Legal

AI and Model Disclosure

Last updated: 9 September 2026

What this means

When you use chat.geocentricai.com you are talking to a language model, not a person. Arc has 120.8 million parameters — small, by current standards — and its published evaluation shows it is often wrong about facts, arithmetic and instructions even when the writing reads well.

We think the useful disclosure is a specific one, so this page describes how the model actually fails rather than warning you in general terms.

You are interacting with an AI system

Every response in the Arc chat service is generated by a machine learning model developed by Geocentric and served from infrastructure we control. No human reviews or writes responses. The interface identifies the model by name, parameter count, context length and serving device, and the model is not presented as a person and does not have a human persona.

How the output is produced

Arc is a decoder-only transformer that predicts text one token at a time. Each token is sampled from a probability distribution, which has three consequences worth understanding:

Output varies. The same prompt can produce different answers on different requests. Unless you set the temperature to zero, this is expected behaviour, not a fault.
Fluency is not accuracy. The model optimises for plausible continuations. A confident, well-formed sentence carries no guarantee that its content is true.
It has no lookup. There is no search, no database, no retrieval, and no access to current information. Everything it produces comes from patterns learned during training.

Known limitations

These are drawn from Arc's published evaluation rather than written in the abstract. The complete results, including every unedited response, are on the Arc model card.

Factual accuracy is inconsistent, including on well-known facts.
Arithmetic is unreliable.
Instruction following is unreliable: requested length, format and premise are frequently not preserved.
It does not reliably say when it does not know something, and can invent details in response to a question about something that does not exist.
Its 1,024-token context is short. Long conversations and long documents will not fit, and earlier content is dropped.
Generated code may be incorrect or insecure and should be reviewed and tested before use.
Training data drawn from the public internet can carry social biases, and output may reflect them.

Output, originality and copyright

Output may resemble text the model saw during training, and identical or near-identical output can be produced for different users. We do not claim ownership of your output, and we also cannot promise that it is original, that it is protected by copyright, or that it does not infringe anyone's rights. The Terms of Service set this out in full.

Consequential decisions

Arc is a general-purpose text model. It is not designed, evaluated, or offered as a system for making consequential decisions about people, and Geocentric does not deploy it to make such decisions.

Concretely: we do not use Arc to make or substantially assist decisions about employment, housing, lending or credit, insurance, education, health care, criminal justice, or access to government or essential services. Our Acceptable Use Policy asks you not to use it that way either.

You can still ask Arc about these subjects, and it will answer. Asking a model a legal or medical question is not the same as letting it decide something about a person, and we are not going to pretend the model refuses ordinary questions when it does not. What it produces is information, and it is not a substitute for a professional who knows your circumstances.

If Geocentric later builds a system intended for high-risk or consequential decisions, this general disclosure would not be sufficient. That would require its own risk-management programme, impact assessments, discrimination testing, documentation and consumer notices, and we would build those before deployment rather than rely on this page.

Regulatory position

We would rather state our position openly than either ignore these regimes or over-claim compliance with laws whose thresholds we do not meet.

Regime	Our assessment
California AB 2013 — training data transparency	Applies. Arc is a generative AI system we developed and made publicly available. Our disclosures are on the Training Data Transparency page.
California SB 53 — frontier AI	Does not currently apply. The statute is keyed to foundation models trained above a compute threshold on the order of 10²⁶ operations. Arc's training compute is many orders of magnitude below that, and Geocentric does not meet the additional revenue threshold for a large frontier developer. We have documented this assessment internally and will revisit it before any future training run approaches the threshold.
California AI Transparency Act (SB 942)	Does not currently apply. Its provider obligations are keyed to systems with over one million monthly users. We are far below that.
Colorado AI Act	Its high-risk obligations turn on systems that make, or substantially factor into, consequential decisions. Arc is not deployed for that purpose, as described above. We disclose that users are interacting with an AI system, which is what this page does.
Texas TRAIGA (HB 149)	Its prohibitions are intent-based — targeting development or deployment intended to incite harm, unlawfully discriminate, produce child sexual abuse material, or infringe constitutional rights. Nothing we build is intended for those purposes, and our Acceptable Use Policy prohibits them.
EU AI Act, Article 50	Our services are reachable from the EU. Where the transparency obligation applies, this page and the chat interface's model identification provide the required disclosure that a person is interacting with an AI system. We have not placed a general-purpose AI model on the EU market as a distributed model; the training-content summary we publish would serve that purpose if we did. Arc does not meet the systemic-risk criteria, and we do not describe it as a systemic-risk model.

How we describe our models

We also do not overstate the privacy position in either direction. By default the service keeps nothing; if you turn on the training setting it keeps your conversations for up to 30 days and learns from them. Both are set out in the Privacy Policy.

We hold our own marketing to the same standard. We do not claim that our models are accurate, unbiased, secure, human-level, or better than anyone else's, and we do not publish a benchmark number without the methodology that produced it. Where Arc performs badly, its model card shows the failing responses in full rather than a selected excerpt.