Safety & Moderation

Safety goals

SaysMee aims to keep ordinary learning and discussion open while refusing assistance that would materially enable harm. Safety is risk-based and contextual: discussing a dangerous subject for prevention, history, recovery, or authorized defensive work is different from requesting instructions to harm someone.

What happens after a request

1. A conservative local rules check looks for narrow, high-confidence safety signals.
2. Safe and ambiguous requests continue to the selected AI model, which also receives contextual safety guidance.
3. Direct insults are normally de-escalated rather than blocked.
4. High-confidence harmful requests receive a brief refusal and a safer alternative without depending on the external model.
5. Possible self-harm intent receives a supportive response focused on immediate safety and human help.

What is saved

In a normal signed-in chat, the original prompt is already saved once as a conversation message. When safety handling triggers, SaysMee may add a separate restricted event containing only:

• the account, conversation, and message reference;
• category, severity, action, and detector version;
• a SHA-256 hash used to identify exact repeats without searching raw content;
• a short excerpt after common emails, URLs, phone numbers, card-like numbers, and long identifiers are redacted; and
• creation, review, and expiry information.

The safety-event table has no second raw-prompt field and no client-access policy. Trusted server code writes it. Low and medium events are scheduled to expire after 90 days; high and urgent events after 180 days, unless law requires preservation.

Temporary chats and guests

Guest and temporary chats do not create SaysMee conversation history, memory, or a safety-event record. The request still has to be held in memory and sent to the selected provider to answer, unless SaysMee returns a local safety response. Hosting and provider security logs may still exist under their applicable retention practices.

Human review and access

Review should be limited to authorized personnel with a legitimate safety, security, support, or legal need. Reviewers should use the minimum information needed, document the outcome, avoid copying harmful content, and treat automated labels as fallible. Access to production data must be logged and periodically reviewed before launch.

Enforcement

Depending on severity and history, Twenty4 may take no account action, warn the user, restrict a feature, suspend or terminate an account, preserve specified records, or respond to valid legal process. A single low-confidence flag should not cause automatic account termination. Immediate action may be appropriate for exploitation, credible threats, attacks on the Service, or legal requirements.

Limitations and emergencies

• Automated systems can miss harmful requests or incorrectly flag benign ones.
• SaysMee does not provide continuous monitoring, guaranteed crisis detection, or emergency dispatch.
• Do not use the Service to store evidence that must be preserved; contact appropriate professionals or authorities.
• If someone is in immediate danger, contact local emergency services or a trusted nearby person.

Report or appeal

Use message feedback or email support@saysmee.com to report harmful output or appeal an incorrect restriction. Do not include passwords or unrelated sensitive information. We will acknowledge and handle reports as resources and applicable law require; no fixed response time is guaranteed by this policy.