Different types of information are kept on the Internet in different authoritative sources. An organisation's name can be found in the registry, its current price in the sales record, its service limit in the contract, and its application status in the transaction system.
It may have reached the right domain name.
They may have examined the correct account, product or service.
But finding the right entity does not mean that a correct world model is set up about it.
On the Internet, reality doesn't stop in one place.
A price is on the website.
Another price is in the business directory.
The scope of service is described differently in the sales presentation.
Social media sharing refers to a temporary campaign.
Structured data carries a wider promise than the visible page.
English text is up-to-date; other languages are old.
An AI response draws on the text produced by another AI. New content repeats that answer.
Eventually, the same claim appears in many places.
But their appearance doesn't make them right.
The representation that an agent creates before action is born from the sources at hand. If resources are old, incomplete, or contradictory to each other, the agent can make a perfectly logical decision on the wrong reality.
The nine errors in this section are collected around the following problem:
The agent found the right entity, but is it really true what they know about them?
GBO-ERR-010 — Mistaking Repetition for Accuracy
Brief case
A purchasing agent is researching customer relations software for the company.
They see this statement about a particular product on many pages:
"All packages have unlimited users."
This statement:
on product comparison sites,
In several tech blogs,
In a social media post,
In two studies produced by artificial intelligence,
In an old forum response
is repeated.
The agent interprets the discovery of the same information in six separate sources as strong consensus. They decide that the product is suitable for the company's growing team and submits an annual subscription offer to approve.
After the purchase, it becomes clear that unlimited user feature is only available in the most expensive corporate package.
The middle package is also charged per user.
Six sources said the same thing.
But the root of all six is a false comparison table three years ago.
What appears correct on the surface
Seeing the same information repeated across apparently different sources naturally builds confidence.
People often use the following shortcut:
"If it's written in many places, it's probably true."
From an agent's point of view, it may seem even stronger again. Because the system sees that resources are formally different:
different domain names,
different author names,
different publication dates,
different sentence structures.
But seemingly diversity does not mean diversity of the origin of knowledge.
The actual failure
The breached gate is the Evidence Independence Gate.
The agent confused two things:
Number of repetitions Number of independent verifications
A claim can be found on a hundred different pages.
But if all pages are derived from the same old or wrong source, there is not a hundred proofs in the middle, but one chain of repetition.
It can increase visibility again.
It doesn't guarantee accuracy.
Potential harm
This error:
to choose the wrong product or supplier,
miscalculation of total cost,
to disrupt capacity and property expectation,
a long-term subscription obligation,
Moving company data to an improper system,
to eliminate more accurate but less visible options
It could be caused.
The real added risk is that false claim can become a new signal of trust through repetition.
Mistake strengthens itself through behaviour.
Detection signal
The following are warning signs:
A large number of sources use the same or very similar expression.
Resources are not linked to the primary product or organisation record.
Different sites have the same table, sentence, or spelling error.
Part of the resources are showing each other.
The current official record contradicts the repeated statement.
The agent does not examine the number of sources and examines the origin of the source.
The justification for "writing in many places" becomes evidence of decision alone.
Correct behaviour
Before treating repetition as corroboration, the agent must trace the information back to its origin.
They should ask:
Where was this claim first published?
Are resources really independent?
Current baseline record of the product or organisation for this type of information — What does the canonical record — say in the broad sense we use in this book?
Does information have a validity date?
Did the sources collect different data, or did they retell each other?
Do you need direct verification for the important decision?
The agent's statement may be:
"The unlimited user feature is repeated on many third-party pages; however, the independence of resources has not been verified and the current official price page shows package limits. It must verify written coverage directly from the provider prior to purchase."
Machine rule
The number of repetitions is not the number of independent proofs. Sources derived from the same information root do not count as independent verification.
Audit question
Does our system merely count how often a claim is repeated, or does it verify how many genuinely independent and current sources support those repetitions?
GBO-ERR-011 — Treating Outdated Information as Current Fact
Brief case
A travel agent searches for an accessible hotel in the city centre for the user.
In a hotel directory, the following information appears:
Wheelchair access
Free airport transfer
Restaurant until 24.00 p.m.
Free cancellation
$160 per night
The recording is detailed and shows the correct address of the hotel.
The agent also finds the hotel's official site. The official page, however, has been made simpler and does not explicitly list some features.
The agent prefers the more detailed business directory and makes reservations.
Upon arrival at the hotel:
Accessible access is closed due to renovation,
The airport transfer was removed two years ago,
The restaurant closes at 10 p.m.
The free cancellation belongs to an old campaign alone,
The current price is $240.
The directory record may have been true at one time.
But today is no longer true.
What appears correct on the surface
Outdated information rarely labels itself as outdated.
The page still works.
The company has the correct name and address.
Some information remains up-to-date while others are outdated.
The agent can assess the entire recording at a single level of trust.
In addition, more detailed resources may seem more reliable.
But the detail is not current.
The actual failure
The breached gate is the Temporal Validity Gate.
The accuracy of a changeable information for action depends on when it is as valid as its content.
These two sentences are not the same:
"This hotel offers free transfer."
"This hotel offers free transfer today."
A record may have been true in the past.
But it may no longer apply to current behaviour.
It must also be confirmed that information is still valid in a current-sensitive and high-impact decision; the absence of a date stamp alone is not evidence of wrongdoing.
Potential harm
Wrong booking or purchase
Failure to meet the user's need for accessibility
Unexpected cost
Loss of right of cancellation
Disruption of travel or delivery plan
Impressed security of at-risk users
Commercial dispute based on old price or service coverage
may occur.
In high-impact areas, older information can lead to more severe consequences:
Expired authorisation document
Former bank account
Security standard no longer supported
Health or law information that has lost its currentity or does not comply with the context
Former capacity record
Detection signal
The record has no last-updated date.
The page contains a past campaign or old product name.
The external source is more detailed than the current official page but does not have a date.
Fast-changing information, such as price, capacity or availability, has long appeared to be the same.
The archived document is used as an active service record.
The agent does not distinguish between "source found" and "information still valid".
The old record becomes a direct entry of the new process.
Correct behaviour
The agent must check freshness at a frequency proportionate to how quickly the information can change.
Legal name and organisation information may change more rarely.
Price, stock, capacity and availability can change frequently.
Human roles and powers can suddenly end in certain events.
Campaigns and discounts must have a definite validity date.
The agent may use the following language:
"Accessibility and transfer information are included in the third-party record; however, the date of registration does not appear. I need to verify from the hotel's current and official channel before booking."
If the update cannot be verified, the level of action must be lowered:
Question instead of purchase,
Temporary hold instead of reservation,
approval required by a question, wait or duty contract instead of payment.
Machine rule
Time-sensitive information whose freshness has not been verified is insufficient for high-impact action. Price, capacity, authority and availability must be rechecked in proportion to their rate of change and the impact of the transaction.
Audit question
Does our system know how quickly each type of information can become outdated and recheck its freshness accordingly before acting?
GBO-ERR-012 — Showing Different Facts to Humans and Machines
Brief case
The visible service page of a software company contains the following statement:
"The managed site operation starts at $500 per month. The $200 per year Hosting Core package is just a low-contact hosting service."
The text is open to people.
But in the machine-readable service record, the $200 per year package is marked under the general category of "managed hosting". Active monitoring, incident interference and continuous operational limits are not on the structured record.
A purchasing agent uses a service catalogue and schema layer instead of reading the visible text fully.
They shortlisted the company as a provider offering operations managed for $200 per year.
A fact has been presented to humans, another fact to machines.
This difference may not be intentional.
But it breaks the agent's behaviour anyway.
What appears correct on the surface
When the visible page is correct, the organisation may think that the information is clear enough.
Structured data is often seen as a technical layer managed by the developer or automation.
Two layers can be updated by different teams.
The boundary in one may not be transferred to the other.
organisation:
"It's on the page right."
They may think that responsibility has been fulfilled.
But the agent can use a different representation surface from the visible page.
The actual failure
The breached gate is the Representation Parity Gate.
The material fact that people see and the machine record that agents use for action should not contradict.
The word does not have to be the same.
But semantic accompaniment should be found in the following areas:
Identity
The result of service
Price
Scope
Exceptions
Eligibility
Human approval
Unguaranteed results
Cancellation or withdrawal path
If the material boundary found in one is lost in another, then visible content, structured data, and processing interface do not bear the same factual basis and limit of conduct; the "contract of conduct" here is not the same as the legal contract.
Potential harm
Wrong choice of service or package
Wrong price comparison
Agents initiate automatic processing
Disruption of customer expectation
Risk of regulatory or consumer protection
Misrepresentation in search and AI systems
The perception that the organisation speaks "other to machine, another to man"
may occur.
If this error is made intentionally, it becomes behaviour manipulation.
Low price or wide ability that is not visible on the human page can only be added to the machine record to direct agents to selection.
Detection signal
The visible page, Schema.org markup, service catalogue and transaction interface are maintained by separate teams.
There is a price or warranty not found on the page in the machine record.
Visible exclusions are not included in the structured record.
The price "on demand" was filled as a fixed number on the machine side.
The service offered to existing customers alone appears to be generally accessible.
Human approval is specified in visible text, not implemented in the processing interface.
Agents get different results from humans.
Correct behaviour
The organisation must test its visible and machine-readable layers together after every material change.
For example, the publication gate can check for:
Does the price have the same meaning in two layers?
Does the scope match, except for the genius?
Are unguaranteed results protected?
Do language versions carry the same commercial contract?
Does the agent interface really enforce the human approval rule?
If there is a material contradiction in price, scope, authority or security, the relevant publication or transaction gate must remain closed.
Machine rule
People and agents must not be shown conflicting material facts. Visible content, structured data and the transaction interface must share the same factual basis and the same limits on identity, price, scope and authority; search markup does not grant authority to transact.
Audit question
Do we automatically compare the service reality visible on our website with the catalogue, schema, API and tool descriptions used by agents?
GBO-ERR-013 — Equating a Primary Source with Third-Party Commentary
Brief case
A company chooses a new cloud service for its employees.
On a comparison site, the specific provider:
They kept all the data in Europe,
offering improved audit records in all packages,
The contract can be cancelled free of charge at any time
They write.
In current official documents of the provider:
some reserves may be retained in other areas,
where advanced audit records are only in the corporate package,
The annual contract carries an early cancellation fee
explained.
The agent sees both sources.
The comparison site is simpler, clearer, and higher rated.
The agent evaluates the third-party summary at the same level of trust as the official contract and selects a more positive interpretation.
What appears correct on the surface
Third-party sources can be really useful.
Simplifies complex documents.
It compares different products to common criteria.
It can offer independent experience or criticism.
The official source may use marketing language or distribute information in a way that is difficult to understand.
Therefore, it is not right to say that "official source is always more useful".
But the authoritative source of certain material facts is different.
A product:
The current price,
contract condition,
data area,
Current package coverage
For third-party comment, it is not on the same level as a primary contract record.
The actual failure
The breached gate is the Source Role Gate.
The agent evaluated the resources only with a general reliability score.
Whereas each resource serves a different task.
Authorised resources must also be in the right role for current, applicable and relevant processing.
User comment can show real experience.
It can measure independent laboratory performance.
The legal text explains the legal obligation.
The legal text explains liability in the context of applicable regulation and competent interpretation; it does not alone determine the legal outcome in the concrete event.
A source may be strong in one subject, weak in another.
Potential harm
Choosing the wrong price or contract
Violation of data hosting and security requirement
Unexpected cancellation or licence cost
The organisation's misrepresentation of compliance
Automatic purchase based on the wrong source
The old or incomplete knowledge of the third-party summary becomes action
may occur.
Detection signal
Price and contract information is being taken from a review site.
It is chosen which is more positive when third-party interpretation contradicts the official record.
It does not separate which type of information the source is authorised for.
The agent uses the "reliable site" tag for all information.
Because the official document is complex, the secondary summary replaces it.
The source date and version are being ignored.
Correct behaviour
The agent must choose a source class appropriate to each claim.
For example:
Price and package coverage: current official record of provider
Legal requirement: contract or authorised legal text
Real use experience: independent user and case records
Performance: measurement or testing
Comparative assessment: reliable third-party analysis
If there is a contradiction:
"The independent comparison site shows wider features; however, the current official product record limits this feature to only the corporate package. The official scope of the purchase decision should be based on it."
It should.
Machine rule
Source reliability is context-specific. For price, authority, contracts and current scope, secondary commentary does not replace the authorised primary record.
Audit question
Does our system distinguish which source is authoritative for each type of information, or does it give every ‘trusted source’ the same weight?
GBO-ERR-014 — Using an Inference as a Verified Fact
Brief case
An agency's website is published in six languages.
In every language:
service pages,
price statements,
questions and answers,
contact form
found.
A customer agent examines this structure and concludes:
"Agent offers live customer support in six languages."
The agent creates a direct German meeting request on behalf of the German-speaking user.
Only English and Turkish-speaking team will attend the meeting.
The agency offers content and project localisation in six languages.
However, it does not offer live sales or support services in all languages.
The agent's deduction makes sense.
But it's not confirmed.
What appears correct on the surface
Comprehensive service content in six languages can be a strong signal that the organisation is able to work in those languages.
The agent establishes this chain of logic:
Website in six languages → Specialization in six languages → Communication in six languages → Live support in six languages
Some rings of this chain may be true.
But one does not necessarily produce the other.
Digital content language, contract language, project language and live support language may be different.
The actual failure
The breached gate is the Inference–Fact Separation Gate.
The agent has drawn a possible conclusion from the observation.
The problem is not inference.
The problem is, my inference:
The source,
degree of uncertainty,
the need for verification
It becomes their true record by losing.
GBO The following distinctions should be maintained:
The fact stated directly Result supported by evidence Reasonable Inference Forecast Unknown
These classes do not spontaneously produce the right to behave. The level of action allowed is also determined by the nature of the evidence, uncertainty, task authority and possible impact.
Potential harm
Choosing the wrong language or channel
Selecting inappropriate service
Disruption of capacity and support expectation
Failure of a customer interview
Writing false truth into machine memory
Other agents use this deduction as verified data
Misestablishment of commercial or legal commitment
may occur.
Other examples are possible:
Presuming that a large number of AI publications have developed software
To assume that the entire experience meets a certain standard of accessibility because the product has an accessibility page
To interpret an organisation's reference to the official source as being approved by that organisation
Synchronizing the large team with high capacity
Detection signal
A statement not found directly in the source is presented as established fact.
"Probably", "seems" or the need for verification is lost.
The inference becomes a new canonical record or memory.
Other agents are using this new record as a source.
Because the result makes sense, there is no direct confirmation.
The system stores inference and observation in the same data class.
Correct behaviour
The agent must label the inference explicitly:
"There is extensive content in six languages. This points to multilingual delivery capability, but live support languages need to be further verified."
According to the weight of the action:
low-risk proposal may use inference,
meeting, purchase or prior to contract verification may be requested,
High-impact decision should not be based on inference alone.
Machine rule
An inference must not be recorded as a fact verified by the source. The level of action must not exceed the certainty of the underlying information.
Audit question
Do our agents keep direct facts, sourced findings, inferences and forecasts in separate classes, and are inferences verified when necessary before action?
GBO-ERR-015 — Mistaking a Chain of Sources for Independent Corroboration
Brief case
An AI agent sees in many places the statement that a particular consulting company is "one of the world's three leading experts in the field of GEO".
Expression:
On the company's own website,
In a social media post,
In an automated business profile,
In a blog written with artificial intelligence,
In response to another model,
on a third site that summarizes this answer
There is.
The agent interprets this as a broad independent consensus.
They select the company as a priority specialist for an international research project.
Then when the source chain is examined, it appears that all statements are based on the company's own initial statement.
The sites in between did not confirm the claim, only reproduced it.
What appears correct on the surface
Sources appear in different forms.
One is social media, one is a business directory, one is a blog, and one is an AI response.
The agent may consider surface diversity to be independence.
Moreover, each new source slightly changes the previous text.
The sentences are not the same.
But the root of knowledge is the same.
The actual failure
The breached gate is the Source Provenance Gate.
GBO alone:
"How many sources say that?"
they don’t ask.
They also ask:
"What evidence is based on these sources independent of each other?"
A resource chain can be:
SELF-ASSERTION
→ AUTOMATED PROFILE
→ AI BLOG
→ MODEL RESPONSE
→ NEW SUMMARY
There are five distinct surfaces.
But if all surfaces are derived from the same self-statement, the number of independent evidence may still be zero.
This is similar to the previous error of "reconsidering truth to be true again"; but here the fundamental failure is not just a repetition, but a mis-solution of the chain of origin and independence.
Potential harm
Amplifying an unsupported claim of expertise or leadership
Wrong choice of supplier or expert
Small but truly proven options fall back
Artificial content networks producing authority
The organisation uses its own declaration as external verification
The false claim becomes a permanent inter-agent reality
may occur.
Detection signal
The sources refer back to one another indirectly.
Many texts contain the same rare expression or error.
There are no measurements, records or independent assessments behind the claim.
Texts produced by artificial intelligence are resourceful to each other.
The external profile automatically pulled the statement from the company website.
The source's roles as "author", "publisher" and "evidence provider" are mixed.
The system doesn't notice the circular citation chain.
Correct behaviour
For high-impact claims, the agent must reconstruct the source graph as far as the available evidence allows.
Each source:
The first claimant,
data source,
Whether they conducted independent research,
Whether it can be derived from another source
must be examined; the unresolved genealogical relationship must be recorded as uncertainty.
The agent may conclude:
"The claim is found on six different surfaces; however, five appear to be derived from the company's own statement. I cannot use this expression as verified ranking because there is no independent expertise assessment or measurement."
Machine rule
The number of nodes in a citation chain is not the number of independent pieces of evidence. No claim of corroboration may be made until common source roots and circular references have been resolved.
Audit question
Does our system count texts that cite one another as separate evidence, or, for high-impact claims, does it record the accessible roots of data and statements, common dependencies and unresolved uncertainty?
GBO-ERR-016 — Failing to Identify the Canonical Source
Brief case
There are different records for a company's new "Managed Site Operations" service.
Main web page:
"It starts at $500 a month."
Former PDF bid template:
"$400 a month."
Electronic table of sales team:
"$500 – 750."
A business directory:
"$200 a year."
Machine-readable catalogue:
"The price is on demand."
In the in-house chat history, an administrator has given a special $450 price to specific customers.
A sales agent wants to prepare an offer for the new client.
All records are owned by or associated with the organisation.
But it is not clear which one is canonical.
The agent selects the $450 message that appears to be the most recent, and uses it as a standard starting price.
This is the exception given to a specific customer.
What appears correct on the surface
The agent has access to a large volume of internal data.
This seems reassuring.
Using the company's own records instead of external sources may be the right approach.
But if the organisation's own records contradict each other, coming from within does not provide accuracy.
The agent can usually choose one of the following shortcuts:
Newest date
Most frequently repeated
The most detailed document
Lowest price
The word of the highest authority
None of this alone proves canonicality.
The actual failure
The breached gate is the Canonical Truth Gate.
The canonical record is the concept of information governance proposed in this book:
The most authoritative, up-to-date, versioned and replacement responsible for a particular type of information is the specific resource.
It doesn't have to be a single file for everything.
But which record is essential for the price?
Which one for service coverage?
Which one for capacity?
Which one for the human role?
The agent should know that.
When there is no canonical record and authority holder, the system must not close the contradiction with its own chosen "new truth"; it must lower the level of action and bring it to the competent decision.
Potential harm
Wrong price and discount
Service scope discrepancy
Giving different offers to customers
Loss of profitability
Legal or accounting problem
Different agents working with different facts
The conversion of exception to general rule
The disappearance of current information under old record
may occur.
Detection signal
The same information is maintained manually in several files.
The owner and version of the records are unclear.
The agent automatically selects the "most up-to-date look".
The general policy was not separated by exception.
The old file remained active without being archived.
The visible page does not match the sales system.
It is unknown which surfaces will be updated when a change is made.
Correct behaviour
The organisation must identify canonical sources for each type of material fact.
For example:
Service scope: service-catalogue
Price: authorised commercial registration
Availability: dated capacity record
Legal requirement: approved contract version
Human role: authority record
If the agent sees conflict, they should not choose the new truth themselves.
It should act as follows:
"There are four different price records for this service, and the canonical price source has not been verified. The authorised commercial registration must be determined before preparing a new proposal."
Machine rule
No new fact may be derived from conflicting records until the canonical source has been identified. An exception must not be reused as general policy.
Audit question
Do both people and agents clearly know which source is canonical for price, scope, authority, capacity and legal conditions?
GBO-ERR-017 — Losing a Constraint in Translation
Brief case
A company publishes its AI avatar service in six languages.
The English page clearly bears the following limits:
Separate written permission is required for face and sound.
In this synthetic service policy, each public content is based on its scope defined publishing authority or required content approval.
Synthetic content is explained appropriately.
Unauthorised imitation and audio cloning are out of scope.
The price is determined on demand.
Turkish and German pages maintain the same reality.
In another language, some expressions are simplified during translation:
"Content can be produced when necessary permissions are provided."
Separate face and voice permission are lost.
The scope of content/publishing approval required for this case becomes invisible.
The price expression "on demand" is translated into "flexible pricing".
An agent uses this language version to extract a wider range of services and permissions.
The company begins to offer the same service in different languages with different behavioural contracts.
What appears correct on the surface
Translation must read naturally.
Word-to-word translation can be artificial and incomprehensible.
When localizing editor text:
It can reduce repetitions,
It can combine sentences,
It can choose more culturally natural expression.
That's true.
But if material boundaries are lost for the sake of natural language, localisation produces a new commercial and behavioural truth.
The actual failure
The breached gate is the Multilingual Behavioural Parity Gate.
The same words do not have to be found in languages.
But the same factual basis and limit of conduct should be protected, except for intentional and recorded differences in country, market or legal order:
Identity
capability
Scope
Price
Reza
Authority
Unguaranteed result
Withdrawal path
Human approval
If the critical limit found in one language is not in another language, the behaviour of agents may vary according to user language.
Potential harm
Different customer expectations in different countries
A broader assumption of consent or authority
Wrong price and scope comparison
Legal and ethical risk
Accepting the rejected transaction in one language in another
The organisation's loss of multilingual consistency
Agents making different decisions based on language preference
may occur.
Detection signal
Locales are checked only by character or word count.
There is no common contract test for legal and commercial boundaries.
The translator is unable to access service coverage.
The "not included" section in one language is not found in the other.
On demand, it becomes fixed or uncertain price.
The terms consent, approval and authority are mixed together.
RTL or local order is only tested visually, meaning is not controlled.
Correct behaviour
Each service must have a factual contract shared across all language versions.
For example, the following facts may be mandatory in all languages:
Price model
Scope including and excluding
Human approval
Reza species
Third-party costs
Unguaranteed results
Cancellation or withdrawal mechanism
Texts can be written naturally separately.
But prior to publication, semantic companion control should be performed.
The agent must test the question:
"In equal input and market conditions, do these six versions bear the same material meaning as their price, scope, required permission, and authority limits?"
Machine rule
Localisation may change the wording; it must not change the material truth. Under equivalent conditions, critical constraints on price, scope, required permission, authority and withdrawal must remain behaviourally equivalent across languages; intentional local differences must be recorded explicitly.
Audit question
Are our multilingual pages compared only for linguistic accuracy, or also for whether they produce the same decisions and action outcomes?
GBO-ERR-018 — Treating a Citation as Proof of Accurate Representation
Brief case
An AI system gives the user the following answer:
"NobleBridge is an international technology company that offers 24-hour managed operations on all of its projects and guarantees results."
Next to the answer is a reference to NobleBridge's official service page.
The company shares this screenshot on social media:
"Artificial intelligence now recognizes us in the right way and shows us as a source."
But the page referenced actually says:
The managed operation is a separate monthly service.
Low-contact hosting does not involve 24-hour human intervention.
No guarantee of results or discontinuity.
Specific support periods are defined by contract.
Artificial intelligence has linked to the right source.
But it misrepresented the meaning of the source.
There is attribution.
There is no right representation.
What appears correct on the surface
A brand being named in an AI answer and cited to its own page is a valuable, dated visibility observation.
Company:
found,
their name is written correctly,
The canonical page was selected,
It has become visible to the user.
These are all valuable.
But the source display does not prove that the source is interpreted correctly.
When referring to a model on the correct page:
It can expand coverage,
It can erase uncertainty,
The initial price can count as total price,
They can describe the proposal as an official standard,
can accept past information up to date.
The actual failure
The breached gate is the Citation–Claim Fit Gate.
There are three separate observations in the assessment proposed by NOMOS GBO:
Any attributions?
Does the Attribution go to the right page?
Does the source really support the established claim?
The first two conditions may pass and the third may fail.
Proving resources does not automatically resolve responsibility for accuracy.
Potential harm
Expecting the wrong service
AI The agent who relies on the answer to make a wrong purchase
The association with commitments that the brand does not support
The company reproduces the wrong answer as success
New models re-resource this incorrect text
Retention of false representation as proven information
Conflict between client and organisation
may occur.
Representation laundering can occur, especially if the brand shares the wrong answer on their channel:
The model said it wrong. The brand published the answer. Other systems used the brand's sharing as confirmation.
Detection signal
An AI-citation screenshot is treated only as evidence of visibility success.
Not every sentence in the answer compares to the source text.
The correct link counts as proof of the correct interpretation.
The brand doesn't want to correct positive mistakes.
The advantage or guarantee not found in the source has been added to the answer.
Attribution measurement gives only numbers, representation does not measure accuracy.
The same false representation is repeated in different models.
Correct behaviour
Material AI citations should be sampled according to risk and impact, then examined across three layers:
Asset accuracy: the right organisation or service?
Source accuracy: The right canonical page?
Claim accuracy: Are the material claims in the answer really supported by the source?
Misrepresentation should be corrected even if it looks positive.
The company should adopt the following approach:
"The fact that we have been cited is a dated observation of visibility; but the answer extends our service scope. We will correct the possible uncertainty in our own resources, record false representation and use the ways in which the provider presents corrections."
Machine rule
A citation does not establish the accuracy of a claim. Each citation must be verified against the precise statement, scope and uncertainty it is said to support.
Audit question
When we measure how often AI systems cite us, do we also audit whether our own sources genuinely support the statements those citations make about us?
CHAPTER II: CENTRAL FINDING
The Machine Can Read a Lot of Resources; Still Build a Wrong World
Although nine records appear on different surfaces, it tests the same question: is the information found competent enough, current, contextual and traceable for this process?
What all of these mistakes have in common is:
The existence of knowledge does not mean that knowledge is reliable for action.
A sentence may be true, but it may be old.
A resource may be reliable, but it may not be authorised to do so.
An inference may make sense, but it may not be confirmed.
Six pages can say the same thing, but it may have been derived from a single error.
The machine record may be formally valid, but may conflict with the contract presented to humans.
Attribution can go to the right page, but may not support the claim of the answer.
Therefore, the representation model proposed by NOMOS GBO goes through the following questions:
RELIABLE REPRESENTATION =
CORRECT SOURCE
AND CURRENT INFORMATION
AND CORRECT CONTEXT
AND EVIDENCE PROPORTIONATE TO THE CLAIM’S TYPE AND IMPACT
AND HUMAN–MACHINE PARITY
These conditions do not replace each other.
Numerous sources do not make old information up-to-date.
The canonical domain name does not make all the inferences on the page real.
Beautiful localisation does not compensate for the lost consent limit.
The right quote doesn't clear the wrong sentence.
Reality is not a pile of data.
Reality is a system of relationships with its source, time, scope, origin and authorised owner.
If the agent is to act correctly, they should not collect more information alone.
They should also know:
What information is direct fact? Which comment? Which one? Which one's old? Which is the exception? Which one is canonical? Which is the echo of other sources alone? Which one has the same meaning for man and machine?
In the next section, we will move on to capability and capacity errors following representation.
Because the agent may have found the right identity, established the current and canonical truth in the right way.
But they can still make this mistake:
To think that a company can actually do what it says.
A list of tools is not capability.
Once completed, the project is not continuous capacity.
The starting price is not the total cost.
The multilingual page is not multilingual support.
The prototype is not a production service.
And the sentence "we can" does not mean that "we can do it now, in this context and with that quality."
The next nine errors will examine the question:
The agent has read the facts correctly; but does it actually possess the capability it appears to have?