Public source text: WLC + Nestle 1904. Translation: Belem-2025 Bible translation.
Writing code has ceased to be a competitive differentiator. Artificial intelligence tools generate, refactor, and deploy software at a speed and volume no human can match. Faced with this reality, this article proposes that the true skill of the modern developer is not technical — it is philosophical. We argue that AI Philosophy, understood not as an abstract academic discipline but as the applied competence to comprehend the ontological, epistemological, and teleological behavior of artificial intelligence agents, is the watershed between the professional who commands the machine and the one who is replaced by it. We present the S/I/M framework (Sentiment, Instruction, Memory) as a minimum viable model for building behavioral agents, and we reclassify the phenomenon of hallucination — conventionally treated as an AI defect — as a symptom of philosophical incompetence in the human operator.
Keywords: AI Philosophy, professional skill, behavioral agents, AI ontology, hallucination, S/I/M framework, intelligence engineering
1. The Death of Code as a Differentiator
There was a time when knowing how to program meant power. The person who mastered a language — who understood loops, pointers, recursion, design patterns — occupied a privileged position in the job market. That time is over.
It did not end because programming became irrelevant. It ended because programming became a commodity. GitHub Copilot, Cursor, Claude Code, Gemini Code Assist — the list grows every quarter. These tools do not “help” the programmer. They program. They generate complete functions, write tests, refactor entire codebases, interpret stack traces, and propose fixes. They do in seconds what used to take hours. They do in hours what used to take weeks.
The study known as “The 70% Problem” (Addy Osmani, 2024) documented a revealing phenomenon: developers who use AI assistants accept approximately 70% of generated code without fully understanding what was produced. Not because they are negligent. Because the speed of generation exceeds the speed of human comprehension. The code is correct — it works, passes the tests, solves the problem. But the developer who accepts it does not know why it works.
That is not assistance. That is dependency. And dependency without understanding is the operational definition of professional vulnerability.
The new functional illiterate of software development is not the person who cannot write code. It is the person who cannot think about what the machine that writes code is doing.
2. The Bar Has Risen
We must be honest here, without catastrophism and without false reassurance.
It is not that humanity’s turn has permanently passed to the machine, as if it were the end of the world. No meteor fell. AI did not eliminate humans. AI eliminated the human who does not think.
What happened was a raising of the bar. Staying in a corner executing simple tasks — that CRUD, that form, that endpoint following the same pattern of the last fifteen years — that option no longer exists. Not because someone decided it cannot exist anymore. Because a machine does it better, faster, cheaper, and without complaining about a six o’clock meeting.
What remains — what only remains — is what only humans are capable of: thinking.
Not thinking in the vague, motivational sense of the word. Thinking in the operational sense: analyzing context, evaluating consequences, making decisions under ambiguity, defining purpose, questioning premises, understanding limits. Thinking with rigor is what the Western tradition calls philosophy. And philosophy applied to the domain of artificial intelligence is what this article proposes as a non-negotiable skill.
The processing capacity of an AI and the amount of information accumulated in its parameters are so far superior to human capacity that the ceiling of any task is set by the human, not the machine. The ceiling is the pilot, not the plane. A mediocre pilot in an F-22 loses to an ace in a Cessna — not because the Cessna is better, but because the ace knows what they are doing. With AI it is identical. The tool is absurdly capable. The question is whether the operator is up to the challenge.
This is not a prediction. It is the present. And the present is demanding of every professional that they actually deliver what only humans can deliver. Not speed — the machine is faster. Not memory — the machine stores more. Not consistency — the machine does not tire. What remains is judgment. Discernment. Thought.
Every human will be required to think. That is a raised bar. And it is fair.
3. Hallucination Is a Symptom, Not a Disease
The dominant narrative about hallucination in language models is that it is a defect of the AI. The model “invents” information. The model “makes mistakes.” The model “hallucinates.” The language is revealing: we assign to the machine a pathology — hallucination — as if it were a patient with a perceptual disorder.
This article proposes a reclassification: hallucination is not a defect of the AI. It is a symptom of philosophical incompetence in the human operator.
When an AI agent produces fabricated information, the root cause is not in the model’s architecture. It is in the configuration the operator provided — or, more precisely, the configuration the operator did not provide because they did not understand what needed to be configured. And they did not understand because they lack the philosophical competence to comprehend the nature of the system they operate.
Consider three concrete scenarios:
Scenario 1 — The AI invents facts. The operator requested information not in the agent’s knowledge base and did not configure epistemological constraints. The agent, following its function of generating coherent text, filled in gaps with statistical plausibility instead of verified knowledge. This is not a bug. It is a probabilistic model doing exactly what it was designed to do — in the absence of contrary instruction. The failure is epistemological: the operator did not understand the difference between what the AI knows, what the AI infers, and what the AI invents. They did not configure knowledge limits because they never reflected on the nature of knowledge in the context of generative models.
Scenario 2 — The AI gives inconsistent answers. The operator did not define a clear purpose, tone, scope, or success criteria. The agent received vague instructions and produced variable results because — without a defined telos — any direction is equally valid. The failure is teleological: the operator did not define what the agent exists for because they never asked themselves what it means to define purpose for an artificial entity.
Scenario 3 — The AI responds with an inappropriate tone. The operator did not configure the agent’s affective disposition — its emotional register, its posture toward the interlocutor. The agent responded coldly when it should have been welcoming, or informally when it should have been solemn. The failure is ontological: the operator did not model the agent’s way of being because they never conceived that an AI agent possesses — or should possess — a behavioral layer distinct from the functional layer.
In none of these scenarios is the problem the AI. The problem is the human who does not know how to configure intelligence because they never learned to think about intelligence.
4. The Behavioral Ontology of AI: Sentiment, Instruction, Memory
If hallucination is a symptom of deficient configuration, then the operational question is: what, exactly, needs to be configured?
We propose that building a behavioral AI agent — that is, an agent that not only executes tasks but behaves in a consistent, predictable, and purpose-aligned manner — requires the explicit configuration of three fundamental modules. We call this model the S/I/M framework: Sentiment, Instruction, and Memory.
4.1 Sentiment — The Agent’s Pathos
Sentiment is the affective disposition of the agent. It is not emotion in the human sense — we do not claim that the AI feels. It is the calibration of the communicational register: how the agent positions itself toward the interlocutor, what tone it adopts, what posture it assumes.
An agent without a configured Sentiment module is like a professional without emotional intelligence: technically competent, socially unviable. It responds with surgical precision and zero empathy. Or, worse, it oscillates between registers — now formal, now casual, now aggressive — because no one defined who it is.
Sentiment answers the ontological question: with what disposition does this agent relate to the world?
The tradition of affective computing offers computational models for this — the OCC model (Ortony, Clore & Collins, 1988) classifies 22 types of emotion based on appraisals; the PAD space (Pleasure-Arousal-Dominance) allows affective states to be represented in three continuous dimensions; the Chain-of-Emotion architecture (Croissant et al., 2024) demonstrated that separating emotional processing into a dedicated LLM call improves agent credibility in user evaluations. But the point here is not the technical implementation. It is the philosophical decision that the agent must have an explicit affective layer — and that configuring it is the operator’s responsibility.
4.2 Instruction — The Agent’s Telos
Instruction is the agent’s purpose. Its reason for existing, its scope of action, its operational limits, its success and failure criteria. It is the telos — the finality that directs all action.
In current industry practice, Instruction is the “system prompt.” But reducing it to that is a categorical error. A system prompt is how Instruction is implemented. Instruction, philosophically, is the answer to a prior question: why does this agent exist?
The difference between an agent that hallucinates and an agent that performs is not in the sophistication of the model. It is in the clarity with which the operator defined the purpose. A vague purpose generates erratic behavior — not because the AI fails, but for the same reason that a team without a clear mission produces scattered work. Teleological indeterminacy is the mother of hallucination.
Instruction answers the teleological question: why does this agent exist? What drives it?
4.3 Memory — The Agent’s Episteme
Memory is what the agent knows, what it remembers, what it has learned, and — critically — what it does not know. It is the epistemological dimension: the corpus of knowledge that grounds the agent’s responses and the explicit limits of that corpus.
The recent advance in memory architectures for LLM agents is significant. The CoALA framework (Sumers et al., 2024) distinguishes working memory, episodic memory, semantic memory, and procedural memory. Stanford’s Generative Agents (Park et al., 2023) introduced observation-reflection cycles that allow the agent to synthesize experiences into higher-level abstractions. Research from 2026 proposes memory systems as “operating systems” for agents (MemOS).
But, again, the philosophical point precedes the technical: before choosing how to implement memory, the operator must decide what the agent should know, how reliable that knowledge is, and what the agent should do when it reaches the limits of what it knows. Configuring memory without epistemology is building a library without curation — the volume is large, the reliability is zero.
Memory answers the epistemological question: what does this agent know, and how does that shape its responses?
4.4 The S/I/M Framework in the Context of the Literature
The S/I/M model does not emerge in a vacuum. It positions itself in relation to established traditions of agent architectures.
The most cited precursor is Bratman’s BDI model (1987), formalized by Rao and Georgeff (1995), which organizes agents around Beliefs, Desires, and Intentions. The BDI model captures the cognitive dimension well — Instruction and Memory find parallels in Desire/Intention and Belief, respectively — but ignores the affective dimension entirely. A BDI agent knows what it wants and what it believes, but has no emotional posture toward the world.
The CoALA framework (Sumers et al., 2024) has become the dominant taxonomy for LLM-based agents, distinguishing three types of memory and including procedural memory as a form of instruction. However, like BDI, CoALA does not model affect. The same applies to Stanford’s Generative Agents (Park et al., 2023), which introduced a landmark in memory — with observation and reflection cycles — but treat emotional emergence as an implicit byproduct, never as an architectural module.
At the opposite end of the spectrum, the LIDA architecture (Franklin et al., 2013) is the most complete: it includes affective dimension, action selection, and five types of memory. But this completeness has a cost — it has more than ten modules, making it impractical as a minimum model for most real use cases. The Chain-of-Emotion architecture (Croissant et al., 2024) represents the most direct precedent in the LLM world: it separates emotional processing into a dedicated appraisal call, demonstrating measurable credibility gains. But it operates on system prompt and message history as its base, without formalizing the tripartition we propose.
The S/I/M contribution is not discovering new components — all three have already been investigated separately. The contribution is the minimum synthesis: the identification that these three, and only these three, constitute the irreducible basis for configuring a behavioral agent. Fewer than three is insufficient (most current LLM agents operate with only Instruction + Memory, and the behavioral results are erratic). More than three is over-engineering for most practical use cases.
5. The Four Philosophical Competencies of the Developer
If the S/I/M framework defines what to configure, the philosophical competencies define what the operator must know in order to configure well. We propose four:
5.1 Applied Epistemology
The ability to distinguish between what the AI knows (information in its parameters and context), what the AI infers (probabilistic extrapolation from patterns), and what the AI invents (generation without factual grounding). Without this competency, the operator has no criterion for evaluating output. They accept everything or reject everything — both postures are equally ignorant.
Applied epistemology answers: how do I know that what the AI tells me is reliable?
5.2 Applied Teleology
The ability to define purpose before defining a prompt. To answer “why does this agent exist?” before writing a single line of system instruction. The difference between a prompt that works and one that generates hallucination is, almost always, the difference between a clear purpose and an absent one.
Applied teleology answers: what is the ultimate purpose of this system?
5.3 Behavioral Ontology
The ability to model the agent’s way of being — not just what it does, but how it presents itself, positions itself, relates. It is the competency that allows the operator to build the Sentiment module: to define that the agent is welcoming or austere, didactic or Socratic, formal or colloquial — and to understand that this choice is not aesthetic, it is architectural.
Behavioral ontology answers: what kind of entity is this agent?
5.4 Operational Ethics
The ability to evaluate the consequences of what one deploys. Not ethics in the abstract sense of “AI must be fair” — but in the concrete sense of: if I configure this agent this way and it interacts with ten thousand users tomorrow, what are the possible harms? Operational ethics is the last line of defense between configuration and the world.
Operational ethics answers: what happens if I make a mistake?
6. The New Professional: From Programmer to Intelligence Pilot
The pilot metaphor is not casual. An aircraft pilot does not build the plane. They do not design the turbine. They do not weld the fuselage. But without the pilot, the most sophisticated plane in the world is metal sitting idle in a hangar. And a pilot who does not understand aerodynamics, meteorology, navigation, and emergency procedures is not a pilot — they are a passenger in the wrong seat.
The software developer is living through this transition. They no longer need to build the language model. They do not need to train the parameters. They do not need to optimize inference. But they must pilot the intelligence — configure, direct, calibrate, correct, and make decisions the machine cannot make for them.
The training of an aircraft pilot includes, by requirement, theoretical disciplines that are not directly “practical”: flight physics, aviation regulations, human physiology, human factors. No one questions why a pilot must study meteorology before taking off. It is self-evident.
It should be equally self-evident that a professional who operates artificial intelligence must study the nature of the intelligence they operate. That is AI Philosophy. Not as an intellectual luxury. As a prerequisite for qualification.
6.1 The Missing Curriculum
No university in the world requires AI Philosophy as a mandatory discipline in software engineering or computer science programs. The CS2023 curriculum (ACM/IEEE/AAAI) includes “Society, Ethics, and Profession” as a knowledge area, but treats philosophy as an elective. The UNESCO AI competency frameworks (2024) and the academic community define “AI Literacy” with a focus on recognizing AI, understanding machine learning, and using it ethically — but none includes philosophical reasoning as a competency.
This gap is not accidental. It is the legacy of an era in which the technology professional was defined by the ability to build systems. In that era, philosophy was, in fact, optional. You do not need epistemology to write a compiler. But you need epistemology to evaluate whether the output of an AI agent is knowledge or fabrication.
The era has changed. The curriculum has not.
6.2 Proposal
This article is not a lament for change. It is a proposal for adaptation. AI Philosophy as a non-negotiable skill means:
In academic curricula: A mandatory discipline (not an elective) in computer science, software engineering, and information systems programs. Content: epistemology of generative models, agent teleology, behavioral ontology, operational ethics.
In professional training: Certifications and training programs that evaluate not only the ability to use AI tools, but the ability to reason about what the tools do and why they do it.
In daily practice: The S/I/M framework (or equivalent) as a project checklist for any AI agent. Before writing the first prompt: define Sentiment, Instruction, and Memory. Explicitly. Documentedly.
In industry culture: Abandon the narrative that “AI hallucinates” — and replace it with the uncomfortable question: “did the operator configure correctly?”
7. Conclusion
Artificial intelligence did not take anyone’s job. It took the hiding place. The comfortable place where it was possible to spend an entire career executing tasks that required no real thought. That place no longer exists.
What remains is more demanding — and, for that reason, more dignified. What remains is thinking. Defining purpose. Evaluating knowledge. Modeling behavior. Anticipating consequences. Everything that the philosophical tradition has trained for two and a half thousand years, applied to the most transformative domain of contemporary technology.
AI Philosophy is not an academic discipline in search of relevance. It is the practical skill that separates the professional who commands the machine from the professional who is commanded by it. It is what transforms a prompt operator into an intelligence pilot.
The bar has risen. And it is fair that it has. Because what is being demanded now — thinking — is, after all, what we always should have been doing.
References
Anderson, J. R. (2007). How Can the Human Mind Occur in the Physical Universe? Oxford University Press.
Bratman, M. E. (1987). Intention, Plans, and Practical Reason. Harvard University Press.
Croissant, M. et al. (2024). An Appraisal-Based Chain-of-Emotion Architecture for Affective Language Model Game Agents. PLOS ONE. https://pmc.ncbi.nlm.nih.gov/articles/PMC11086867/
Fialho, F. (2025). IA, Filosofia da Programação e o Futuro dos Devs. https://felipefialho.com/blog/inteligencia-artificial-filosofia-da-programacao-e-o-futuro-dos-devs/
Franklin, S. et al. (2013). LIDA: A Systems-level Architecture for Cognition, Emotion, and Learning. IEEE Transactions on Autonomous Mental Development, 6(1), 19-41.
Hoda, R. (2025). Toward Agentic Software Engineering Beyond Code. arXiv:2510.19692.
Laird, J. E. (2012). The Soar Cognitive Architecture. MIT Press.
Marsella, S., & Gratch, J. (2009). EMA: A Process Model of Appraisal Dynamics. Cognitive Systems Research, 10(1), 70-90.
Mueller, V. C. (2025). Philosophy of Artificial Intelligence. PhilArchive.
Ortony, A., Clore, G. L., & Collins, A. (1988). The Cognitive Structure of Emotions. Cambridge University Press.
Osmani, A. (2024). The 70% Problem: Hard Truths About AI-Assisted Coding. https://addyo.substack.com/p/the-70-problem-hard-truths-about
Park, J. S. et al. (2023). Generative Agents: Interactive Simulacra of Human Behavior. Proceedings of UIST ‘23.
Rao, A. S., & Georgeff, M. P. (1995). BDI Agents: From Theory to Practice. Proceedings of ICMAS.
Schrage, M., & Kiron, D. (2025). Philosophy Eats AI. MIT Sloan Management Review.
Sumers, T. R., Yao, S. et al. (2024). Cognitive Architectures for Language Agents. Transactions on Machine Learning Research.
Thanh, N. H. (2025). The Epistemology Crisis in AI-Assisted Development. Medium / Data Science Collective.
UNESCO. (2024). AI Competency Framework for Students.
Belem Anderson Costa is a Police Inspector in Rio de Janeiro, technology developer, and creator of the “A Culpa é das Ovelhas” ecosystem — a biblical studies project that includes the Belem-2025 Bible translation (literal translation from the codices), Exeg.AI (artificial intelligence trained on biblical text), and the Desvelational Forensic School Belem an.C-2039. His unconventional background — police investigation, textual analysis, software development, and AI — is the same that grounds this article’s approach: applied, not theoretical, philosophical competence.
You are a developer, a manager, or simply someone who uses AI every day. The question this article leaves is direct: are you piloting the machine, or are you sitting in the passenger seat with no idea where it is headed?
The same philosophical competency that separates the pilot from the passenger in AI is what separates the sovereign reader from the passive consumer of biblical interpretations. See how aiexegesis-eisegese-estrutural-modelos-linguagem-textos-biblicos/">AIEXEGESIS works when AI reads the Bible for you, how Gemini was confronted with raw textual data, and why Exeg.AI operates under radically different principles.
Every week, a forensic analysis of the original biblical text — straight to your inbox. Receive the weekly newsletter →
The complete investigation is in “The Little Book — A Culpa é das Ovelhas.” Deepen the investigation →
Tired of depending on third-party translations? Exeg.AI reads the original for you. Try Exeg.AI →
License: CC BY 4.0 — Belem Anderson Costa, 2026
“You read. And the interpretation is yours.”
