CoChat

Hallucination or Confabulation? Neuroanatomy as metaphor in Large Language Models

Open access
Smith, Andrew L., Greaves, Felix, Panch, Trishan · 2023 · PLOS Digital Health
Cited by 92

The DOI matches its Crossref record; 3 of the 3 references it cites were found in a registry or an index.

Abstract

As Large Language Models (LLMs) and their capabilities become an increasingly prominent aspect of our workflows and our lives, it is important that we are thoughtful and deliberate with the words we use to refer to the inner workings and outputs of this technology.We think that conveying the complex functions (and malfunctions) of LLMs using metaphorical language that is precise and accurate can lead to a better understanding of these powerful tools among both the academic community and the public.If we are meticulous in our choice of metaphors, we open ourselves up to the possibility of achieving a better shared understanding of the complex concepts in this exciting new field.Here, we give the specific example of AI "hallucinations" and demonstrate how a change in metaphorical language can lead to new ways of understanding and may even foreshadow future directions in the development of artificial intelligence systems.In psychiatry, hallucinations are a relatively well-defined perceptual phenomenon referring to sensory experiences without associated external, or 'real-world' stimuli.Clinically, hallucinations are commonly associated with conditions such as schizophrenia, bipolar disorder, and Parkinson's disease [1].Employing the term "hallucination" to characterize the inaccurate and non-factual outputs generated by LLMs implies acceptance of the notion that LLMs are engaged in perceiving, that is, becoming consciously aware of a sensory input.While this is a subject of some ongoing debate, there is currently no evidence that AI has gained conscious awareness [2].LLMs do not have sensory experiences, and thus cannot mistakenly perceive them as real.As such, we believe the term "hallucination" misrepresents the nature of the process occurring within LLMs which it has been used to describe.The model is not "seeing" something that is not there, but it is making things up.More accurate terminology is found in the psychiatric concept of confabulation, which refers to the generation of narrative details that, while incorrect, are not recognized as such.Unlike hallucinations, confabulations are not perceived experiences but instead mistaken reconstructions of information which are influenced by existing knowledge, experiences, expectations, and context.Confabulation can occur in various clinical conditions including dementia, Wernicke-Korsakoff's syndrome, schizophrenia, traumatic brain injury (TBI), and cerebrovascular accidents (CVAs) [3].Confabulation is frequently associated with a generalised lack of awareness of one's deficits often seen in right sided CVAs or TBIs, as well as in bipolar disorder, schizophrenia, and the dementias [4].When answering questions, LLMs generate responses based on learned patterns in very large datasets [5].The output of LLMs can

Overview

The claims and statistics of this paper haven’t been checked yet.

Free: it reads the open-access text and checks the first 5 claims that cite a source we can check. Create a free account to check every claim.

A few free checks a day · papers already verified open instantly · results are public

Integrity signals

Registry

Crossref · 10.1371/journal.pdig.0000388

Cites

3 of 3 references verified0 citation errors · 0 not checkable

Statistics

Verification

Claims the paper makes

The claims and statistics of this paper haven’t been checked yet.

Free: it reads the open-access text and checks the first 5 claims that cite a source we can check. Create a free account to check every claim.

A few free checks a day · papers already verified open instantly · results are public

Self-check

Percentages, sums and reported differences re-computed from the paper’s own numbers.

The self-check needs the paper’s full text.

Citations

Cites

Cited by

See all 92

Metadata

Authors
Smith, A. L., Greaves, F., & Panch, T.
Venue
PLOS Digital Health
Published
2023-11-01
Volume
2
Issue
11
Pages
e0000388
ISSN
2767-3170 (ISSN-L)
OpenAlex
W4388142096
Cited by
92
References
3
FWCI
80.57
Type
journal article
Language
en
Open access
gold

About these checks

Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.

Statistics and arithmetic are recomputed, not judged.

Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.

What counts as a claim ›How we measure confidence ›How we check this ›

Are you an author, or do you see a problem with this page? Request a review

Data sources

Bibliographic data: Semantic Scholar, OpenAlex, Crossref.

Contains information from the Semantic Scholar Academic Graph Datasets, made available under the ODC Attribution License (ODC-BY 1.0).

Semantic Scholar’s data is described in: Kinney, R., et al. (2023). The Semantic Scholar Open Data Platform. arXiv:2301.10140.