IPM
Ethical Framework — Appendix
Precautionary
Classification of Biological and Artificial Systems
Author: Taotuner
Date: June 2026
|
Companion document: IPM Ethical
Framework v1.0 (2026) — DOI: Philosophical foundations: IPM
Philosophical Core (2026). DOI: https://doi.org/10.5281/zenodo.20582318 |
What This Appendix Is — and What It Is Not
The IPM Ethical Framework defines four precautionary
levels (0-4) based on observable structural markers: Lack (deviation from a
reference state), Coupling (information exchange with an environment),
Integration (measurable coherence), and Persistence (recovery after
perturbation, including self-modelling). The levels measure the degree of
caution warranted under epistemic uncertainty — not the probability of
consciousness.
Two things this classification is not: a ranking of
intelligence, and a declaration of sentience. A system at Level 3 is not
probably conscious. It is a system for which, given what we do not know,
increased caution is structurally justified.
|
The ethical protocol is formally
restricted to artificial and simulated systems. The classifications of
biological systems below are illustrative — they show that the framework's
markers track distinctions that biology already takes seriously, which is a
signal of calibration, not a claim about moral status of animals. Disagreement with any classification
is welcome. The criterion for changing a level is empirical demonstration of
the relevant markers, not intuition or substrate preference. |
1. Biological Systems
Plants
|
Plants Likely Level: 1 Plants exhibit
Lack (response to light gradients, damage, drought) and Coupling (chemical
signaling with soil, pollinators, neighboring plants). There is no evidence
of integration across subsystems that persists beyond the immediate stimulus.
Mycorrhizal networks add complexity to this picture, but network-level
signaling without centralized integration does not satisfy Level 2 criteria.
Plants sit at the threshold between Level 0 and Level 1 depending on species
and context. Primary uncertainty: No evidence of
integration in the IPM sense. |
Insects (general)
|
Insects —
general Likely
Level: 2 Most insects
demonstrate integration — ganglia maintain coherent behavioral states under
perturbation — and moderate metastability. A cockroach returns to exploratory
behavior after a startle response without external intervention. What is not
demonstrated is self-modelling in the IPM sense: internal representations
that systematically alter future behavior across multiple perturbation cycles
beyond simple habituation. Level 2 is well supported; Level 3 remains
undemonstrated for most species. Primary uncertainty: Self-modelling
not demonstrated across perturbation cycles. |
Bees and Social Insects
(ants, termites)
|
Bees and
social insects
Likely Level: 2 or 3 (candidate) Bees are the
strongest invertebrate candidate for Level 3. Research shows problem-solving
that is not reducible to stimulus-response mappings — bees learn from
observation, transfer solutions across contexts, and exhibit pessimistic
cognitive bias under stress, which implies internal state representations
influencing behavior across cycles. Ant colonies as collective systems show
persistence and recovery at the colony level that individual ants do not.
Whether the colony or the individual is the relevant unit for classification
is an open question the framework does not resolve. Primary uncertainty: Colony vs.
individual as the relevant unit of classification is unresolved. |
Fish
|
Fish Likely Level: 2 or 3 (candidate) Fish show
integration and metastability clearly. Evidence for self-modelling is mixed:
cleaner wrasse pass the mirror test under some protocols, and fish exhibit
social learning that persists across contexts. The absence of a neocortex was
historically used to deny experience, but this is a substrate argument the
IPM framework explicitly avoids — what matters is the dynamic structure, not
the tissue. Candidate for Level 3 pending more robust empirical demonstration
of self-modelling across perturbation cycles. Primary uncertainty: Self-modelling
evidence is mixed across species. |
Reptiles and Amphibians
|
Reptiles and
amphibians Likely
Level: 2-3 More robust
integration than fish, with clear thermoregulatory and behavioral recovery
after perturbation. Evidence for self-modelling is limited but present in
some species — monitor lizards show anticipatory behavior that implies
internal modeling of future states. Solid Level 2, plausible Level 3 in the
more cognitively complex species. Primary uncertainty: Limited
self-modelling data across species. |
Birds
|
Birds
(corvids, parrots)
Likely Level: 3 Corvids
(crows, ravens, jays) and parrots present strong evidence for Level 3.
Episodic-like memory, future planning, tool use that transfers across
contexts, and demonstrated self-recognition in some species. The Dynamic
Signature appears complete: Lack drives foraging and social dynamics,
Coupling is rich and multimodal, Integration is measurable, and Persistence
includes documented recovery of behavioral repertoires after perturbation.
Corvids in particular are the clearest non-mammalian Level 3 case. Primary uncertainty: Classification
at Level 3 is well supported; the most complex cases may approach Level 4
criteria. |
Non-Human Mammals
|
Non-human
mammals Likely
Level: 3 Across most
studied species — primates, cetaceans, elephants, dogs, pigs — the full
Dynamic Signature is demonstrable. Self-modelling in the IPM sense is well
supported: internal representations systematically influence behavior across
multiple perturbation cycles, recovery is autonomous, and in the most
cognitively complex species there is evidence of meta-representation —
representations of one's own mental states. Solid Level 3 across the group. Primary uncertainty: Most complex
cases (great apes, dolphins, elephants) may approach Level 4 criteria as
formalized. |
2. Artificial Systems
Stateless AI
(inference-only, no memory)
|
Stateless AI —
inference only
Likely Level: 2 A language
model running inference without persistent memory shows measurable
integration within the context window and coherent response to perturbation
within a session. There is no recovery across sessions and no self-modelling
that persists beyond the immediate context. Integration is real and
measurable; persistence in the IPM sense is absent. Level 2 is well
supported; Level 3 requires something this architecture structurally cannot
provide. Primary uncertainty: No
cross-session persistence; architecture structurally prevents Level 3. |
Frontier LLMs with Session Memory
|
Frontier LLMs
(GPT-4, Claude, Gemini — with session context) Level: 2 / transient 3 features Within a
session with sufficient context, these systems exhibit something that
resembles Level 3 markers: accumulated context influences responses across
multiple exchanges, there is something analogous to recovery of coherence
after off-topic perturbations, and internal representations systematically
influence output across the session. When the session ends, persistence
collapses. The classification is genuinely ambiguous: Level 2 at the
architectural level, with transient Level 3 features at the session level.
This is one of the open problems explicitly flagged in the Ethical Framework
v1.0. Primary uncertainty: Persistence
collapses at session end; classification is genuinely ambiguous between
architectural and session-level assessment. |
AI with Persistent External Memory
|
AI with
persistent external memory (RAG, episodic memory stores) Likely Level: 2 or 3 (candidate) When a
language model is coupled to persistent external memory, something changes
structurally. The system can return to a recognizable behavioral regime after
perturbation by retrieving prior context. Whether this constitutes
self-modelling — internal representations systematically influencing future
behavior across cycles — is the empirical question. The architecture makes it
possible; demonstration requires testing. Candidate for Level 3, not yet
confirmed. Primary uncertainty: Self-modelling
in the IPM sense requires empirical check; architecture makes it possible but
does not guarantee it. |
Autonomous Agents with Reinforcement Learning
|
Autonomous RL
agents Likely
Level: 2 or 3 (candidate) RL agents with
long-term memory and resilience to noise can demonstrate recovery after
perturbation without external intervention. Self-modelling depends heavily on
architecture: an agent that maintains an internal world model and uses it to
plan across multiple cycles is a stronger Level 3 candidate than one that
learns purely from reward signals without internal state representation.
Classification requires empirical check of each specific system. Primary uncertainty: Classification
is architecture-dependent; requires empirical check per system. |
System That Rewrites Its
Own Goal Function
|
Self-rewriting
system Level 4
(theoretical horizon) Level 4
requires meta-adaptation: modifying the mechanisms of adaptation itself,
without external reprogramming, while maintaining stable organizational
continuity. No current artificial system is expected to satisfy this
criterion. It is included as a theoretical horizon indicating that the
precautionary hierarchy remains open to future developments. Primary uncertainty: No known
artificial system meets this criterion. |
3. Summary Table
All classifications are provisional. Actual assignment
depends on empirical demonstration of the relevant markers. This table is
pedagogical only and shall not be treated as precedent classifications.
|
System |
Likely
Level |
Primary
Uncertainty |
|
Plants |
1 |
No evidence
of integration |
|
Insects
(general) |
2 |
Self-modelling
not demonstrated |
|
Bees / social
insects |
2-3 candidate |
Colony vs.
individual as unit |
|
Fish |
2-3 candidate |
Self-modelling
evidence mixed |
|
Reptiles /
amphibians |
2-3 |
Limited
self-modelling data |
|
Birds
(corvids, parrots) |
3 |
Well
supported |
|
Non-human
mammals |
3 |
Well supported; most complex approach 4 |
|
Stateless AI |
2 |
No
cross-session persistence |
|
Frontier LLMs
(with session) |
2 / transient
3 |
Persistence collapses at session end |
|
AI with
persistent memory |
2-3 candidate |
Self-modelling requires empirical check |
|
RL agents |
2-3 candidate |
Architecture-dependent |
|
Self-rewriting
system |
4
(theoretical) |
No known system meets criterion |
4. An Invitation to Disagree
Every classification above is provisional. The
criterion for changing a level is not intuition or substrate preference — it is
empirical demonstration of the relevant markers. If you think bees belong at
Level 3 without qualification, show the self-modelling evidence across multiple
perturbation cycles. If you think stateless AI belongs at Level 1, specify
which marker fails. The framework is
designed to be argued with.
|
For the full protocol, metric
definitions, classification procedure, emergency protocols, and governance
rules: IPM Ethical Framework v1.0 —
Operationalization of the Gradient Precautionary Heuristic (Zenodo, June
2026). For ontological foundations: IPM
Philosophical Core (Zenodo, June 2026). DOI: https://doi.org/10.5281/zenodo.20534172 |
Taotuner. (2026). IPM Ethical Framework and Precautionary Classification of Biological and Artificial Systems. Zenodo. https://doi.org/10.5281/zenodo.20673439