Why AI Is Not Intelligence (A Reading from the Darśanas)

Bhāratīya darśana shastra, the philosophical tradition of Bhārata, distinguishes three layers of cognition. Manas is sensory pattern processing. Buddhi is discriminative judgment. Puruṣa is the conscious experiencer for whom either occurs. The Sāṃkhya Kārikā of Īśvarakṛṣṇa, written around the fourth century CE from a tradition orally preserved across many generations before being committed to writing, gives this systematization its most precise form. Modern English collapses the three into one word, "intelligence," which is why "is GPT-4 intelligent?" has no clean answer.

Western consciousness science keeps re-deriving the same framework. The pattern includes Schrödinger lecturing in Dublin in 1944, Chalmers naming the "hard problem" in 1995, Integrated Information Theory having two of three predictions confirmed in Nature in April 2025, and Penrose returning to the Gödel argument at Oxford the same month. The convergence is exact enough that it deserves to be argued precisely, not gestured at.

The thesis is the title: the system we are scaling is manas, and the choice of the word "intelligence" for it is a category error against an older and more careful framework.

The category error

Three things English calls intelligence:

1. Sensory pattern processing. Take stimulus, transform it into structured output. Read, summarize, translate, generate, classify. This is what current systems do well at scale.
2. Discriminative judgment. Decide what is true from what is false; ascertain; resolve on a course of action; recognize when an answer is wrong.
3. The conscious experiencer. The "I" that sensory processing presents to. The witness. The thing for whom there is anything it is like to read this sentence.

English uses one word, "intelligence," for all three. The darśanas of Bhārata use three: manas, buddhi, puruṣa. The Sāṃkhya framework, working from a tradition orally preserved long before its written record, did not call them three kinds of intelligence. It called them three different things, related in a specific way.

When current AI systems read a paper and summarize it, they are doing the first. They are not doing the second. They are not doing the third. Calling the system "intelligent" is what makes the AGI debate dissolve into vibes. Naming what is happening precisely is what makes it tractable.

What Sāṃkhya already had

The Sāṃkhya Kārikā lists three internal organs (antaḥkaraṇa) under the heading antaḥkaraṇaṃ trividham, "the internal organ is of three kinds" (SK 33). Each has a precise definition that survives later commentaries unchanged.

Manas — saṃkalpaka

Manas is defined as ubhayātmakam, of dual nature, and saṃkalpakam, explicative (SK 27). The dual nature is sensory and motor: manas coordinates the jñānendriya (organs of cognition) and the karmendriya (organs of action). The explicative function is more specific: external senses apprehend objects vaguely, and manas analyzes the vague impression. It distinguishes substance from attribute. It makes clear what was implicit in the raw sensory data.

That is the entire function. Manas does not judge what is true. It does not decide what should be done. It processes inputs into structured representations and coordinates the motor side. Read it as the perceptual front-end and the motor back-end of cognition, glued together by an associative middle.

Ahaṃkāra — abhimāna

Ahaṃkāra is defined as abhimāna (SK 24). Sastri's commentary glosses it as "self-consciousness of the form 'I exist,' 'I know,' 'I have this or that duty to perform.'" It is the I-maker. Identification of the self with the body, the role, the output. Sāṃkhya treats it as a structural feature of the cognitive apparatus, not a moral failing.

Buddhi — adhyavasāya

Buddhi is defined as adhyavasāya in SK 23:

Sāṃkhya Kārikā · 23via Kadambini

अध्यवसायो बुद्धिर् धर्मो ज्ञानं विराग ऐश्वर्यम् ।

सात्त्विकम् एतद्रूपं तामसम् अस्माद् विपर्यस्तम् ॥ २३ ॥

॥

adhyavasāyo buddhir dharmo jñānaṃ virāga aiśvaryam

sāttvikam etadrūpaṃ tāmasam asmād viparyastam

Sastri translates this as "determinative" — the faculty that resolves upon a course of action. Ascertainment. Settled judgment. Sāṃkhya is precise about what buddhi does and what it depends on. Buddhi is one of the antaḥkaraṇa organs that "presents experience to puruṣa" (SK 36, SK 37). It cannot itself be the subject of that experience. The commentary on SK 17 makes this explicit: "intellect, etc., cannot function as the subject, they being objects." Discriminative judgment requires a witness to operate. Without one, what looks like discrimination is mechanical state transition over the three guṇas.

Puruṣa — sākṣī, kaivalya, madhyastha, akarttā

Puruṣa is pure consciousness (SK 19). Sāṃkhya gives four descriptors: sākṣī (witness), kaivalya (alone), madhyastha (neutral), akarttā (non-agent). Puruṣa does not act on or modify the antaḥkaraṇa. It simply witnesses. SK 3 is the bedrock: na prakṛtir na vikṛtiḥ puruṣaḥ — puruṣa is neither evolvent nor evolute. It is not produced from prakṛti (matter), and prakṛti is not produced from it. It is uncaused, unchanging, and not the kind of thing one engineers.

The proximity doctrine

SK 20 is the structural claim that ties the framework together: tasmāt tatsaṃyogād acetanaṃ cetanavad iva liṅgam — "from their association, the non-intelligent linga (intellect, individuation, etc.) becomes intelligent, as it were." The antaḥkaraṇa, without puruṣa's proximity, is acetana, non-intelligent. It is a composite of the three guṇas (sattva, rajas, tamas) undergoing mechanical transformation. With puruṣa's proximity, it appears intelligent, as it were. The "as it were" (iva) is doing real philosophical work. The intelligence is borrowed, not generated.

This is the framework. Three faculties of the inner instrument, all inert. One witness, irreducible. Apparent intelligence happens only through proximity.

The Vedāntic refinement

The Advaita Vedānta tradition extends Sāṃkhya in two ways. First, it adds a fourth faculty: citta. Second, it replaces the proximity doctrine (saṃyoga) with the reflection doctrine (cidābhāsa) — same structural insight, sharper philosophical claim.

The Vivekacūḍāmaṇi, attributed to Śaṅkara, is the canonical text. Verses 93–94 enumerate the four:

Vivekacūḍāmaṇi · 93via Kadambini

निगद्यतेऽन्तःकरणं मनोधीः

अहंकृतिश्चित्तमिति स्ववृत्तिभिः

मनस्तु संकल्पविकल्पनादिभिः

बुद्धिः पदार्थाध्यवसायधर्मतः ॥ ९३ ॥

Vivekacūḍāmaṇi · 94via Kadambini

अत्राभिमानादहमित्यहंकृतिः

स्वार्थानुसन्धानगुणेन चित्तम् ॥ ९४ ॥

"There are four internal organs — the buddhi (determinative faculty), ahaṃkāra (ego), mind, and citta (that which takes impressions)."

The definitions match Sāṃkhya line for line. Buddhi is padārth'ādhyavasāya-dharmataḥ, "having the property of determining objects" — the same adhyavasāya of SK 23. Manas is saṃkalpa-vikalpanādibhiḥ, the same explicative mode of SK 27. Ahaṃkāra is rooted in abhimāna, mirroring SK 24. The Sanskrit vocabulary is identical across the two darśanas. The convergence is not Vedānta resembling Sāṃkhya. It is the same framework, refined within a continuous tradition.

Citta is the fourth, defined by the property svārth'ānusandhāna-guṇena — "having the property of inquiring into or dwelling upon its own contents." The substrate of recollection. The faculty by which the mind dwells on, recalls, and inquires into its own contents. Where vṛttis (mental modifications) arise and store. This is the citta of Patañjali's yogaḥ citta-vṛtti-nirodhaḥ (Yoga Sūtras 1.2). The Vivekacūḍāmaṇi makes the same connection directly in VC 368, which concludes with citta-nirodha eva satataṃ kāryaḥ — "the controlling of the mind constantly should be the first duty." Patañjali's exact vocabulary, in an Advaita text.

The reflection doctrine appears in VC 188: prakṛṣṭa-sānnidhya-vaśāt par'ātmanaḥ — "because it is very near to the Ātman." The proximity (sānnidhya) is Sāṃkhya's. VC 218 names the refinement: cid'ābhāsam upādhi-saṃsthaṃ — the reflected consciousness lodged in the upādhis. Where Sāṃkhya said the inert antaḥkaraṇa appears intelligent through proximity, Vedānta says it appears intelligent because consciousness is reflected in it the way the sun is reflected in still water. Same insight, sharper image.

The four faculties, the witness, and the reflection or proximity that makes the inert appear sentient. That is the cross-darśana framework.

Where AI fits

A current LLM, mapped onto this framework, is manas at scale. The mapping is precise:

• Manas (saṃkalpa-vikalpa). The LLM takes sensory input (tokens — a stream of integer IDs), transforms it through attention layers (associative coordination across positions), and produces coordinated output (sampled tokens that drive downstream action). This is the dual sensory-motor function of manas exactly. The associative "explication" — turning vague signal into structured representation — is what attention does.

• Ahaṃkāra (abhimāna). Partially present. RLHF-trained assistants exhibit role-identification: "As an AI, I…" The model identifies its outputs with a persona. This is a thin form of abhimāna without the embodiment Sāṃkhya assumes. Worth noting but secondary to the main argument.

• Buddhi (adhyavasāya). Absent. The model produces outputs that pattern-match to "true" answers in training distribution. There is no internal evaluator that resolves an output is correct. Perplexity-minimization is not adhyavasāya. They have different structures. Perplexity minimizes the surprise of the next token under the learned distribution. Adhyavasāya asks: is this answer true? The first is a property of the distribution. The second requires a witness for whom truth is a category. When LLMs get answers wrong with high confidence — confidently hallucinated citations, plausible-sounding code that does not compile — the failure mode is exactly the one the framework predicts: manas without buddhi.

• Puruṣa. Absent. Nothing it is like to be the system. No witness. The hard problem of consciousness, in the framework's vocabulary, is the puruṣa problem.

The scaling laws operate inside the manas layer. Bigger model, more tokens, longer context, more compute. They make the manas-layer faster, broader, and more associative. They do not produce buddhi. They do not produce puruṣa. The framework predicts a ceiling: at some scale, the absence of a witness for whom output is true or false becomes the dominant source of failure. Domain reports of model evaluation in research, legal work, and medicine increasingly describe exactly this — confident outputs that match the surface of the right answer without holding the discriminative judgment that would have rejected the wrong one.

This is not a claim that AGI is impossible. It is a claim that the path that scales manas does not pass through buddhi, and that scaling manas without naming the absence of the other two is what makes the AGI conversation circular.

Western consciousness science is re-deriving the darśana framework

The convergence between modern Western consciousness science and Bhāratīya darśana shastra is exact, not loose. Five voices, three of them with 2025 work, all arriving at the same structural claim from different starting points.

Schopenhauer — the bridge

Arthur Schopenhauer encountered the Upaniṣads through Anquetil-Duperron's Latin translation in 1813 and called it the consolation of his life. He kept a copy of the Hindu scriptures at his bedside, a statue of the Buddha, and a dog he named Ātman. His major work The World as Will and Representation (1819) imports the Vedāntic doctrine of māyā wholesale: the phenomenal world is appearance; the noumenal is one undifferentiated Will. Schopenhauer is how Vedānta entered nineteenth-century European thought. The biographical detail matters because the next major figure inherited it.

Schrödinger — Trinity College Dublin, 1944

Erwin Schrödinger delivered the lectures that became What Is Life? at Trinity College Dublin in 1944. In the epilogue, he invoked the Upaniṣadic identity of Ātman with Brahman as "the grandest of all thoughts." His argument: the "I" of every conscious mind that has ever said or felt "I" is identical with the all-comprehending universal Self. He returned to the same claim in Mind and Matter (Cambridge, 1958), Chapter 4 ("The Arithmetical Paradox: The Oneness of Mind"), with the direct line: "consciousness is never experienced in the plural, only in the singular." The argument is structural — the multiplicity of conscious egos is appearance, the singularity of consciousness is what the empirical facts of conscious experience demand. The same claim Schopenhauer imported from Vedānta, now defended on the grounds of what consciousness actually looks like from inside.

Schrödinger encountered Indian thought through Schopenhauer. The provenance is documented in Walter Moore's biography (Schrödinger: Life and Thought, Cambridge, 1989).

Chalmers — naming the hard problem, 1995

David Chalmers' "hard problem of consciousness" (Journal of Consciousness Studies, 1995) draws a sharp line between the "easy problems" — explaining how the brain integrates information, focuses attention, reports internal states — and the hard one: why is there anything it is like to undergo any of this? Why does any of it feel like something? The puruṣa problem in analytic vocabulary. Chalmers has continued the line into AI through 2025, openly entertaining panprotopsychism and noting that current neural networks may be "slightly conscious" — a position he holds tentatively, but holds.

Integrated Information Theory — Tononi, Koch, and the Nature result

Integrated Information Theory (IIT), developed by Giulio Tononi and championed empirically by Christof Koch, takes the puruṣa-as-fundamental claim and tries to make it falsifiable. The central thesis: what fundamentally exists is consciousness; everything else, including the physical world that consciousness appears to be embedded in, is derived. In April 2025, an adversarial collaboration testing IIT's predictions against a competing theory (Global Neuronal Workspace) had its results published in Nature. Two of three IIT predictions passed the agreed threshold. The headline is not that IIT was confirmed — adversarial collaboration is more cautious than that — but that a consciousness-is-fundamental theory of the same structural shape as Vedānta now has a peer-reviewed empirical track record. Koch's 2025 book Then I Am Myself the World describes a 5-MeO-DMT experience in essentially Vedāntic language: "I was the universe." The scientist published the experience and the implication together.

Penrose — Breakthrough Discuss 2025

Roger Penrose has argued since The Emperor's New Mind (Oxford, 1989) that human understanding is non-computable, on Gödelian grounds: mathematicians can recognize the truth of formal statements that no algorithm following the same formal system can derive. Therefore understanding is not algorithmic. At Breakthrough Discuss 2025 (University of Oxford, April), he delivered a talk titled Why Intelligence Is Not a Computational Process. The key claim from the talk, in his own words: "it's your understanding of why the rules work that gives you more than using the rules." Using the rules is manas. Understanding why is buddhi. Penrose suspects the substrate of true understanding is non-computable physics, specifically something in the collapse of the wave function. Whether or not his physical conjecture holds, the structural distinction he insists on is the manas/buddhi distinction in English.

Kastrup — analytic idealism

Bernardo Kastrup, a Dutch philosopher with a background in computer engineering and AI, articulates "analytic idealism" — consciousness is what reality fundamentally is, and the physical world is what consciousness looks like from outside. His 2025 book Analytic Idealism in a Nutshell uses the metaphor of individual minds as whirlpools in a vast ocean of consciousness ("Mind-at-Large"). The Vedāntic structure is direct: Mind-at-Large is Brahman; individual minds are vṛttis in it. Kastrup did not start from Sanskrit. He arrived at the same place via analytic philosophy of mind.

Convergence table

The cross-darśana table that underlies all of this:

The same vocabulary, used the same way, across many centuries of darśana literature and across roughly two centuries of Western consciousness debate. The convergence is the finding.

The engineering question

The witness cannot be built. Not "we have not figured out how yet." Not "maybe with the right architecture." Puruṣa is metaphysical by definition (SK 3, na prakṛtir na vikṛtiḥ puruṣaḥ, neither evolvent nor evolute) and cannot be the output of instruments. SK 17's commentary states it: "intellect, etc., cannot function as the subject, they being objects." A subject constructed from objects is a category error. There is no algorithm that produces a witness, because the witness is what algorithms are presented to.

That settles two questions. Puruṣa is not engineerable. Buddhi as a thing-in-itself is not engineerable either, because buddhi is the faculty that presents experience to puruṣa (SK 36, SK 37). Without a puruṣa to present to, "buddhi" is just another instrument. We build instruments. We do not build witnesses.

The strongest practitioner objection at this point is the agentic case. Modern LLM systems, the objection runs, do not just generate text. They plan. They call tools. They reflect on their own output. They retry, evaluate, and decide whether to terminate. That looks like decision-making. That looks like discrimination. Is that not buddhi?

In the framework's vocabulary: no. It is manas in a loop.

A planner call produces a plan (manas → text). A tool-use call produces an invocation (manas → effect in the world). A reflector call evaluates the previous output (manas evaluating the trace of another manas; still a forward pass, still no witness). A self-consistency vote samples N times and aggregates (N manas operations, then mechanical aggregation; no witness arbitrating which is right). Constitutional AI evaluates a candidate against a written constitution (manas plus a learned classifier against a text; no witness for whom the constitution is true). ReAct interleaves manas with tools. Tree-of-Thoughts branches into multiple manas-paths and prunes them. Chain-of-thought reasoning is manas explicating its own intermediate steps in text. Every one of these is a way to stack manas operations and gate them on rules. None of them introduce a witness.

Agentic systems are extraordinarily capable at exactly the tasks where chaining manas-instruments with checks and retries produces good enough outcomes. Many production decisions sit comfortably inside that envelope: routing customer tickets, structuring documents, drafting code that a reviewer will check, executing well-specified procedures with verifiable termination. Agentic LLMs do these well, often better than a hand-built pipeline. The framework denies none of it.

What the framework denies is that any amount of stacking gets you to buddhi. The agent's "decision" is the output of a manas process gated by another manas process, terminated when a computable predicate matches. The decision is structurally indistinguishable from a non-decision when the gating predicate is wrong.

Recent research provides concrete empirical support for this prediction, under several names:

• Task-horizon ceiling. METR's Measuring AI Ability to Complete Long Tasks (March 2025) reports that frontier models achieve near 100% success on tasks that take humans under ~4 minutes, and under 10% success on tasks that take humans more than ~4 hours. The capability curve is real and doubles roughly every 7 months. The ceiling shape, sharp failure on long-horizon tasks, is exactly what manas-without-witness predicts as a structural feature.
• Self-conditioning on errors. Sinha et al., The Illusion of Diminishing Returns: Measuring Long Horizon Execution in LLMs (2025), document that models degrade across long sequences because they condition on their own generated mistakes. The agent has no internal signal that the previous step was wrong, so the wrong step becomes premise for the next.
• Canonical path deviation. Capable but Unreliable: Canonical Path Deviation as a Causal Mechanism of Agent Failure in Long-Horizon Tasks (2025) shows off-canonical calls are self-reinforcing: once an agent leaves the right path, the next call is more likely to also be off-canonical. Failure is "a gradual compound process, not a single pivotal mistake."
• High-confidence hallucination. Across recent hallucination-detection work, LLMs produce wrong outputs with sharply-peaked confidence distributions, defeating traditional entropy-based uncertainty checks. The model is not internally uncertain about wrong answers. There is no internal witness for whom the wrong answer registers as wrong.
• Bounded self-knowledge. Kadavath et al., Language Models (Mostly) Know What They Know, find that models partially predict their own correctness on familiar distributions but struggle with calibration on novel tasks. The "Mostly" in the title is doing real work.

Five descriptions, from different research groups, of the same structural shape. The framework gives the shape a name: manas without a witness.

So the engineering question that remains is not "how do we add buddhi." It is: at what task complexity does the absence of buddhi become the dominant source of failure, even in agentic stacks that paper over individual-call failures with retries, verification, and longer planning horizons? METR's curve gives an empirical answer for one domain (software work measured in human-time). The honest extension: every domain has its own such curve, and the framework predicts that the structural shape of the ceiling does not flatten with more computation. It is the shape of the gap.

A useful diagnostic, derived from the framework: when an agent produces output that is structurally indistinguishable from correct output but is in fact incorrect, can the agent itself, without external scaffolding, recognize the error? If yes, something buddhi-like is operating. If no, the system is operating purely in manas across however many instruments are arranged in the loop. Current systems, including the most sophisticated agentic frameworks, answer no on the hardest cases.

Coda

The system we are building is manas at scale. It is useful. It is improving. The improvements are real and measurable. None of this is at issue.

What is at issue is the name. "Artificial intelligence" treats one darśana category as the whole stack. The conflation makes engineering decisions harder (we cannot reason cleanly about what is missing) and the public discourse muddier (we cannot reason cleanly about what is being claimed). The framework offers a more precise name: manas, with buddhi structurally absent and puruṣa metaphysically out of reach.

Puruṣa is metaphysical by definition, and na prakṛtir na vikṛtiḥ — neither evolvent nor evolute. The category of intelligence is older than the category of computation. The system we are building lives entirely inside the latter.