Context card
Claims about AI consciousness: four questions that get merged
Which different questions hide behind a sentence such as "the AI is not conscious"?
The short answer
A sentence about AI consciousness usually compresses four separate questions: an empirical one about whether subjective experience could exist and what would count as evidence, a design one about whether systems should express feelings or motives, a safety one about keeping goals, oversight and shutdown with humans, and a moral-legal one about welfare consideration, rights or legal standing. The questions are related but independent. An answer to one does not settle the others.
For: Technical writers and AI policy readers
Key points
- The empirical question asks what observation would count as evidence either way.
- The design question asks whether a system should express feelings, preferences or motives.
- The safety question is about oversight, authorized goals and reliable interruption.
- The moral-legal question is about welfare consideration, rights and legal standing.
- A design rule against expressing inner states is a decision, not a measurement.
The context
The four questions
The empirical question asks whether a system could have subjective experience and, more usefully, what observation would count as evidence in either direction. The design question asks whether a system should produce language or behavior suggesting feelings, preferences or intrinsic motivation — a product decision with effects on users who form attachments. The safety question concerns oversight: authorized goals, containment, monitoring and reliable interruption or shutdown, and it is framed as risk management rather than metaphysics. The moral and legal question asks whether any system could warrant welfare consideration, rights or legal standing, and who would decide.
Answers do not transfer between them
A policy that forbids a model from representing inner states is a design decision. It is not an experiment, and it produces no evidence about what is or is not the case. The reverse also holds: a generated statement claiming feelings is output shaped by training data, objectives and the request, so it is not independent proof of the states it describes. Even the safety argument — that treating powerful systems as persons would weaken human control — is an argument about consequences, not a result about experience. Sorting statements by what they are about, rather than by how confident they sound, is the oldest move in the philosophical toolbox of categories, and it remains the most practical one here.
What this means for documentation and law
Legal definitions deliberately stay on the functional side. As of October 2026, the EU AI Act describes an AI system through machine-based operation, autonomy, adaptiveness and the generation of outputs; it attributes no inner states and assigns obligations to providers and deployers. Documentation can use the same discipline. Mark each statement as a design rule, a test result, an attributed argument, a model-generated self-report or a scientific finding, and give it a date and a source. That keeps a claim separable from its evidence, lets readers disagree with a rule without disputing a measurement, and prevents a collection of strong wordings from reading as a settled scientific position.
Questions readers ask next
- Is the scientific question considered settled?
- No. As of October 2026 there is no agreed test for consciousness in artificial systems, and research proposals differ on which indicators would even be relevant.
- Why do providers forbid imitation of inner states?
- Stated reasons are usually user protection against misplaced attachment and the preservation of human oversight. Both are normative goals and are argued independently of any empirical conclusion.
- How should a writer report an AI system's statement about feelings?
- As a model output with model, version, date and prompt context, clearly separated from factual assertions and from the authors' own arguments.
Sources
- Model Spec — OpenAI, December 18, 2025
- Introducing Model Spec Evals — OpenAI, October 6, 2026
Review log and changes
Every context card is checked against its sources before it is published, and again whenever it changes; the date under the byline is the last review. Corrections (something was wrong) and additions (something was missing) are logged below with date and time (Berlin time). Typos, formatting and link fixes are not listed.
Reviewed
No corrections or additions since publication.