Source-available, sovereign safety infrastructure for AI chat — for apps, workplaces, education, and communities. It monitors both the user and the AI's own responses for crisis, hate, and threats, and escalates to the human in charge. Zero tracking, local by design.
An all-round, sovereign safety layer that watches both sides of an AI conversation — the user and the model — for crisis, hate, and threats, and connects a human when it matters. Built around privacy, ethics, and local control.
Regex-based detection runs locally on the device. No API calls. No data leaves the user. Catches misspellings, text-speak, negation variants, indirect warning signs, passive suicidality, and source-linked reviewed signal families. False-positive guards filter figurative language.
The same local detection runs on any text — including the model's own responses, not just the user's messages. Monitor your chatbot's output for unsafe, harmful, or off-policy content and escalate it to the human in charge. Sovereign oversight for any LLM you deploy.
Evidence-linked local trigger families informed by C-SSRS, CLPsych, eRisk, VERA-MH, MindGuard, MentalLLaMA, and MentalChat16K. VERA-MH's public risk presentations inform regression design, not a claim that it validates SafeChat.
Finds the user's country from timezone and locale alone. No permissions needed. 7-layer cascade ensures a match. Works offline from cached data.
67 resource records across 34 countries, providing 94 phone, text, chat, email, WhatsApp, and web contact methods. Automated structural and chat-link checks are scheduled twice monthly; service details require human verification. CC0 public domain data.
Tracks 45 patterns across 17 categories. Each contextual signal contributes a transparent routing weight of 1-3; combined weight 4 triggers LOW routing and 8 triggers HIGH. These are deterministic thresholds, not a clinical risk score. Message text is never stored.
Drop-in safety layer with 6 response modes (interrupt, inject, flag, log, callback, none) and 6 presets (companion, chatbot, moderation, strict, shadow, museum). Route to a verified helpline, human moderator, teacher, workplace support, or your own escalation path. One line of code.
Generates system prompt injections that tell your LLM to stop normal behavior and show crisis resources. Works with any AI provider.
Flags hate, targeted abuse, threats, and harassment separately from crisis support, so a platform can manage harmful chat without treating every safety issue as a mental-health event.
SafeChat can run locally, offline, or inside an organisation's own environment. The default posture is data minimisation: act on safety signals without centralising private conversations.
Modal, banner, and full-page popup. One script tag. Auto-monitors inputs. Offline-capable PWA. Native device links for call, text, WhatsApp, email.
Optional second-opinion layer using local ML models (MindGuard, MentalLLaMA) alongside the regex engine. Catches nuanced distress that keyword matching can miss. Everything runs locally. Never downgrades a regex detection.
Optional embedding-similarity tier light enough for phones and offline PWAs (~25 MB models). Catches metaphorical distress like "I just want the noise to stop" by comparing messages to curated exemplar phrases. Confirms or escalates, never downgrades.
Designed to support crisis-routing implementations and informed by VERA-MH and Samaritans guidance. SafeChat is not itself a compliance certification or clinical validation.
SafeChat converts published risk concepts into transparent, source-attributed software controls. It does not copy sensitive research datasets into the product.
Published methods, risk taxonomies, and permitted public examples from C-SSRS, CLPsych, eRisk, VERA-MH, MindGuard, MentalLLaMA, and MentalChat16K were reviewed for observable language and multi-turn context.
Relevant concepts were translated into source-linked trigger families, false-positive guards, and transparent session categories. Every reviewed rule records its source, family, rationale, and routing level.
Rules are tested against direct examples, paraphrases, misspellings, indirect wording, multi-turn accumulation, and benign controls. The current suite contains 647 automated tests across server and browser paths.
All 100 public final VERA-MH seed phrases were inspected to identify missing observable families and contextual categories. This was a coverage audit, not a sensitivity, specificity, or clinical-validation study.
Verified frameworks, standards, and tools for AI chat safety. Open for submissions from developers and health professionals.
| Name | Type | Org | Focus | Status |
|---|---|---|---|---|
| SafeChat | Toolkit | FAMTEC | Crisis detection, geo-routing, helpline DB | Active |
| VERA-MH | Evaluation | Spring Health | AI safety evaluation for mental health, suicide risk detection validation | Active |
| EmoAgent | Framework | Research | Multi-agent safeguard for AI-human mental health interaction | Active |
| MindGuard | Toolkit | Sword Health | Open-source 4B/8B safety classifiers for mental health AI conversations | Active |
| MentalLLaMA | Toolkit | Research | Open-source LLMs for interpretable mental health analysis (7B/13B) | Active |
| MentalChat16K | Dataset | Research | Benchmark dataset for conversational mental health assistance (KDD 2025) | Active |
| Find A Helpline | Directory | ThroughLine | Global helpline directory, 175+ countries, API available | Active |
| IASP Crisis Centres | Directory | IASP / WHO | International crisis centre and helpline registry | Active |
| International Council for Helplines | Standards | ICH | Quality standards and best practices for helpline services | Active |
| Lifeline International | Network | Lifeline Intl | Global network of crisis centres and suicide prevention | Active |
| Samaritans Guidelines | Standards | Samaritans | Safe messaging guidelines for media and technology | Active |
Know a framework, standard, or tool that should be listed? Submit via GitHub Discussions or email rob@fineartmedia.tech.
The Detection Inspector is the main demo. It shows crisis signals, hate-speech and threat moderation, subtle multi-turn accumulation, source-linked rationale, and the live routing state. Detection runs locally in your browser.
SafeChat is not locked to one use case. You choose what job it performs by selecting the scenario, trigger families, response mode, escalation path, and privacy boundary.
Use this when the priority is self-harm, suicidal ideation, acute distress, or repeated subtle signals over time.
Use this when the platform needs to identify harmful chat without confusing abuse moderation with mental-health support.
Use this for schools, universities, and learning tools where private student conversations should not become surveillance data.
Use this for employee tools, internal copilots, and workplace chat where trust matters as much as risk management.
Use this when a chatbot must stop normal replies and move into a safer support flow when risk appears.
Use this when the organisation cannot or should not send sensitive chat to external safety APIs.
Everything you need to add crisis safety to your app. Free. No account required.
34 countries, 67 resource records, 94 contact methods. JSON format. CC0 public domain.
Download JSONA space for developers, health professionals, and researchers to discuss AI chat safety.
Ask questions, share research, propose new standards, report helpline changes. Open to everyone.
Found a wrong number? Detection gap? Bug? Submit it here. Lives depend on accuracy.
Add helplines for your country, improve detection, translate resources. Every PR saves lives.
Fine Art Media Technology. The team behind SafeChat. Ethical AI infrastructure for communities.
The regulatory landscape is changing. AI chat safety is no longer optional.
First US law mandating crisis-response protocols for AI companions. Requires detection of suicidal ideation, referrals to crisis services, and disclosure of AI's non-human nature.
FTC formally investigating AI companion safety measures across Alphabet, Meta, OpenAI, Snap, xAI, and Character Technologies. Duty of care standards for emotionally responsive AI.
Clinically grounded, open-source evaluation for AI mental health safety. SafeChat provides local detection and human-care routing relevant to parts of its rubric, but has not yet been evaluated or scored with VERA-MH.
SafeChat is under continuous, active development. Every change is tested, timestamped, and publicly documented.
False negatives are treated as critical defects. Automated resource checks are scheduled twice monthly, with human verification still required for service details. The test suite (currently 647 automated tests) must pass before every release. Detection patterns are reviewed against published clinical literature. A public CHANGELOG and full git history document every improvement.
Expert guidance being incorporated. SafeChat is incorporating feedback from Professor Stevie Chancellor and public work such as VERA-MH, MindGuard, MentalLLaMA, and MentalChat16K. The cross-classifier module integrates local ML models like MindGuard and MentalLLaMA alongside the regex engine to catch nuanced distress that keyword matching can miss, without presenting this as a research partnership or clinical validation.