The Confabulation Problem (AI)

In the AI Story

The clinical literature on confabulation—Korsakoff syndrome, split-brain patients, right-hemisphere stroke victims—establishes consistent features. Confabulated claims are contextually appropriate (they fit the conversational situation). They are delivered with normal fluency and confidence (the patient does not experience them as fabrications). They resist correction (because the patient believes them). And they fill gaps—when asked why the paralyzed arm is not moving, the patient generates a plausible narrative ('I don't want to,' 'I already did,' 'my arm is tired') that satisfies the coherence requirements of their self-narrative. The fabrication is not strategic deception. It is the brain's automatic gap-filling, operating below conscious awareness. The parallel to AI is precise. Large language models fill gaps continuously—that is what next-token prediction does. Given context, the model extends it in the statistically most likely direction. If the context requests information the model does not possess, the model does not report absence of knowledge (or does so only when trained to). It generates the statistically most likely completion. The completion is often accurate. Sometimes it is confabulated. The surface features are identical.

Confabulation differs from simple error in ways that matter epistemologically. A simple error—'the capital of France is London'—conflicts with known facts and triggers detection through coherence-checking alone. The claim contradicts established knowledge. The evaluator notices the contradiction and rejects the claim. Confabulation produces no such conflict. The confabulated claim extends established knowledge in a direction that feels natural, that does not conflict with the evaluator's existing beliefs. The Deleuze fabrication Segal caught was not a simple error. It was a sophisticated extension of two genuine bodies of thought (Csikszentmihalyi's flow theory and Deleuzian philosophy) that cohered beautifully. Every internal feature signaled quality. Only anchor-checking—consulting Deleuze's actual positions—revealed the fabrication. The confabulation was designed (by architecture, not intention) to pass coherence tests. Detection requires the specific, effortful work of grounding verification that the model's fluency discourages. The evaluator who relies on coherence alone accepts the confabulation, because the confabulation is coherent.

Retrieval-augmented generation (RAG) is the dominant engineering response to confabulation. The model retrieves verified documents before generating—anchoring output in sources that have been independently checked. RAG reduces confabulation rates measurably. Studies comparing RAG to unaugmented models show significant improvements across domains. But Haack's framework reveals residual danger. RAG produces output that is grounded at some points (where retrieval occurs) and ungrounded at others (where the model fills gaps between retrieved documents with pattern-based completions). The grounded and ungrounded elements are woven together seamlessly. The seam is where the evaluator could distinguish reliable from unreliable—and seamlessness erases the seam. The result is partial grounding that creates false security. The evaluator knows some content is anchored and lowers epistemic guard. Confabulated extensions enter through the opening. Haack's independent security dimension applies: the quality of RAG's grounding depends on the quality of retrieved documents, which depends on database curation. A RAG system grounded in peer-reviewed literature is more reliable than one grounded in blog posts. But reliability is invisible to the user unless the system transparently distinguishes retrieved from generated content—which current systems do not.

The confabulation problem is not a temporary bug awaiting an engineering fix. It is an architectural feature of systems that generate output through statistical pattern completion rather than evidential reasoning. Future models will confabulate less frequently as training improves and grounding mechanisms become more sophisticated. But the structural gap between the model's ability to produce coherent claims and its inability to verify those claims correspond to reality is epistemological, not engineering. Closing the gap requires human epistemic labor: checking anchors, tracing claims to sources, maintaining the distinction between 'sounds right' and 'is right.' The labor does not scale. The model generates at computational speed. The human checks at cognitive speed, constrained by domain knowledge, available time, and tolerance for tedious verification. The asymmetry is structural. Most AI-generated claims go unverified. The ones that get through may be true or confabulated—without checking, the evaluator cannot tell. Haack's prescription is demanding: treat AI output as proposed crossword entries requiring clue-checking before acceptance. The presence of grounded claims strengthens the grid but does not justify accepting ungrounded claims by proximity.

Origin

The term 'confabulation' entered AI discourse around 2022–2023 as an alternative to 'hallucination,' which implies perceptual failure. Confabulation—borrowed from clinical neuropsychology—more accurately describes the mechanism: narrative gap-filling by a system that produces coherent fabrications without experiencing them as fabrications. The clinical literature traces to Sergei Korsakoff's 1880s descriptions of memory confabulation in alcoholic patients, extended by Morris Moscovitch and others across the twentieth century. The application to AI recognizes that language models fill gaps the same way damaged brains do—automatically, coherently, without epistemic access to whether the filling corresponds to reality. The Susan Haack—On AI simulation applies Haack's foundherentist framework to diagnose why confabulation is harder to detect than error: confabulations satisfy coherence (intersections) while violating grounding (clues), and human evaluators are cognitively calibrated to trust coherence as evidence of truth.

Haack did not develop foundherentism to address AI—Evidence and Inquiry (1993) predates the current moment by three decades. But her framework's application is not retrofitting. It is recognition that the epistemic structure she diagnosed—the necessity of both anchoring and coherence, the insufficiency of either alone—applies with intensified urgency to outputs produced by pure coherence engines. The model's confabulations are, in Haack's terms, beliefs that pass the coherentist test (internal consistency, mutual support) while failing the foundationalist test (experiential grounding). The evaluator applying only one test accepts fabrications. The evaluator applying both—checking clues and intersections—catches them. The framework makes visible what the fluency conceals.

Key Ideas

Gap-filling is architectural. Next-token prediction extends context in statistically likely directions—filling informational gaps with pattern-based completions that may or may not correspond to reality.

Confabulation vs. error. Errors conflict with known facts and trigger coherence-checking detection. Confabulations cohere with existing knowledge and evade detection without anchor-checking.

RAG's false security. Partial grounding creates the illusion of comprehensive grounding—retrieved documents anchor some claims, ungrounded extensions ride the coattails of reliability.

Detection requires domain knowledge. Catching confabulations demands independent expertise to recognize the gap between what the model claims and what the evidence supports—a requirement that does not scale.

Unverified confabulations degrade the commons. Each undetected fabrication entering the shared informational environment becomes part of the web against which subsequent claims are checked—compounding error across iterations.