The Difference Between a Forgery and a Guess

Opening Scene

An appraiser distinguishes carefully between a deliberate forgery — something manufactured specifically to deceive — and an honest but mistaken attribution, where a genuine object was simply misidentified through a reasonable, if incorrect, chain of reasoning. These require genuinely different responses. Hallucination in language models carries this same important distinction, and understanding it shapes how you address the problem.

In Plain English

Fabrication is when a model generates something with no genuine basis at all — an entirely invented citation, a fictional statistic. Confabulation or extrapolation-based hallucination is when a model reasons from partially correct information toward an incorrect but superficially plausible conclusion — closer to an honest mistake than an outright invention. Both are genuinely problematic, but they call for different mitigation strategies: fabrication is best addressed through grounding, covered in Article 4, while confabulation often requires better reasoning verification and self-consistency checks, covered in Article 7.

The Old Way

Before this distinction was well understood, hallucination was often treated as one undifferentiated problem:

Hallucination was sometimes treated as one single, undifferentiated failure mode, without recognizing that different underlying causes call for different mitigation approaches.
There wasn’t yet a well-established practice of diagnosing whether a specific hallucination was outright fabrication or a more subtle, reasoning-based confabulation.
Mitigation efforts sometimes applied the same technique universally, missing gains available from matching the specific mitigation to the specific type of error.

Distinguishing fabrication from confabulation, and matching mitigation deliberately to each, reflects a maturing, more precise understanding of hallucination as a genuinely varied phenomenon.

What’s Changing (and Why AI Is the Reason)

Evaluation practices increasingly diagnose hallucination by type, distinguishing outright fabrication from reasoning-based confabulation, informing more targeted mitigation.
This connects directly to the grounding techniques covered in Article 4, most effective against fabrication, and the self-consistency checks covered in Article 7, most effective against confabulation.
As this distinction has matured, mitigation strategies increasingly get matched deliberately to the diagnosed type of hallucination risk a specific application faces.

The Metaphor, Fully Extended

The Antiques Appraiser	Fabrication vs. Confabulation Concept
A deliberate forgery manufactured specifically to deceive	Outright fabrication with no genuine basis at all
An honest but mistaken attribution from reasonable inference	Confabulation reasoning from partial truth toward a wrong conclusion
Genuinely different responses required for each case	Genuinely different mitigation strategies required for each type
Diagnosing which one you’re dealing with before responding	Diagnosing which type before choosing a mitigation approach

For Beginners: What to Actually Do

Practice examining a hallucinated output and diagnosing whether it looks like outright fabrication or a reasoning-based confabulation from partial truth.
Learn to apply grounding, covered in Article 4, as the primary defense against fabrication specifically.
Get comfortable recognizing confabulation as a distinct, subtler failure mode requiring different verification techniques.

For Practitioners and Leaders: The Deeper Layer

Build diagnostic practices that distinguish fabrication from confabulation when evaluating hallucination incidents.
Match mitigation strategy deliberately to diagnosed type: grounding for fabrication, self-consistency checks for confabulation.
Track hallucination incidents by type over time to identify which mitigation investments would have the most impact.

Quick Recap

Fabrication is hallucination with no genuine basis at all; confabulation reasons from partial truth toward an incorrect conclusion.
These are genuinely different failure modes calling for different mitigation strategies.
Grounding is most effective against fabrication; self-consistency checks are most effective against confabulation.
Diagnosing the type of hallucination informs more targeted, effective mitigation.

Where This Fits in the Series

Article 5 covered the distinction between fabrication and confabulation. Article 6 turns to tracing the chain of ownership: source attribution and citation practices.