The Yoga tradition classifies mental activity into a small number of distinct kinds. One of them is the case where the mind, prompted by language, generates a representation of something that does not exist. The classic example is “the horns of a hare.” The words combine grammatically; the mind constructs a coherent image; there is nothing for the image to be of. This is its own cognitive event, separable from valid perception, from being wrong about something real, from memory, and from sleep.
The reason the classification matters is that it tells you which kind of failure you are looking at. A representation generated by language with no referent calls for one kind of fix. A representation with a real referent that has been wrongly identified calls for a different kind. The shared English word “error” treats these as the same problem and produces single-solution thinking.
This cognitive event is also a tool. Poetic objects, mathematical entities, design fictions, hypothetical scenarios — all involve the deliberate generation of representations whose referent is empty by design. The discipline here is recognition: knowing when the event is happening and treating it for what it is.
Where English Falls Short
“Hallucination,” in the LLM sense, names the phenomenon but does no analytic work. The Yoga frame forces a distinction the AI literature collapses. A model that invents a non-existent paper is producing one kind of cognitive event — the linguistic form is generating a representation without any referent. A model that misattributes a real paper to the wrong author is producing a different kind — there is a referent in play, and the model has the wrong one. These call for different fixes: grounding interventions for the first, retrieval improvements for the second. The single word “hallucination” hides this split, and the field’s debugging practice suffers for it.
Where it Shows Up
LLM output that fabricates citations, court cases, API endpoints, library functions. The linguistic form is generating cognition without grounding. The fix is structural — retrieval, grounding, evidence anchors at generation time.
AI-generated research synthesis summarising across no actual studies. Representation without referent, with the masquerade carried by the genre conventions of research writing. The fix is to require evidence anchors and to mark generated content as such.
Persona work. Almost all personas are referentless representations. Treating them as if they were derived from real users is the failure. The honest discipline is to acknowledge their actual epistemic status.
Design fictions and speculative artefacts. Referentless representation used intentionally. The failure mode here is generating these without marking them — letting the audience receive them as if they were grounded.
Debugging model output. Before reaching for a fix, classify the failure. Empty-referent failures and wrong-referent failures look identical on the surface and have different treatments.
Diagnostic Question
“Is this an empty-referent failure, a wrong-referent failure, or valid cognition — because the fix differs in each case.”
IKS Roots
The Sanskrit term is vikalpa (विकल्प). Patanjali’s definition at Yoga-sutra 1.9 reads śabda-jñāna-anupātī vastu-śūnyo vikalpaḥ — cognition that follows from word-knowledge and is empty of object. Vikalpa is one of five vṛttis (modes of mental activity), the others being pramāṇa (valid cognition), viparyaya (mistaken cognition with a real referent), nidrā (sleep), and smṛti (memory) — see card Y1 for the framing of the system as a whole, and Y3 for viparyaya as the partner concept. The canonical example, śaśa-śṛṅga (शशशृङ्ग), “the horns of a hare,” names something whose linguistic form is well-formed and whose extension is empty.
Further Reading
Edwin Bryant, The Yoga Sutras of Patanjali (North Point Press), gives accessible translation and commentary. Christopher Chapple’s work for technical philosophical context. The Yoga-bhashya of Vyasa is the foundational commentary; Vacaspati Misra’s Tattva-vaisaradi extends it.