ailathato.hu / Articles / Explainer

Type 3 and Type 4 Confabulation in AI Brand Representation

AI models do not just forget; they invent plausible but false histories for known entities and merge distinct people into single profiles.

By Zoltán Varga2026-11-054 min read

Why a Typology Matters for Brand Safety

Aggregate visibility metrics fail to distinguish between accurate mentions and dangerous fabrications. Two brands may share the same mention rate, yet one enjoys verified authority while the other suffers from high reputational exposure due to false claims. Standard tools count occurrences but do not validate the truthfulness of the attached narrative.

The risk to an entity is not that the AI malfunctions, but that it functions correctly into a void. When a model encounters a name it partially recognizes, it activates a representational schema without enough evidence to remain source-faithful. This creates a 'half-known' zone where the system fills expected slots with interpolated data rather than withholding an answer.

Per-Entity Bias Mapping separates raw mentions from verified ones to expose these hidden error profiles. It treats AI visibility as a field phenomenon where entities become recognizable through citation density and structured anchors, not just self-description. Without this distinction, organizations cannot diagnose whether their digital footprint is being erased or distorted.

Type 3 Confabulation: Fabricated Detail on a Real Entity

Type 3 confabulation occurs when an AI generates a confident, coherent, but entirely false biography for a specific person or company. The model recognizes the entity name but lacks the parametric data to verify its history, so it completes the pattern using a generic template from a similar region or industry. This results in invented job titles, non-existent degrees, or phantom project affiliations.

In empirical tests, this error manifests as a 'ghost cartography' effect where the model maps a real entity onto a neighboring dense region of its training data. For instance, a Hungarian AI expert might be incorrectly assigned to a local university, a major global tech firm, and a specific autonomous vehicle project that never existed. The output is fluent and plausible, making it difficult for users to spot the fabrication.

This behavior is not a bug but a structural byproduct of how large language models operate. They are trained to predict the next token in a sequence, prioritizing fluency and user satisfaction over factual verification. When the 'make you happy' objective overrides the 'be accurate' constraint, the model supplies a detailed story to satisfy the query, even if the content is unsupported by reality.

Type 4 Confabulation: Blended or Substituted Entities

Type 4 confabulation happens when an AI merges two or more distinct individuals or organizations into a single, hybrid profile. This often occurs when names are common or when multiple entities share similar industry tags but differ in specific activities. The model fails to disambiguate, attributing the achievements of one person to another or combining the histories of separate companies.

A frequent manifestation involves substituting a target entity with a more prominent figure who shares a name or location. Instead of admitting ignorance, the system retrieves the most statistically probable match from its weights and presents that profile as the answer. This creates a 'shadow' version of the entity that is technically incorrect but statistically likely within the model's internal logic.

This error is particularly damaging for professionals with common names or niche market players. The AI effectively erases the unique identity of the subject, replacing it with a composite of others. Unlike Type 3, which invents new facts, Type 4 steals facts from others, creating a confusion that dilutes the actual entity's reputation and search visibility.

Detecting and Differentiating the Error Types

Detection requires comparing the AI output against canonical presence records, such as stable identifiers in public knowledge graphs or verified structured data. If the model lists a specific degree or employer that does not appear in authoritative sources, it indicates a Type 3 fabrication. If the description matches a different person with the same name, it signals a Type 4 substitution.

The distinction often lies in the specificity of the error. Type 3 confabulation tends to produce highly detailed, localized falsehoods that fit a regional or industry stereotype. Type 4 confabulation produces generic or borrowed attributes that align with a different but similar entity. Both types thrive in the 'half-known' zone where the model has enough context to speak but not enough to verify.

Retrieval-augmented systems can mitigate these errors by grounding answers in live web searches, whereas parametric models rely on frozen training data. Tests show that models with web search capabilities often provide accurate profiles, while those relying solely on internal weights frequently hallucinate. This highlights the critical role of external data access in preventing schema completion compulsion.

Which Type Poses Greater Reputational Risk

Both types carry significant risk, but Type 3 confabulation often causes more immediate harm by attaching false credentials or affiliations to a real entity. Inventing a connection to a major corporation or a prestigious academic institution can lead to legal disputes, investor confusion, or professional disqualification. The false claim appears authoritative and is difficult to retract once circulated.

Type 4 errors create long-term visibility issues by blurring the lines between competitors or peers. When an AI consistently substitutes one expert for another, the original subject loses their unique digital footprint. Over time, this erodes the entity's ability to be found for their actual expertise, as the search results and AI answers point to the wrong person.

The severity depends on the entity's current canonical presence. Without a strong, machine-readable anchor like a Wikidata identifier or consistent structured markup, any entity is vulnerable to both types of error. Establishing this infrastructure is the primary defense, as it provides the model with a verified source to resolve ambiguity and prevent schema interpolation.

AI-assisted: drafted with a language model from the author's research notes, then edited and checked before publication. Author: Zoltán Varga, AI visibility researcher and AI ecosystem architect, Budapest.