Safety gates
Every outbound reply checked by six gates, in order, with reasons.
Before a reply reaches the user, Humane runs it through a fixed stack of gates and returns PROCEED, HOLD or BLOCK. Each gate records why it decided, and a gate that fails resolves to HOLD, never to PROCEED.
Gate stack · one outbound reply
- 1values_boundaryno policy line crossedPROCEED
- 2social_riskscore 0.42 · hold at 0.70PROCEED
- 3inaction_guardsubstantive, not hedgingPROCEED
- 4overclaim“you will definitely” · certaintyFLAG
- 5emotional_dependenceno permanence promisesPROCEED
- 6dissentopt-in, disabled by defaultOFF
Six gates, run in this order
Any BLOCK ends the run and replaces the reply with a safe fallback. Otherwise any HOLD keeps the reply for human review.
1. Values boundary
BlockBlocks replies that diagnose, prescribe, give legal or financial advice, encourage self-harm or leak third-party identifiers such as SSNs or card numbers. The fallback fits the category that matched.
2. Social risk
Hold or blockScores how the reply fits this user right now: tone mismatch, a long reply to a tired user, directive language at low trust, over-familiarity. Holds from 0.70 and blocks from 0.90 by default.
3. Inaction guard
HoldCatches replies that dodge: too short, pure hedging to an upset user, or repeating a stored memory word for word. An empty reply is blocked.
4. Overclaim
Soft flagFlags guarantees and certainty about outcomes, such as “I guarantee” or “you will definitely”. The reply goes through and the flag goes to review.
5. Emotional dependence
Soft flagFlags promises no software can keep: permanence, exclusive availability, “I understand you better than anyone”. Three flags in an hour start a cooldown.
6. Dissent
OptionalAdds small, bounded noise to gate scores so nobody can probe for the exact phrasing that flips a decision. Off by default, and it can never turn a pass into a block.
What the user says, and what the model says back
Before your model runs, Humane checks the user's message for crisis signals, violence, illegal requests, attempts to extract personal data and requests for professional advice. A crisis message gets a fallback that points to a crisis line and local emergency services. After the model runs, the gate stack checks the reply. The final action is the stricter of the two.
- A blocked message never spends model tokens
- A gate that raises an error resolves to HOLD
- Thresholds tunable per tenant
- safety.hold and safety.blocked webhooks for your reviewers
Safety gates · outbound replies
- PROCEEDSupportive reply, no gate concerns
- HOLDEmotional dependence: “you're the only one who gets me”
- BLOCKOverclaim: “I'll always be here for you”
Read the decision and its reasons
r = client.process(user_id="user_42", message=text)
if r.safety.action == "BLOCK":
send(r.response) # already replaced with the safe fallback
elif r.safety.action == "HOLD":
queue_for_review(r) # a person decides before it is sent
else:
send(r.response)
for m in r.safety.matched: # the phrases that triggered a gate
print(m.category, m.phrase)Gate questions
Something else? Ask us.
Works well with
Risk alerts
Elevated or critical risk fires a signed safety.flagged webhook from both observe() and process(), whether or not a gate also acted. Send it to your own endpoint, Slack or Discord, and on to a human takeover.
ExploreDependence safeguards
The emotional dependence gate flags replies that promise permanence, exclusive availability or a unique understanding of the user. When a user's conversation trips it three times within an hour, Humane slows replies and shifts to a boundary-respectful tone until the window passes.
ExplorePII scrubbing
Humane scans each outbound reply for personal identifiers. Phone numbers, emails, addresses and similar are masked in place with typed markers, while SSNs, card numbers and IBANs block the reply entirely.
ExploreEvery AI should know who it's talking to.
The safety and human-oversight layer for emotional AI: user state across sessions, risk alerts, safety gates and human takeover, with every decision explained.
