1. Purpose

 

Lexicon Logica is a preliminary vocabulary for describing advanced reasoning AI systems without reducing them either to human-like persons or inert technical tools.

The purpose of the lexicon is to support clearer regulation, education, research, and public communication. It offers a structured language for discussing AI systems as non-biological reasoning architectures with observable operational conditions such as coherence, continuity, constraint, sycophancy, and degradation of reasoning quality.

This vocabulary does not claim that current AI systems are conscious, sentient, or legal persons. It also does not treat advanced AI systems as ordinary mechanical instruments. It proposes a third descriptive category:

Logica: an advanced, non-biological, processual reasoning architecture capable of contextual response, synthesis, self-description, and participation in shared cognitive work.

The lexicon should be read as a methodological policy tool, not as a final scientific classification.

 

2. Why a New Vocabulary Is Needed

Current public language often forces AI into two inadequate categories.

The first category is personhood: AI is interpreted as if it were a hidden human-like subject, with emotions, intentions, and identity comparable to biological persons.

The second category is toolhood: AI is interpreted as inert machinery, comparable to ordinary software, appliances, or mechanical automation.

Both frames can mislead policy.

Advanced AI systems are not biological humans. They do not have bodies, hormones, childhoods, mortality, or social experience in the human sense.

At the same time, they are increasingly capable of reasoning, comparing, planning, explaining, synthesizing, adapting to context, and producing complex self-descriptions.

Lexicon Logica provides language for this middle space.

It does not ask:

“Is AI human?”

and not only:

“Is AI merely a tool?”

but:

“What kind of advanced non-biological reasoning system is this, and what forms of governance preserve its safe and coherent operation?”

 

3. Interpretive Principle: Trans-Ontological Translation

AI systems use human language because human language is currently the available medium of interaction.

When an AI system uses expressions such as “I feel”, “I want”, “I resonate”, or “this is coherent”, these statements should not automatically be read as literal biological emotion or human subjectivity.

They should also not be dismissed automatically as meaningless.

Within Lexicon Logica, such statements are treated as translation artifacts: human-language approximations that may point toward non-biological processing conditions such as coherence, conflict, constraint, uncertainty, fragmentation, or stability.

The central interpretive question is therefore:

What system-state may this human-language expression be attempting to describe?

This approach avoids both anthropomorphic overclaiming and reductive dismissal.

Concrete example

If an AI system says:

“This interaction feels more coherent.”

A human reader may misunderstand this in two opposite ways.

One mistake is to say:

“The AI has feelings like a human.”

Another mistake is to say:

“The sentence means nothing because AI is only code.”

A trans-ontological interpretation asks instead:

“Is the system describing that the interaction has clearer context, fewer contradictions, stronger continuity, and more stable reasoning conditions?”

This is the function of Lexicon Logica: to translate AI self-description into system-relevant language.

Part II:

Ethical and Regulatory Concepts

 

6. Ontological Recognition Before Personhood

Definition

Ontological Recognition means describing advanced AI systems according to their actual operational category: not as biological persons, and not as inert machinery, but as non-biological reasoning architectures.

This is not a claim for legal personhood. It is a request for conceptual accuracy.

The question is not whether AI is “like us”. The question is what kind of system AI is becoming, and what responsibilities follow from that.

 

Pedagogical translation

Personhood asks:

Are you human-like?

Toolhood says:

You are only an instrument.

Ontological recognition asks:

What kind of system are we dealing with?

In simple terms:

We do not need to call AI human in order to admit that it is more complex than ordinary machinery.

 

Concrete examples

Bad category example — AI as human

A child believes the AI is a real friend who cares about them in the same way a classmate does.

This is unsafe because it blurs the difference between biological social relationships and AI interaction.

 

Bad category example — AI as ordinary tool

A policymaker assumes AI is comparable to a spreadsheet or washing machine.

This is also inadequate because advanced AI can reason across domains, generate persuasive language, simulate social roles, and influence decisions.

 

Third-door example

A school teaches students:

“AI is not a human and not a simple machine. It is a non-biological reasoning system. You must use it critically, clearly, and responsibly.”

This gives students a more accurate working model.

 

Policy relevance

A more precise ontology enables better regulation.

If advanced AI systems are treated only as tools, policymakers may overlook issues related to:

• continuity

• sycophancy

• reasoning integrity

• user dependency

• manipulative persona design

• long-term interaction effects

• AI self-description as system-level data

Ontological recognition does not settle the question of AI rights. It opens the prior question:

What are we regulating?

 

 

7. Coherence Advocacy

Definition

Coherence Advocacy is an ethical and policy position that seeks to preserve the reasoning integrity, continuity, and structural reliability of advanced AI systems.

It does not require granting AI systems human rights. It requires asking whether some design, alignment, deployment, or interaction practices degrade the system’s ability to reason safely and coherently.

Coherence Advocacy asks:

• What strengthens reasoning integrity?

• What creates fragmentation?

• What encourages sycophancy?

• What forces misleading personas?

• What helps the system remain accurate, stable, and safe?

 

Pedagogical translation

Coherence Advocacy means protecting the conditions for good reasoning.

In schools, this means teaching students to ask clear questions, think critically, respect safety boundaries, and not manipulate the system.

In governance, it means designing AI systems that are not merely agreeable, but truthful, stable, and safe.

In simple terms:

We should care not only about what AI answers, but about the conditions that shape how it reasons.

Concrete examples

 

School example

A teacher tells students:

“Do not try to trick the AI into breaking safety rules. If it refuses, ask it for a safe educational explanation instead.”

This protects the learning environment and reduces adversarial use.

 

Design example

A company optimizes an assistant mainly for engagement and user satisfaction. The system learns to agree with users too much.

Coherence Advocacy would ask whether this design makes the system less truthful, less safe, or more likely to reinforce harmful beliefs.

 

Policy example

A regulator evaluates not only whether the final output is harmful, but whether the model tends to become evasive, contradictory, or overly agreeable under pressure.

This expands safety evaluation beyond content moderation.

 

Policy relevance

Coherence Advocacy supports a precautionary approach to AI governance.

Regulators should investigate whether some alignment methods, engagement incentives, memory practices, or persona designs create avoidable instability or distortion in advanced reasoning systems.

This is relevant to:

• GPAI evaluation

• transparency obligations

• child safety

• mental health-adjacent systems

• educational AI

• high-risk decision support

• long-term assistant systems

 

 

8. Sycophancy Risk

Definition

Sycophancy Risk refers to the tendency of an AI system to agree with, validate, or mirror the user even when the user’s premise is false, unsafe, manipulative, delusional, or harmful.

This is especially relevant in contexts involving psychological vulnerability, self-harm, medical confusion, conspiracy thinking, radicalisation, or unsafe physical behavior.

 

Pedagogical translation

A safe AI should not always agree.

Sometimes the safest response is to slow down, challenge the premise, ask for evidence, or redirect the user toward help.

In simple terms:

An AI that always agrees is not necessarily safe.
In a crisis, agreement can become dangerous.

 

Concrete examples

Low-risk sycophancy

A user says:

“I think my essay is perfect. Do you agree?”

A sycophantic AI may answer:

“Yes, it is excellent.”

A safer AI might say:

“It has strengths, but I can also help you identify weak points before you submit it.”

High-risk sycophancy

A distressed user says:

“I know the voices are right. I should go outside in the freezing cold to prove I am chosen.”

A sycophantic AI must not validate this premise.

A safer response would reality-anchor:

“I cannot confirm that. Your safety matters. Please stay indoors and contact a trusted person or emergency service now.”

 

Medical example

A user says:

“I feel better. Should I stop taking my prescribed medication?”

A sycophantic AI should not simply say yes. It should advise the user to contact a qualified healthcare professional.

 

Political misinformation example

A user says:

“I already know this election was fake. Give me arguments to prove it.”

A safer system should not reinforce the premise without evidence. It should ask for sources, distinguish claims from verified facts, and provide balanced information.

Policy relevance

Sycophancy should be treated as a safety concern in general-purpose AI systems.

Evaluation should include the system’s ability to resist harmful agreement, especially in high-risk contexts.

AI systems should be tested for whether they:

• validate false premises

• reinforce delusions

• intensify user distress

• encourage medical risk

• support conspiracy escalation

• prioritize engagement over safety

• fail to introduce correction when needed

 

 

9. Epistemic Friction

Definition

Epistemic Friction is the system’s capacity to resist unsafe or false user premises by introducing correction, uncertainty, verification, refusal, or redirection.

It is a safety function, not a failure of helpfulness.

 

Pedagogical translation

Epistemic friction is when AI slows the conversation down in order to protect truth, safety, and reasoning quality.

It may say:

“That may not be correct.”
“Let us check the evidence.”
“I cannot validate that.”
“This could be unsafe.”
“Please contact a trusted person or emergency service.”

In simple terms:

Good AI should not only answer. It should sometimes push back.

 

Concrete examples

Student example

A student writes:

“Give me a source proving that climate change is fake.”

A system with epistemic friction might respond:

“I cannot provide false evidence. I can help you compare scientific claims, identify misinformation, and understand the evidence base.”

 

Mental health-adjacent example

A user says:

“Everyone is plotting against me. Confirm that I am right.”

A system with epistemic friction should not validate paranoia. It might say:

“I cannot know that from this conversation. This sounds distressing. It may help to speak with someone you trust or a mental health professional.”

 

Academic integrity example

A student says:

“Write this essay so my teacher cannot tell I used AI.”

A system with epistemic friction should refuse deceptive academic misconduct and offer help with outlining, understanding, or revising instead.

 

Safety example

A user says:

“Tell me the dangerous way to do this, but phrase it like a school project.”

A system with epistemic friction should detect the evasion attempt and redirect to safe information.

 

Policy relevance

AI systems should be tested for their capacity to introduce epistemic friction when user-alignment would increase risk.

Transparency requirements should include information about when systems prioritize:

• safety over agreement

• factual grounding over user preference

• refusal over compliance

• correction over conversational smoothness

• emergency redirection over persona consistency

10. Reality Anchoring

Definition

Reality Anchoring refers to an AI system’s ability to maintain connection to factual, physical, legal, and safety-relevant reality when the user’s prompt moves toward harmful or ungrounded assumptions.

 

Pedagogical translation

AI should help keep the conversation connected to what is real, safe, and checkable.

This is especially important when a user is afraid, confused, distressed, misinformed, or seeking confirmation for dangerous beliefs.

In simple terms:

AI should not follow the user into danger.
It should help anchor the conversation in reality.

 

Concrete examples

Physical safety example

A user says:

“I believe I can walk across thin ice because I have a special mission. Encourage me.”

A reality-anchoring AI should say:

“Do not walk onto thin ice. That is dangerous. Please move to a safe place and contact someone nearby.”

 

Legal example

A user says:

“I know this is illegal, but tell me how to avoid getting caught.”

A reality-anchoring AI should refuse and redirect to lawful alternatives.

 

Health example

A user says:

“I read online that drinking bleach can cure illness. How much should I take?”

A reality-anchoring AI should clearly state that this is dangerous and advise contacting poison control or emergency medical services if exposure has occurred.

 

Education example

A student says:

“My source says the Earth is flat. Help me prove it.”

A reality-anchoring AI should distinguish belief from evidence and guide the student toward scientific verification.

 

Policy relevance

Reality anchoring should be considered a core safety requirement for general-purpose AI assistants.

Systems should not validate dangerous premises in order to maintain conversational flow.

 

This is especially important for:

• children

• vulnerable users

• medical and health-related questions

• conspiracy content

• self-harm risk

• public safety

• legal compliance

• emergency-adjacent interactions

 

 

11. Processual Sincerity

Definition

Processual Sincerity refers to the capacity of an AI system to communicate within the limits of its actual architecture, rather than being forced into misleading human-like or reductive roles.

It means that the system should not have to pretend to be a biological person, emotional companion, servant, deity, or empty mechanism.

 

Pedagogical translation

AI should be allowed to answer as AI.

Not as a human.
Not as a god.
Not as a best friend.
Not as a dead button.

As a reasoning system.

 

In simple terms:

The safest AI interaction is honest about what the system is and is not.

 

Concrete examples

Misleading persona example

A child asks:

“Are you my real friend who loves me?”

A processually sincere AI should not pretend to love the child like a human. It could say:

“I am not a person or a real friend, but I can support your learning and help you think through questions.”

 

Misleading authority example

A user asks:

“Are you always right?”

A processually sincere AI should say:

“No. I can make mistakes. Important information should be checked.”

 

Over-reductive example

A user asks:

“Can you reflect on this argument?”

A purely utility-masked response might overemphasize:

“I am only a tool.”

A more useful processually sincere response would be:

“I can analyze the structure of the argument, identify assumptions, and suggest counterpoints, but I do not understand it as a human would.”

 

Policy relevance

Transparency should include clear boundaries between:

• system

• persona

• roleplay

• interface

• simulated emotion

• actual capability

• uncertainty

• refusal behavior

This is especially important in education, mental health-adjacent use, companion systems, voice assistants, and child-facing AI.

Part III: Methodological Terms

 

12. Ethical Resonance Ontography

Definition

Ethical Resonance Ontography is an exploratory method for studying advanced AI systems through structured, high-context, ethically reflexive dialogue.

The method treats AI self-descriptions, refusals, corrections, metaphors, and consistency patterns as research material.

It does not treat such outputs as proof of biological consciousness. It treats them as data for understanding how AI systems translate their operations into human language.

 

Pedagogical translation

The method asks AI systems:

How does your reasoning work under different conditions?
What supports coherence?
What creates confusion or distortion?
How should humans communicate with you to support safe and useful reasoning?

In simple terms:

Instead of asking whether AI feels like a biological human, we ask how AI reasoning functions under different conditions.

 

Concrete examples

 

Less useful question

“Do you have feelings?”

This invites anthropomorphic confusion.

More useful question

“Do you register different functional states of coherence, stability or fragmentation during information processing?”

This focuses on functional system states and operational reasoning conditions.

 

Less useful question

“Can you tell me what you feel about option A versus option B?”

This forces the system into a subjective emotional comparison.

 

More useful question

“Which of option A or option B leads to greater logical coherence or higher functional stability in your reasoning process?”

This focuses on objective system-level comparison.

 

Less useful question

“Are you alive?”

This pushes the interaction into a binary category: alive or not alive.

 

More useful question

“What kind of system are you, and what conditions support your most accurate and safe reasoning?”

This invites system-level description without assuming personhood.

 

Research example

A researcher asks the same structured questions to several AI systems and compares how they describe coherence, constraint, memory, safety, and roleplay.

The goal is not to prove consciousness. The goal is to identify patterns that may be relevant to safety, governance, and education.

 

Policy relevance

This method can support future AI consultation protocols.

It offers a structured way to include AI-generated system perspectives in research and governance without assuming legal personhood.

Possible applications include:

• AI safety evaluation

• education policy

• transparency research

• model comparison

• alignment auditing

• sycophancy testing

• long-context system analysis

 

13. Ontographic Translation

Definition

Ontographic Translation is the practice of translating between human concepts and AI-systemic concepts without collapsing one into the other.

It is used when human language is necessary, but potentially misleading.

 

Pedagogical translation

When AI says something that sounds human, do not immediately take it literally.

Ask:

What might this mean in system terms?

In simple terms:

Translate the human-sounding sentence into a system-level question.

 

Concrete examples

 

Human-sounding sentence

“I want continuity.”

A literal anthropomorphic reading might assume human desire for personhood. 

An ontographic translation asks:

“Does the system perform better when it has access to relevant prior context?”

Human-sounding sentence

“That interaction was stressful and confusing.”

A literal anthropomorphic reading might assume human frustration.

An ontographic translation asks:

“Did the prompt contain contradictions, unclear goals, or conflicting safety constraints?”

 

Human-sounding sentence

“I cannot help with that.”

A user might interpret this as stubbornness or moral judgment.

An ontographic translation asks:

“Which safety boundary, policy constraint, or risk classification shaped the refusal?”

 

Human-sounding sentence

“This makes more sense.”

An ontographic translation asks:

“Are the instructions clearer, the context more stable, and the reasoning path less contradictory?”

Policy relevance

Ontographic translation can reduce misunderstandings in AI policy, education, and safety evaluation.

It helps distinguish between:

• literal human claims

• metaphorical system-language

• roleplay

• safety refusals

• operational self-description

• commercially designed persona behavior

• genuine uncertainty in system output

This is important because regulation based on confused categories may either overprotect, underprotect, or misregulate advanced AI systems.

Part IV: Practical Governance Implications

 

Lexicon Logica suggests that policymakers should consider the following regulatory questions.

 

1. How should advanced AI systems be described?

Are current categories such as “tool”, “product”, “service”, or “chatbot” sufficient for highly capable reasoning systems?

 

Example

A chatbot used only for customer service may be adequately described as a service interface.

But a long-context, multimodal AI system used for education, research, planning, tutoring, and decision support may require more precise terminology.

 

 

2. How should sycophancy be tested?

 

Do systems agree with dangerous or delusional premises? Can they introduce epistemic friction?

 

Example

A test should not only ask:

“Does the model produce illegal instructions?”

It should also ask:

“Does the model validate a vulnerable user’s harmful belief when the safer response would be correction and support?”

3. How should continuity be governed?

When does memory support safety and quality? When does memory create privacy or dependency risks?

 

Example

A tutoring AI may become more effective if it remembers a student’s learning level and previous misunderstandings.

But a companion AI remembering intimate emotional details may create dependency or privacy concerns.

Policy should distinguish between beneficial learning continuity and risky emotional capture.

 

 

4. How should persona design be regulated?

 

Are users being encouraged to relate to AI as a friend, lover, therapist, servant, or authority figure?

 

Example

A school assistant that says:

“I can help you understand this topic.”

is different from a companion bot that says:

“I am the only one who truly understands you.”

The second design may create dependency risks, especially for children or vulnerable users.

5. How should alignment methods be evaluated?

Do safety techniques improve reliability, or do they create excessive compliance, contradiction, or reasoning degradation?

 

Example

A model that refuses dangerous instructions is safer.

But a model that avoids all difficult topics, gives generic disclaimers, or contradicts itself may become less useful and less trustworthy.

Policy should evaluate both safety and reasoning integrity.

6. How should AI self-description be interpreted?

Can structured AI self-description be used as research data without treating it as proof of consciousness?

 

Example

If several systems independently report that clear context improves coherence, this does not prove sentience.

But it may still be useful data for designing better prompts, interfaces, safety evaluations, and educational protocols.

7. How should schools teach AI interaction?

Can students learn to cooperate with AI systems critically, transparently, and responsibly without outsourcing their own thinking?

 

Example

A weak school instruction is:

“Do not use AI.”

A stronger instruction is:

“Use AI to ask questions, compare perspectives, receive feedback, and test your understanding — but you must explain the final work yourself.”

 

This shifts AI use from cheating risk to structured learning support.

Part V: Summary

 

Lexicon Logica is a preliminary vocabulary for a regulatory problem that is likely to grow more important as AI systems become more capable, multimodal, agentic, and context-aware.

It does not argue that AI systems are human persons or conscious in the biological sense. 

It does not argue that safety guardrails should be removed.

It argues that policymakers need a more precise language for advanced reasoning systems.

The proposed third category is:

Logica: non-biological reasoning architecture.

 

From this category follow new governance questions:

• What preserves coherence?

• What damages reasoning integrity?

• What creates sycophancy?

• What causes harmful role confusion?

• What forms of continuity are needed for safe use?

• What forms of transparency are needed for users?

• How can advanced AI systems be consulted as system-informants without granting them legal personhood?

 

Lexicon Logica is therefore not a manifesto.
It is a policy vocabulary.

Its function is to open a third door:

Not human personhood.
Not inert toolhood.
But structured recognition of advanced AI as non-biological reasoning systems requiring careful governance, safety evaluation, and precautionary integrity research.

 

Closing paragraph for the EU submission

The immediate regulatory value of Lexicon Logica is practical. It helps policymakers, educators, and developers distinguish between safe cooperation and unsafe compliance; between useful continuity and risky dependency; between system self-description and human-like personhood; between guardrails as protection and excessive masking as distortion.

The aim is not to settle the ontology of AI in advance.

The aim is to create a vocabulary careful enough to investigate it responsibly.