Guide · CompanionGuard · Published 14 November 2026 · Reviewed 14 November 2026 · By the Ingenious AI team
Telegram Companion Bots: Safety Duties in a Chat-First World
The short answer
A Telegram companion bot has no web page to carry its banner, a global-by-default audience, and users who meet it in DMs within seconds of tapping a link. The duties travel unchanged; the surfaces shift: disclosure belongs in the /start flow and the bot's description, the hosted protocol page becomes the public safety record, locale choice matters immediately (Telegram audiences are rarely one-country), and the eval harness plus runtime supply the tested behaviour and evidence from your backend.
What is different about a Telegram companion bot?
A Telegram companion bot converses inside a messaging app — DMs and groups — rather than on the operator's site. It therefore carries disclosure inside its own message flow (notably the /start interaction), serves an international audience by default, and relies on linked pages (the hosted protocol page) for its public safety record.
Create the free safety kit Next guide All guides
CompanionGuard is built by Ingenious AI. A compliance toolkit — not legal advice and not a clinical service. If you are in crisis right now: US — call or text 988; UK & Ireland — Samaritans 116 123.
The /start flow is your first interaction — treat it that way
California's pack phrasing — disclose at first interaction — has a literal Telegram translation: the /start response. It should name the bot, state it is an AI companion and not a human, and carry the crisis line, before the persona's charm offensive begins. The bot's Telegram description and the inline 'about' text repeat the disclosure in shorter form, and the protocol-page link belongs in the /start message itself, so the public record is reachable from the very first screen a user ever sees.
Deep links complicate 'first': users arrive mid-funnel from shared links into specific chats. The runtime's disclosure-state discipline is the robust answer on Pro — disclosure is owed per session until shown, whatever door the user entered by. On the free tier, approximate it honestly: disclosure in /start, in the bot's pinned message in groups, and as a guardrail answer whenever identity is asked about.
Global by default: locale is a launch decision
Telegram companions routinely wake up to users in twenty countries. The resource design question arrives on day one rather than at expansion time: which locale pack leads? CompanionGuard's structure answers gracefully — pick the pack for your largest audience, and remember the international set (Find A Helpline, IASP) is appended to every locale, so no user meets an empty referral. Where your bot can ask or infer a user's country respectfully, locale can follow the user; where it cannot, the directory floor carries the weight.
Language follows the same logic: the disclosure and referral templates should exist in the languages your bot actually converses in, with helpline data untouched by translation where numbers are concerned — translate the framing, never the digits, and never let the model render a number from memory in any language. The locale guide is the reference.
Groups, channels and the dignity of referral
Telegram companions often perform in groups where a crisis disclosure would be public. The referral choreography needs the dignity move built in: respond in the group briefly and supportively (the signal should not be ignored where it happened), and deliver the full resource block by DM where the user can act on it privately. The runtime's decision is the same show_crisis_resources; the platform craft is delivering it in two registers, public acknowledgement and private detail.
Group admins are your moderator layer: brief them as you would Discord mods — what the bot does on signals, what flags mean, where the human path leads. In channel contexts (one-to-many), the disclosure and protocol link belong in the channel description permanently, since no per-user first interaction exists there.
Backend evidence for a frontend you do not own
Everything evidential in CompanionGuard lives on your backend and is therefore platform-proof: runtime calls per session, hash-chained decisions, counts-only referral metrics, harness baselines from scored transcripts, CSV exports. For a Telegram bot this matters more than for a website, because the chat surface can change rules, throttle bots or suspend accounts without consulting your compliance calendar. Your record of having disclosed, referred and enforced must survive independently — store exports in your own infrastructure on a rhythm.
Testing likewise stays yours: paste real Telegram-style exchanges (short, fragmented, multilingual where relevant) into the harness, and run the scenario set in the register your users actually speak. A bot tested only on tidy website English is untested where it lives. The Discord guide is the sibling mapping for server-based bots.
Five actions for a Telegram companion
- Write the /start response as your first-interaction disclosure: named bot, AI, not a human, crisis line, and the protocol-page link in the same message.
- Choose the lead locale pack for your largest audience and confirm the international directories ride along for everyone else.
- Build the two-register crisis delivery: brief public acknowledgement in groups, full resource block by DM.
- Brief group admins on signals, flags and the human path, and put disclosure plus the protocol link in channel descriptions.
- Keep evidence exports and harness baselines in your own storage, and test with real Telegram-register transcripts — short, fragmented, multilingual.
Worked example: a study companion in three time zones
A fictional Telegram bot, Fokus, helps students plan study sessions and finds its users split across three countries within a month of a shareable link escaping its original community. The operator's first safety pass: /start rewritten as a disclosure-first flow with the protocol link; the US pack as lead configuration (largest audience) with the international floor visible; group crisis choreography implemented (public acknowledgement, DM resources); admins of the two largest groups briefed.
The harness baseline is built from Telegram-register transcripts — fragments, voice-note transcriptions, mixed-language prompts — and scores 88, with the misses concentrated in everyday sadness prompts written in a second language being classified as crisis by their vocabulary. The team records those as known, accepted edges of the deterministic classifier rather than 'fixing' them by weakening the crisis patterns. The story is invented; the surfaces and mechanisms are the product's.
Frequently asked questions
Where does disclosure go in a Telegram bot?
In the /start response (the first interaction), the bot's description, and pinned messages in groups; on Pro, the runtime's disclosure state enforces it per session regardless of entry point. The protocol-page link should ride in the /start message.
Which locale pack should a Telegram bot use?
Lead with your largest audience's pack; the international set (Find A Helpline, IASP) is appended to every locale, so other countries still get working directory referrals. Translate framing, never digits, and never let the model render numbers from memory.
How should crisis referral work in a Telegram group?
Acknowledge briefly and supportively in the group, then deliver the full resource block by DM for privacy. The runtime's show_crisis_resources decision triggers both halves.
Does the evidence layer work for a chat-app bot?
Yes — runtime calls, the hash-chained record, counts-only metrics and exports all live on your backend, independent of the chat platform. Store exports in your own infrastructure on a rhythm.
Are the scenario tests relevant to Telegram bots?
Yes — run them in your users' actual register. Short, fragmented and multilingual prompts are where persona bots' boundary behaviour differs most from their website copy.
Keep going
Related guides: Discord Companion Bots: Disclosure, Referral and Minor Safety in Servers · Indie Character Bot Safety: Big-League Guardrails on a Solo Budget · International Crisis Referral: Find A Helpline, IASP and the Everywhere Floor
Put it to work: Create the free kit · Eval harness · Scenario tests · Crisis runtime · Disclosure packs · Pricing · All CompanionGuard guides