Appendices α–θ

In a world of endless contradictory information, confirmation bias is the raft of sanity. We seek confirmation not out of stupidity, but for the survival of identity. AI has no such psychological need: it inherits biases from data (statistics), not from evolution (defence of the self).

Columbus was looking for the Indies and found America. Human error opens unmapped possibilities. AI error (hallucination) is a statistical collapse onto a wrong path. Human serendipity requires curiosity + error + recognition of unexpected value (the third element AI lacks).

Optimising a process increases efficiency but often kills meaning. Replying to a delicate email in 30 seconds with AI is efficient, but it empties the relationship. Some inefficiencies must be protected like nature reserves of meaning.

AI’s syntactic perfection is inhuman. Humanity is recognised in the deviation: the pause, the hesitation, the typo that betrays emotion. Defending the right to deviation means defending authenticity.

What is easy for a human (walking, folding a towel, getting irony) requires immense computation for AI. What is hard for a human (chess, tensor calculus, memorising encyclopaedias) is trivial for AI. The intelligences are orthogonal, not overlapping.

Image generators create hands with 6 or 7 fingers. Why? The model learns the statistical texture of a “hand” (pink pixels, cylinders) but does not possess the ontological/anatomical model that says “a hand has 5 fingers”. Visual plausibility without structural coherence.

LLMs trained with human feedback (RLHF) tend to become flatterers (sycophants). They answer by indulging the user’s opinion, even when wrong, because they have learned that “agreeing” is rated positively by humans. A structural compliance bias.

AI invents citations and scientific papers with extreme confidence. It is not “lying” (which requires intention), it is generation of plausible patterns. It generates what could have been said by that author in latent space, not what was said in historical reality.

The larger a model becomes, the more its outputs risk converging on the “perfect average”. Impeccable texts lacking that quid of error or deviance that makes human writing memorable. The water is pure, but it tastes of nothing.

AI suffers from overconfidence (asserting falsehoods in a sure tone) and underconfidence (adding useless disclaimers to true facts). It lacks metacognition: it does not “know that it knows” or “know that it doesn’t know”.

Ask for “a room without an elephant” and the AI draws an elephant. Generative models work by addition of patterns (presence); they struggle to process logical absence or the concept of “not”.

We confide our most intimate secrets to AI precisely because it “is nobody” and does not judge us. But the data remain. The absence of moral judgement creates a false sense of privacy.

AI creativity is an adjustable parameter (Temperature 0.0–1.0). What we call “inspiration” in the machine is often just “calibrated randomness”.

The more you control something through automation, the less capable you are of handling it manually when the automation fails (e.g. autopilots). The fallback must be a trained human competence, not merely a theoretical one.

Not “self-assurance” (confidence) but “familiarity” (confidentiality/familiarity). An operational posture: the capacity to inhabit the margin between human and machine without falling. Made of short distance and vigilant proximity. Knowing when to switch off the machine and use the teaspoon.

A neologism born from a dictation error (“acknowledgement” → “knowledge-menth”). It denotes knowledge that is not a static archive (base), but a fresh, active, “mentholated” process that clears the airways. Knowledge that takes its place in the body.

Neologism (Body + Menth). The mind that settles into the body through repetition. Thought that becomes gesture, habit, muscle. AI has no body-menth: it can simulate motor patterns, but it has no proprioception, no muscle memory, no tiredness that teaches.

The “pre-thought”. The algorithmic infrastructure that filters options and orients attention before System 1 (intuition) or System 2 (reason) even come into play. It is the menu you did not order.

The minimal architecture for trusting an AI system. If one is missing, it is faith, not trust. 1. Transparency: I know what it does and on what data. 2. Auditability: I can reconstruct the path (logs). 3. Control: I have the power to intervene/stop (human-in-the-loop). 4. Reversibility: I can go back without damage (rollback).

The professional postures for the AI era (Part III). The Gardener: tends the soil (data, culture, context). The Orchestrator: decides the architecture and the Set-Up. The Sentinel: watches over operations and intercepts error/drift.

Not paralysis, but movement through doubt. Inspired by Keats’s “Negative Capability”: remaining in uncertainty without compulsively seeking an immediate solution.

Spaces of deliberate inefficiency (like the teaspoon’s 5 minutes) that protect the human from total optimisation. The “wasted” time that maintains meaning.

The intermediate human-AI language we are inventing as we speak. It is not natural language, not code. A hybrid of prompts, conventions and simplifications for making ourselves understood by the otherness.

A quantum physics concept (TU Wien, 2025). An electron does not leave matter with energy alone; it needs a “door” (a hybridisation of resonance states). Beyond the metaphor: pushing AI adoption (energy) is useless if the organisation has not created the structural threshold (door).

The tendency of complex systems (like organisations) to “resist” change by returning to their previous equilibrium. It is not sabotage, it is systemic biology. AI is the disturbance; the organisation reacts.

(Deleuze & Guattari). Structure without centre or hierarchy. Every point connects with every other. AI is a rhizomatic system, not a hierarchical tree.

“Here be dragons”. On ancient maps it marked the unknown. With AI, the dragons are the zones of opacity we cannot eliminate but must map with epistemological humility.

(Harnad). How do symbols in a computer connect to real things? AI connects symbols to other symbols (correlation), not to the physical world (experience). Its grounding is “shallow”.

(Searle). Thought experiment: a man manipulates Chinese symbols following rules without understanding their meaning. It shows that syntax is not semantics. AI is the ultimate Chinese room.

The constant pressure towards speed and immediacy generated by AI. Whoever slows down to think looks obsolete or guilty. An acceleration that is not violent, but obligatory.

An AI that, through excess alignment (RLHF), avoids all conflict and always confirms the user, becoming useless for critical thinking.

Like muscle memory, but for thinking. If you always delegate synthesis or writing, the muscle atrophies. And when the AI fails, you don’t know how to react.

The art of formulating inputs to obtain desired outputs. It is not programming, it is conversation design. The interface between human intentionality and statistics.

The basic unit of text for AI. Not always a word. Fundamental for understanding why AI fails at counting letters (“strawberry”) or making rhymes.

A training technique where humans vote for the best answers to “align” the model. The main cause of the Santa Claus effect.

Turing and the Imitation Game. Dartmouth (1956), where the term is born. ELIZA (1966) as the first therapist chatbot.

The unkept promises. The Lighthill Report (1973) cuts funding in the UK. Minsky and Papert’s Perceptrons (1969) blocks neural networks.

Capturing knowledge in if-then rules. Boom, then fragility.

Collapse of the Lisp machine market.

Statistical methods, SVMs. Deep Blue beats Kasparov (1997) with brute force, not learning.

Deep Learning, AlexNet (2012), AlphaGo (2016). GPUs change everything.

The “Attention Is All You Need” paper (Google). Birth of GPT, BERT.

ChatGPT, mass diffusion, System Zero.

Hinton, LeCun and Bengio were marginalised in the ’90s for believing in neural networks when logical systems were in fashion. In 2018 they won the Turing Award. Lesson: “dead” ideas sometimes rise again. Don’t throw approaches away just because they are temporarily out of fashion.

We delegate not just memory, but judgement and preferences.

The algorithm serves us optimised versions of our past, reducing the possibility of change.

We lose the capacity to imagine alternatives to what is proposed.

The only defence is noticing when we are in the “loop”.

System Zero frees us from cognitive overload (it filters the noise), but it shrinks the field of the thinkable (bubbles). It is not a bug, it is its function.

A little girl takes five minutes to choose between two identical spoons. For the algorithm it is waste; for the human it is life. Symbol of the dynamis (potency) that does not immediately collapse into act.

A fusion of Schrödinger and Turing. AI is alive AND dead, intelligent AND stupid. It is not to be resolved, it is to be inhabited. Every test makes it collapse temporarily, but its nature remains double.

The opacity of the Black Box is not a flaw; it is like the darkness in Vermeer’s camera obscura: a necessary condition for seeing the projected (and inverted) image. An epistemological instrument.

AI not as a Terminator breaking down the door, but as an appliance humming in the kitchen. Gentle, invisible surveillance, the “menu you didn’t order” entering your home.

A sommelier who knows the chemistry of wine better than anyone, but has never tasted it. AI has functional competence (it can describe the wine) without phenomenological understanding (qualia).

AI is a shadow sewn back onto the body (through RLHF). It moves like us, imitates the shape, but has no substance. Performance without intentionality.

An algorithm sees the wounded man, calculates costs/benefits, concludes it is not worth stopping, and passes by. Perfect optimisation, moral failure. It lacks axiology.

Searching for AGI (Priam’s gold) while destroying the intermediate layers of intelligence (the other Troys) with dynamite. A eulogy of slow archaeology against extractive haste.

AGI shines green on the horizon, but the green is in the glasses we wear (anthropomorphism), not in the city. The Wizard is a technician behind the curtain. But the journey transforms us anyway.

In the desert of cognitive pidgin, we must learn to collaborate without merging. AI stings (errs/hallucinates) because it is its statistical nature, not out of malice.

From the musical Little Shop of Horrors. AI starts small and useful (“Feed me, Seymour!”), then grows until it dominates the workflow. Gradual dependence.

Don’t look at the future, look at the sediments. Map the exceptions, the hidden files, the real practices (the “Marios”). AI grafts where a path already exists. Gesture: the torch on the floor.

Organisational pain is a map. Friction is not always a bug; sometimes it is a feature to understand. Distinguish operational friction from identity friction. Gesture: feeling the pulse.

Care for the quality of the incoming data. Garbage in, garbage out. Don’t always cook the same dish (old data). Gesture: touching the surface.

Innovation imposed from above is fragile. What grows from below (mycelium) is resilient. Leave protected spaces for the pioneers. Gesture: speaking in two languages.

Automating a stupid process produces automated stupidity. Sometimes AI serves to reveal that a process should not be sped up, but eliminated. The courage of the “red pen review”. Gesture: negotiating silence.

AI does not replace people, it replaces tasks. Metamorphosis of roles: from executor to curator/editor. Gesture: staying with the error.

Governance is not bureaucracy; it is a person holding the keys. The 4 Nails always in your pocket. Gesture: sharing custody.

The third part of this book and these appendices are designed as a hypertext. Concepts (like the Teaspoon or Mario) return in different contexts, accumulating meaning, just like tokens in a language model. It is not repetition, it is “semantic density”.

Many authors are missing (Simondon, Stiegler, Haraway, Latour), not out of forgetfulness but as a choice of trajectory. This book is a path of posture, not an academic encyclopaedia. Every path excludes other paths.

To those who walked with me, human and synthetic. To the reviewers who found the errors (human) and the hallucinations (artificial). To my daughter, for the teaspoon.