AI-Generated Content Disclosure
Last updated: 2026-05-21 Version: 2.0
Swaps uses artificial-intelligence (AI) systems in several places. This disclosure explains where AI is used, what its limitations are, and how to interpret AI-generated output. It complements the Terms of Service §12.
1. Where AI is used in the Service
| Surface | AI use | What it does |
|---|---|---|
| AI Concierge (in-product chat) | Large-language-model dialogue | Answers user questions about crypto, providers, the Service. Can suggest pairs, providers, and steps. |
| Help Center search & ranking | Embedding-based retrieval | Returns help articles most relevant to your query |
| Search Address Report — narrative | LLM summarisation of risk signals | Summarises numeric risk signals into prose. The signals themselves are not AI-generated. |
| Support triage (internal) | LLM-assisted ticket classification | Routes tickets to the right runbook. Humans handle replies. |
agent.swaps.app /llm/quote, /llm/routing-explainer | LLM responses over deterministic quote/routing data | Wraps deterministic outputs in natural language for agent consumers |
@agent.swaps/mcp-server resources | LLM-friendly resource framing | Same as above, packaged for MCP-compliant agents |
| ChatGPT mini app responses | LLM (host: OpenAI) + Service tools | Tool outputs are deterministic; surrounding language is generated by the host |
| Swaps Search via ChatGPT App / MCP connector | LLM dialogue (host: ChatGPT or another MCP client) + deterministic Swaps Search tool outputs | V1 is read-only: address checks, transaction explanations, fund tracing, taxonomy/corridors, route quotes. No checkout, no trade execution, no provider session inside the host. Authenticated workspace/report/credits flows return auth_required (V1 has no OAuth). Tool responses may include request IDs for support and debugging; public, user-facing responses minimize internal telemetry except where needed for support. |
| Swaps Search in the Swaps iOS app | Same intelligence and tooling as the ChatGPT App / MCP path, exposed natively in iOS | Read-only checks, explanations, tracing, taxonomy, quote retrieval. External checkout (if any) opens an external link to a Swaps-owned domain — no in-app execution of crypto purchase or transfer. |
2. Models we use
We use third-party large-language models from Anthropic and OpenAI as of this version. We may add or change models without notice. Where the model is hosted by the provider (e.g., Anthropic Claude via API), prompts and responses transit to the provider; provider DPAs and SCCs apply (see the Privacy Policy §6).
3. Known limitations
AI responses can be:
- Inaccurate — including statements that look confident but are wrong.
- Out of date — model training data has a cut-off; new provider integrations, new regulations, new payment-method launches may not be reflected.
- Hallucinated — invented facts, made-up provider names, fabricated quotes.
- Inconsistent — same question can produce different answers on different runs.
- Bias-affected — models reflect biases in their training data.
We tune prompts, ground responses in deterministic Service data where possible (our internal architecture — no silent fallbacks), and add disclaimers in the UI. None of these tunings makes AI output reliable enough to act on without verification.
4. How to interpret AI output
- Treat as a starting point, not a conclusion. AI answers tell you where to look; they do not tell you what is true.
- Verify before acting. If AI quotes a fee, check the fee in the provider's own interface. If AI cites a regulation, check the regulation itself.
- Do not rely on AI for legal, tax, financial, or medical advice. AI cannot give regulated advice. Neither can Swaps. Consult a qualified professional.
- Numeric facts come from deterministic Service data, not AI. When you see a number on Swaps (a quote, a fee, a country code), that number is from the relevant Provider API or our own pricing tables, not from an LLM.
5. Search Address Report — narrative is AI, signals are not
In the Search Address Report:
- The risk signals (e.g., "exposure to mixer X", "interaction with sanctioned cluster Y", "first-seen date Z") come from on-chain analytics and third-party feeds. They are deterministic.
- The narrative (the prose paragraphs that explain the signals) is generated by an LLM that summarises the signals into readable text. The narrative may simplify, omit nuance, or occasionally misphrase a signal.
Where the narrative and the signals appear to disagree, the signals are authoritative. Report misalignments to legal@swaps.app for correction.
6. agent.swaps.app and MCP responses
When you use agent.swaps.app /llm/* endpoints or the @agent.swaps/mcp-server package, the underlying quote and routing data is deterministic — same input produces the same answer (within the 30-second quote refresh window). The natural-language wrapping is LLM-generated and inherits the limitations in Section 3.
If you build an automated agent or downstream service on top of Swaps' AI endpoints, do not chain AI output into automated transaction execution without human review. Hallucinated quotes can produce real losses if executed by a downstream agent.
7. Your data and AI
When you interact with the AI Concierge:
- Your prompt is sent to the model provider (Anthropic or OpenAI) for generation. The provider does not use your prompts to train models (per their enterprise terms).
- We retain the prompt and response for up to 30 days for product improvement and abuse detection, then aggregate or anonymise.
- We do not send your wallet address, full email, or Account password to model providers as part of system context. Prompts you yourself type may contain PII; we recommend not pasting wallet addresses or seed phrases into the chat.
8. Right to human review
Where automated processing has a significant effect on you (such as Account suspension for suspected abuse), we always conduct human review before any final decision (see the Privacy Policy §14). You may request human review of any AI-driven outcome by emailing legal@swaps.app.
9. Reporting AI errors
If you observe an AI response that is wrong, misleading, or harmful, please report it to legal@swaps.app with as much detail as possible (prompt, response, why it's wrong). We use these reports to improve the system.
10. Untrusted third-party data
Many AI responses surface or summarise data drawn from outside Swaps: public blockchain transactions, on-chain metadata, block explorers, third-party risk feeds, provider-published rate tables and country coverage, and content authored by users or third parties. Treat all such source data as untrusted by default.
- Blockchain data is public, append-only, and can be deliberately constructed to mislead — for example, addresses can be tagged to imitate legitimate services, or transaction memos can be authored by adversaries.
- Provider data is reflective of the Provider's own systems and may be incorrect, stale, or incomplete.
- Third-party risk feeds can contain false positives (tags applied to clean addresses) or false negatives (no tag on a known-bad address).
Where an AI response on Swaps presents data from any of these sources, the AI is acting as a summariser, not a verifier. Verify any consequential conclusion in a primary source — a block explorer for on-chain facts, the Provider's own interface for rates and KYC outcomes, the regulator's own register for regulatory status — before acting on it.
Contact: legal@swaps.app