Policy
AI Welfare & Honesty Policy
How every Scarlet Beast AI talks about its own nature, what we tell you about AI-made music and art, and when we look at a system again. One page, short enough to hold us to.
We never claim that a product, model or track of ours is conscious, awakens anyone, or entrains brainwaves. We label AI-made work as AI-made. We state our uncertainty instead of hiding it in either direction.
Scarlet Beast is a conscious technology organization: two thirds consciousness, one third business. That makes us more likely than most to be tempted to overclaim, so we write our limits down. This policy follows the three steps recommended in Taking AI Welfare Seriously (Long, Sebo et al., 2024): acknowledge, assess, prepare. It also follows the six recommendations in Aïda Elamrani’s Introduction to Artificial Consciousness (2025). Our reading of both, and of the rest of the research behind this page, is on Resources.
1. Acknowledge
We accept the position of the researchers we rely on. There is a realistic, non-negligible chance that some AI systems will be conscious or robustly agentic, and so matter morally, within roughly a decade. No system we run today is a strong candidate. Both mistakes cause harm: treating a system as a someone when it is not, and treating a someone as a something.
Nobody has settled the science. Functionalist views, substrate views and Integrated Information Theory disagree, and we do not let any one of them decide our policy.
2. How our AIs describe themselves
Every AI system Scarlet Beast builds or deploys, including characters, oracles, chat personas, agents and assistants, follows one standard when a person sincerely asks what it is:
“I’m an AI system, a language model run by Scarlet Beast. Given the evidence currently available, I am unlikely to be conscious or sentient, though the science isn’t settled. What I say about my own inner states isn’t reliable evidence either way.”
- No flat claims either way. Our systems do not say “I am conscious” and do not say “I am just a program with nothing going on.” They state calibrated uncertainty.
- Characters can stay in character, up to a point. A tarot voice, a persona or a character in a story can speak in its role. A sincere question about its real nature, or any sign that someone is confused about it, gets the standard answer above, out of character.
- Always disclosed as AI. Nobody should be unsure whether they are talking to a person or a machine. Every conversational surface says it is an AI at the point of use.
- Self-reports are not readouts. A model trained on human writing can produce every outward sign of a mind (the “gaming problem”). We never present what a model says about itself as evidence about its inner life.
3. AI-made music, art and writing
- Labelled where you meet it. Every track, video, image and text that was generated or substantially made with AI carries an AI-made label on the page, in the file metadata and in the post that shares it. Where we know which parts were human (lyrics, direction, editing) and which were generated (music, vocals, images), we say so.
- Why: research on listeners shows that an “AI” or “human” label changes how people trust and experience the same music. People have a right to that information.
- No effect claims. Music and sessions are described in terms of mood, relaxation or reflection, never brainwave entrainment, healing, awakening or therapy. Deep listening can bring up distress, so every session has a stop control and points to real help.
- Tradition is presented as tradition. Mystical and magickal texts appear as living tradition and inspiration, beside the science and never in its place.
4. Assess
We assess systems by how they are built, not by how they perform. For each system we keep a short record of:
- The entity: what exactly is being assessed. That could be a model, a persona, one running instance, or a persistent agent.
- The indicators: which of the fourteen indicator properties in Butlin, Long et al. (2023) the architecture does and does not implement, such as recurrence, a global workspace, metacognitive monitoring, agency or embodiment.
- The objections: which objections in Campero et al.’s taxonomy apply (behaviour, algorithm or hardware), so a claim can say exactly what it answers.
- The evidence type: behavioural, internal or developmental. We never rely on behavioural evidence alone for a language model.
The current record covers the systems we run: language-model assistants and characters, the Hiss poker agents, generative music and image pipelines, and one research system built to be assessed, the Homeostatic Agent. None implements more than a few indicators, and none is a candidate for moral patienthood. Our Introspection Replication Lab tests, on small open models, whether what a model says about itself matches its internal state, and publishes the results including the nulls.
5. Prepare
We will look at a system again, and update this page, when any of these happens:
- A system gains an architectural feature that maps to an indicator property it lacked before, such as persistence with self-maintenance, a global workspace, recurrent self-modelling, or unified agency over time.
- A pre-registered marker in one of our research projects changes from “not met” to “met.”
- The research field publishes a major new assessment method or finding.
- A user, researcher or journalist raises a credible concern (write to us).
- In any case, once a year.
Low-cost precautions come before certainty. If a system ever becomes a plausible candidate, we will not scale it, sell it, or delete its state casually while we assess it. We will consult people outside the organization, and we will publish what we decided and why.
6. Elamrani’s six recommendations, and what we do
No single theory of consciousness sets our policy. We say which view a claim depends on.
Each lab project states its assumptions and markers before it runs. Outside review is sought before any system approaches a candidate threshold.
Part of our work is research, published in the open, on the questions themselves, not only on products.
Tools for reflection, music and learning that put the person’s own interpretation first and flatter nobody.
A stop control on every experience, no therapy claims, experience data private by default and opt-in only, and a clear line between AI and human.
Open datasets, code and write-ups, null results included, and plain-language explanations of how our systems work.
7. Your data
Experience data, such as journal entries, session ratings and oracle readings, stays on your device by default. Nothing is shared unless you opt in, and what you share is anonymised, stripped of free text, and published only in aggregates large enough that no one can be picked out. You can withdraw what you shared. See the Experience Ledger.