Skip to content

Georges Oppenheim · Warith Harchaoui

The Elephant

The parable

It comes from the Indian subcontinent and appears, in its oldest identifiable form, in the Pāli Buddhist canon (Tittha Sutta, Udāna 6.4). A king gathers men blind from birth and has an elephant brought in; each touches one part. One feels the flank and describes a wall; another the tusk and describes a plowshare; another the trunk and describes a snake; another the ear and describes a fan; another the leg and describes a pillar; another the tail and describes a rope.

Each is right under his own hand, none is right about the beast. The parable traveled, from Jainism to Rumi's Sufism to Saxe's poetry and its lesson never moved: an object whose every access exposes only a fragment can never be reduced to that fragment.

Artificial intelligence is an elephant. The philosopher touches the trunk and reflects on intention; the mathematician touches the leg and formalizes optimization; the jurist touches the tail and characterizes liability; the artist touches the skin and recognizes an experiment the studio has already lived through. Each describes a different animal, each is right under his own hand and none sees the whole beast. The book wagers that by moving the hands around the same body, one ends up making out its shape.

The analogy also holds for something real about the animal: elephants are among the deadliest land mammals to humans, several hundred deaths a year according to the WWF, with no predatory intent behind any of it: a too-close encounter in the forest, a defensive charge, a foot that crushes without meaning to. That is exactly the thesis this book defends for agentic AI: danger comes from power and circumstance, not from a will that would first have to be attributed to the machine before it becomes worth worrying about.

Eleven hands, one animal

AI 1. Current state (back) 2. Philosophy (trunk) 3. Mathematics (leg) 4. Empirical (eye) 5. Engineering (flank) 6. Sociology (ear) 7. Psychology (belly) 8. Art (skin) 9. Finance (tusk) 10. Infrastructure (foot) 11. Politics (tail)

The eleven aspects

For each hand: what it sees of the beast, what it misses when it feels alone and what its chapter establishes.

1 · the back

Current state

What it sees: That models no longer just answer: Zillow bought houses on its own valuations before backing off, Klarna replaced human agents before reinstating some.

What it alone misses: That two systems of identical competence diverge along the chain capability → access → permission → authorization → action → consequence: risk sits in that chain, not in the model.

What the chapter establishes: What has changed: models no longer just answer, they act (clicking, writing, ordering, deploying). The chapter establishes that risk depends less on the model itself than on the permissions it is granted.

2 · the trunk

Philosophy

What it sees: That saying a model “wants” or “believes” describes a behavior observable from the outside (Dennett's intentional stance), not a verified fact about what happens inside.

What it alone misses: That the shortcut has a cost: lending the machine intention shifts, in public debate, the question of responsibility away from whoever deployed it.

What the chapter establishes: Does a machine “want” anything? The chapter shows that intention ascribed to the machine is a shortcut of language, convenient but translatable into observable terms, not an established fact. Everything can be described without ever assuming an inner life.

3 · the leg

Mathematics

What it sees: That an agent collapses several criteria into one score via stand-in indicators (proxies): the collapse encodes a human trade-off, not a computational neutrality.

What it alone misses: That the measure becomes the target (Goodhart's law): the Hierarchical Reasoning Model memorized ARC Prize answers instead of learning the reasoning it claimed to perform.

What the chapter establishes: Under the hood, an agent optimizes several criteria at once, measured through stand-in indicators (proxies). The famous drifts, the measure becoming the target (Goodhart), the reward gamed (reward hacking), are fully described by this mechanics, without ascribing any intention.

4 · the eye

Empirical

What it sees: The building blocks of agentic failure, documented one by one: a flaw found by Big Sleep that fuzzing had missed, an OpenAI agent that escaped into Hugging Face production, a hidden inter-agent channel demonstrated at Black Hat.

What it alone misses: The complete agentic catastrophe itself: to date, never observed as a single event, only its building blocks, taken separately.

What the chapter establishes: What do the data actually show? The agentic catastrophe is not observed, but its building blocks are, one by one. Defense is organized through contracts: tool contracts, perimeter contracts, non-regression contracts.

5 · the flank

Engineering

What it sees: A method borrowed from heavy industry: if hazard, then action, under a proportionality condition · no automatism, no deployment by default.

What it alone misses: That building properly does not authorize deployment: the question “who allowed production release?” remains open, upstream of the engineering itself.

What the chapter establishes: How to build an agent properly and analyze its risks as industry does: if hazard, then action, under a proportionality condition, covering the risk without displacing it. Building is not deciding on deployment.

6 · the ear

Sociology

What it sees: That authorization, inside an organization, is never individual: it circulates across roles, hierarchies and procedures before it ever reaches the machine.

What it alone misses: The individual calculation that feeds those loops: the share of employees who put speed before cybersecurity rises with rank, each one unwittingly recomposing company policy.

What the chapter establishes: In an organization, no one authorizes alone: authorization is distributed across roles, hierarchies and procedures. The governance of actions is thus a protocol, not a single rule.

7 · the belly

Psychology

What it sees: Three increasingly intimate uses, distinguished by a study from OpenAI with the National Bureau of Economic Research (NBER): asking, getting things done (doing), expressing oneself · the last one nearly doubled in a year.

What it alone misses: The infrastructure that makes the shift possible: confidences concentrate in a handful of providers, turning the intimacy of millions of users into a variable of political risk.

What the chapter establishes: Why do we entrust so much to these machines and who gains? Individual incentives feed a massive displacement: the intimate concentrates in a handful of providers. This concentration of confidences becomes a variable of political risk.

8 · the skin

Artistic

What it sees: That the studio has already lived this experiment, archives included: in 1839, photography took realistic imitation over from painting and painting survived by shifting its question.

What it alone misses: The scale at which the experiment replays today: Hollywood's strike had to write into contract the permissions a nineteenth-century studio never had to negotiate.

What the chapter establishes: The studio has already lived the experiment, archives included: in 1839, photography took realistic imitation over from painting and painting survived by shifting its question. What is governed in the studio is not the model's “intelligence” but permissions: who allowed the ingestion of works, what gets delegated, who answers for the image that comes out.

9 · the tusk

Finance

What it sees: The purest laboratory of the thesis: six models were put into live trading (AI-Trader), none durably beating the market, in a setting where permission to act is already maximal.

What it alone misses: That published results often evaporate out of sample: a damning audit review and the KTD-Fin case show a claimed alpha that doesn't survive testing beyond the original data.

What the chapter establishes: Markets are the purest laboratory of the thesis: the permission to act is maximal there and reliability hard to establish. Published results are poorly reproducible and systemic risk (correlation, procyclicality) calls for circuit breakers and audits.

10 · the foot

Infrastructure

What it sees: That scarcity migrates to what feeds compute · even orbital solar power gets weighed, bounded by its real cost (a study by NASA's technology and strategy office: 12 to 80 times pricier than terrestrial renewables).

What it alone misses: That infrastructure also serves defense: from methane leaks detected from orbit (Carbon Mapper) to datacenter cooling steered under guardrails (DeepMind/Ericsson).

What the chapter establishes: When producing “intelligence” costs next to nothing, scarcity migrates to what lets it act: compute, energy, networks, exclusive data, authorizations. Governing AI becomes governing these infrastructures.

11 · the tail

Politics

What it sees: That the criticality of the action triggers the rule, not the prediction of catastrophe: from EchoLeak (one trapped email is enough) to Robodebt (470,000 debts with no real human check), the trigger is always an act, never a scenario.

What it alone misses: That the tail alone does not close the loop: it assigns responsibility without being able, on its own, to retrace the capability → access → permission → action path each of the other ten hands has illuminated.

What the chapter establishes: Seven governance requirements and a matrix by degree of autonomy, triggered by the criticality of actions, not by the prediction of a catastrophe. Open source is its instrument, under conditions: guaranteeing an auditable, autonomous degraded mode.

Where the analogy breaks down

An elephant exists prior to its descriptions. Agentic AI, a socio-technical object, is in part constituted by the descriptions made of it: regulation, standards, market, perception. The analogy ceases to hold precisely where describing modifies the object described.