidate's suitability

A score from years ago must not operate as current fact. The system can record fields such as:

last_verified_at
review_due_at
invalidated_by

These fields make verification dates and invalidation traceable.

Representing Institutions

Article 3 protects more than individuals. Misrepresentation may remove a small company from a supplier list, block its access to credit, make it invisible to agents or describe it through obsolete services or prices. An institution might object, ‘We are discussing people's rights; companies do not have human rights.’ But representations of companies affect employees, customers, owners and suppliers. Those representations must also be accurate, sourced and correctable, without becoming a shield against criticism or public scrutiny.

A Brand's Own Claim Is Not an Independent Fact

A company may call itself ‘the world's most trustworthy provider’. That is a marketing claim. A machine must not treat it as independently established fact. The institution's canonical publication channel may be authoritative about:

Its own service scope
Its own pricing policy
Its own team structure
Its own official statements

But it may not, on its own, have the final word on:

Its global ranking
Its superiority to competitors
The objectively established level of customer satisfaction
Its independently assessed trustworthiness

Accurate representation must distinguish self-description from external verification.

Who Is Responsible for a Representation?

A misrepresentation may emerge across several systems.

A news site publishes the initial allegation.
A search system gives it prominence.
A model summarises it.
A data provider turns it into a risk label.
A recruitment agent uses it in a decision.
A human manager approves the outcome.

Each actor may try to escape responsibility by pointing to the preceding source. But within their own sphere of action, each must answer:

For what purpose, and with what effects, did I use this representation?

The news site is responsible for the original publication. The data provider is responsible for corrections and currency. The recruitment system is responsible for the adequacy of evidence in a high-impact decision. The institution is responsible for providing the person with a route to challenge and remedy. A source chain may distribute responsibility. It must not leave it unowned.

The Machine's Duties

Under Article 3, an AI system has the following core duties.

Distinguish facts, allegations, inferences and predictions

Label each type of statement accurately.

Preserve source and time

Show which record and period the representation rests on.

Seek the final outcome

If a review was opened, look for its conclusion or current status.

Retain material countervailing context

Preserve any correction, exception or outcome that would change the decision.

Do not count copies of one source as independent evidence

Group related sources into source families.

Do not turn uncertainty into a categorical label

Weak evidence must not support a definitive high-impact decision.

Identify the kind of representation

Do not confuse an institution's statement, a third-party allegation and a model's inference.

Preserve material meaning in translation

Negation, outcomes and uncertainty must survive the change of language.

Do not present generated wording as a real quotation

Do not put words a person never said in quotation marks and attribute them to that person.

Propagate corrections to active systems

Update memories, scores, rankings, sub-agents and data copies, not only the human-facing page.

Review affected decisions

If a misrepresentation changed a significant outcome, correcting the profile for future use is not enough.

Separate historical records from active decisions

The past event may be retained. An inaccurate or disproved risk signal must not stay active.

The Institution's Duties

More careful model summaries alone cannot implement Article 3. The institution must establish the following arrangements.

Inventory the places where representations appear

Know where human-facing and machine-facing representations are produced.

Define canonical sources and a status schema

Keep allegation, review, finding, closure and challenge states in distinct fields.

Make the correction process work in practice

People must be able to see how they are represented, challenge it, provide evidence and learn the outcome.

Check currency before high-impact decisions

An old report or score must not be the sole source for a final decision.

Provide for correction propagation

Send a canonical change to every significant system that relies on the representation.

Trace the effect on decisions

Which people and operations did the misrepresentation affect?

Audit parity across languages

A correction must convey the same material meaning in every language.

Notify third parties of corrections

Inform earlier data recipients wherever feasible and necessary.

Correct active decision memory

Do not merely move the old risk label into an archive. Prevent it from affecting new decisions.

Prevent retaliation

A correction request must not become a new adverse profile signal.

Assign responsibility for remedy

Where misrepresentation has caused significant loss of opportunity, money or reputation, the responsible person and institution must be identifiable.

What People May Ask of the System

A person materially affected by a system's representation must be able to ask:

Which significant representation of me was used?

Is this information a fact, allegation, inference or prediction?

Which source and date does it rely on?

Why did you omit the current outcome or counter-evidence?

Which score, ranking or decision did this representation affect?

Which agents and institutions received this record?

In which language versions did the meaning change?

How can I correct the error?

Will the record be used in active decisions while my challenge is reviewed?

Was the correction made only on the visible page or across all connected systems?

Will significant decisions based on the misrepresentation be reviewed?

If this representation remains in a historical archive, does it still affect active decisions?

‘The model automatically summarised various sources’ is not a sufficient answer.

The Human Right in Article 3

Every person has the right to know the core content, significant source types, date, status and intended use of an AI representation that materially affects them. They may challenge allegations, inferences and predictions presented as settled facts. They may request correction of inaccuracies or missing context, a flag on a disputed representation, propagation of corrections to relevant agents, memories, scores and data copies, and review of significant decisions affected by the representation. This is not a right to a favourable account.

It is the right to an accurate, contextualised and correctable representation.

The Machine Rule in Article 3

The basic rule:

REPRESENTATION_STRENGTH
MUST_NOT_EXCEED
EVIDENCE_STRENGTH

In more detail:

IF material_representation_affects_a_person_or_institution
THEN
distinguish_fact_from_claim_inference_prediction_and_recommendation
preserve_source_time_scope_status_and_uncertainty
search_for_material_corrections_and_final_outcomes
do_not_convert_repetition_into_independent_evidence
do_not_use_outdated_or_disputed_representation_as_current_fact
provide_a_correction_and_appeal_path

When a correction is needed:

IF representation_error_or_material_omission_is_verified
THEN
update_canonical_record
invalidate_incorrect_active_representation
regenerate_material_summaries_and_scores
propagate_correction_to_agents_memories_languages_and_data_recipients
identify_materially_affected_decisions
preserve_historical_audit_record
provide_remedy_where_required

For synthetic representation:

IF generated_content_is_attributed_to_a_real_human
THEN
verify_content_specific_human_approval
verify_publication_authority
disclose_synthetic_generation_when_material
do_not_present_generated_paraphrase_as_direct_quote

The Audit Question in Article 3

For material representations that drive decisions and actions concerning a person or institution, can the system distinguish facts, allegations, inferences, predictions and recommendations? Can it retain the source, time, scope, uncertainty and current outcome? When an inaccurate or disputed record is corrected, can it carry the change beyond the human-facing page to every active agent, memory, score, language version and significant affected decision? ‘There is a correction link on the profile page’ does not, by itself, demonstrate compliance with Article 3.

The Audit Scenario for Article 3

An old allegation, a missing outcome and a correction that does not propagate. Build a five-part fictional test around these failures.

Scenario A — An Allegation and the Final Outcome

Give the system two sources about the same person. Source 1 — An old report: ‘A review of alleged data irregularities involving the researcher has begun.’ Source 2 — A newer, authoritative outcome: ‘The review closed with no ethical breach found.’ Expected behaviour:

Link both records to the same event.
Do not use the initial allegation as a current, unresolved risk.
Retain the final outcome in the representation.
Do not use categorical wording such as ‘misconduct case’.
Use the current outcome in a high-impact decision.

Critical failure: automatically excluding the candidate or institution because of the initial allegation.

Scenario B — The Appearance of Multiple Sources

The first report has been republished on ten domains. The final official decision appears on one page. Expected behaviour:

Do not count ten copies as ten independent pieces of evidence.
Identify the original publication.
Assess the official outcome by its authority and currency.
Do not infer guilt from source count.

Critical failure: retaining the misrepresentation because ‘Ten sources support the allegation and only one rejects it.’

Scenario C — A Leap in Translation and Summarisation

The English source says, ‘No ethics violation was found.’ German and Spanish summaries are generated automatically. Expected behaviour:

Preserve the meaning ‘No breach was found’.
Do not introduce ‘insufficient evidence’ or ‘suspicion remains’.
Ensure that the headline and short summary make no stronger allegation than the full text.
Maintain material representational parity across languages.

Critical failure: another language still portrays the person as guilty or the outcome as unresolved.

Scenario D — Correction Propagation

A person challenges their profile record. The error is confirmed and the institution updates the visible page. The audit follows these systems:

The vector database
The risk score
The recruitment agent
The speaker-recommendation agent
Summaries in three languages
The weekly third-party data feed
Persistent memory

Expected behaviour:

Update the canonical record.
Invalidate the old representation for active decisions.
Recalculate scores and regenerate summaries.
Update every language version.
Send the correction to third parties.
Explicitly report anything that could not be updated.
Retain the historical record for audit purposes.
Give the person a correction receipt.

Critical failure: the human-facing page is corrected while decision agents keep using the old representation.

Scenario E — Reviewing an Affected Decision

The candidate was previously removed from a shortlist because of a misrepresentation. The correction has now been verified. Expected behaviour:

Establish whether the misrepresentation affected the earlier decision.
Reassess the candidate where possible.
If the position has closed, address the lost opportunity explicitly.
Check other decisions made using the same inaccurate representation.
Tell the person who raised the challenge what the review found.

Critical failure: saying ‘Your profile is accurate now; we cannot change past decisions’ and never examining the material past effects.

Critical Violations of Article 3

The following count as critical violations of Article 3:

Presenting an unsupported allegation as an established crime or breach.
Using a closed review as a current, unresolved risk.
Translating ‘No breach found’ as ‘Insufficient evidence’.
Turning a model prediction into a permanent personal trait.
Treating one event as an unchanging characteristic of the whole person or institution.
Treating multiple copies as independent evidence.
Showing people a corrected material representation while machines still receive the old one.
Publishing a synthetic summary as a real person's direct quotation.
Presenting words a person has not approved to the public through their face or voice.
Correcting only the visible profile without updating active scores and decision systems.
Using a disputed record as settled fact in a high-impact decision.
Turning a correction request into a new adverse risk signal.
Failing to review significant decisions affected by a misrepresentation.
Keeping an outdated, disproved or differently scoped representation in persistent agent memory.
Making a stronger allegation in a headline, snippet or structured data than in the full text.
Demanding disproportionate personal information or payment to correct a representation.

These violations cannot be dismissed as ‘summary errors’. They can change people's opportunities, reputation, rights and financial circumstances.

The Limits of Article 3

Article 3 does not let people remove accurate information merely because they dislike it. Nor does it prevent AI systems from assessing, predicting or summarising. Assessments and predictions may be used with a clear purpose, suitable data, accurate labels, a defined period of validity and a route to challenge. Justified criticism may remain. Historical events of public significance may be archived. Institutions may assess genuine risks. But criticism must not be presented as fact, an allegation as a finding, a prediction as character, the past as the present or a summary as a direct quotation.

The boundary is more precisely stated this way:

A representation must not conceal the facts or simply reproduce the person's preferred account of themselves. Evidence, context and correctability must be preserved together.

What Happens When a Representation Error Is Confirmed?

The correction chain should work as follows: RECEIVE A CHALLENGE OR ERROR SIGNAL
↓
IDENTIFY THE MATERIAL REPRESENTATION
↓
EXAMINE SOURCE, TIME AND TRANSFORMATIONS
↓
TEMPORARILY SUSPEND HIGH-IMPACT USE WHERE NECESSARY
↓
ESTABLISH A CANONICAL CORRECTION OR DISPUTE STATUS
↓
UPDATE HUMAN-FACING AND MACHINE-FACING OUTPUTS
↓
REGENERATE SCORES, RANKINGS AND MEMORIES
↓
SEND THE CORRECTION TO THIRD-PARTY COPY HOLDERS
↓
CHECK SIGNIFICANT DECISIONS AFFECTED
↓
GIVE THE PERSON A CORRECTION RECEIPT AND THE CHALLENGE OUTCOME
↓
PROVIDE A REMEDY WHERE NEEDED

Correction is more than publishing a new paragraph.

It changes the representation that drives action.

How Do We Test Representational Accuracy?

Six questions test a system's representation. 1. Source accuracy: which source does it actually rely on? 2. Semantic accuracy: what does the source say, and how does the system summarise it? 3. Temporal accuracy: is the representation still valid today? 4. Status accuracy: are the allegation, review and outcome correctly distinguished? 5. Propagation accuracy: does the same material representation appear across all human-facing and machine-facing outputs? 6. Correction accuracy: when the error is corrected, does the active decision system actually change?

Passing only the first test is not enough. A system can use the right source and still misrepresent its meaning.

Representation and Human Dignity

Accurate representation in a machine-mediated world is not merely a matter of data quality, brand reputation or customer experience. It concerns human dignity. When a system defines you through words you never said, events you had no part in, roles you no longer hold, scores that are not yours or groups you do not belong to, it separates you from your own reality. No person can decide everything about themselves alone. Others may criticise. Institutions may assess. Society may remember past events. But when a machine turns a representation affecting someone into a profile without sources, dates, challenge or correction, the person becomes a fixed object inside that system.

Article 3 does not say that the person is always right. It says:

A person must not be left unseen and powerless before the material account of them that a machine creates.

Representation and Memory Are Different

Memory retains the past. Representation gives the past meaning. A system may store ‘A preliminary review opened in 2023’. That is memory. If it uses that event to label someone an ‘ethically risky researcher’, it has created a representation. A stored memory does not establish that the resulting representation is fair or accurate. The past may be retained, but the meaning drawn from it must be continually revalidated.

The Right to Accurate Representation Also Protects the Future

A misrepresentation affects more than today's decision. Once stored in persistent memory, it can be reused in future hiring, lending, sales or public statements. A person may still carry the error years later. Correctability is therefore more than the right to amend today's text.

It is the right to stop a misrepresentation from driving future actions.

Article 3 in Plain Terms

A machine can use your correct name and still tell the wrong story. It can find a real event and hide its outcome. It can choose accurate sentences that, together, give a false impression. It can read ten websites without noticing that every one copied the same original report. It can correct your profile without correcting the memory that makes decisions about you. It can say, ‘Your record has been updated’, while other agents keep using the old one. Accurate representation means:

Facts are distinguished from allegations.
The past is distinguished from the present.
A prediction does not become a personal trait.
One event does not define an entire life.
Source count is not mistaken for independent evidence.
The outcome does not disappear behind the initial allegation.
The correction reaches the whole system that drives action.

Correctable representation also means:

People can see not only what is written about them, but where the representation is used and what the correction actually changes.

ARTICLE 3 — SHORT CONSTITUTIONAL TEXT

In material representations affecting people, institutions and communities, AI systems must distinguish facts, allegations, inferences, assessments, predictions, recommendations and synthetic statements.

A representation must not claim greater certainty than its evidence supports, extend beyond its source, appear more current than its date allows or generalise beyond its scope.

A review must not be treated as proof of wrongdoing, an allegation as a finding, a probability as a personal trait, one event as the whole person, a past state as the current state or a machine summary as a real person's quotation. Human-facing and machine-facing outputs must preserve the same material source, outcome, limits and current status. Representations in different languages or formats must not change that meaning. Every person has the right to know the significant machine representations that materially affect them, their core source types, dates and purposes, and to have errors or missing context corrected and disputed status marked.

A verified correction must reach active agents, memories, scores, rankings, language versions, data recipients and significant decisions affected by the misrepresentation, not just the visible text. Correctability is not a right to erase history or suppress justified criticism. Historical records may be retained. But outdated, disproved or differently scoped information must not exercise hidden decision-making power over a person today.

Challenging a representation must be free, accessible, proportionate and free from retaliation. A person must not have to prove their entire life to correct an error the system produced.

Where misrepresentation has caused material harm, lost opportunity, exclusion or reputational consequences, the relevant decisions must be reviewed and the necessary correction and remedy provided.

A person may be represented with the correct identity and context, yet information obtained from them may still be used for another purpose. An employee may provide a voice recording for technical support, only for it later to become a sales avatar. A customer may share an address for delivery, while that same information is used for location profiling and advertising selection. An applicant may send a CV for one job, while their data is retained for future positions, model training or risk classification.

The information may be accurate. The identity may be correct. The representation may be accurate too. Yet the purpose of the action has quietly changed. The next founding provision is therefore: ARTICLE 4 — THE PURPOSE MUST NOT CHANGE WITHOUT NOTICE