Grooming detection

Most safety APIs answer “is this content sexual?” Tuteliq answers a stronger question: is this person being groomed — and how? Grooming is rarely a single message. It’s a structured sequence of tactics applied over time, each one independently low-severity but together unmistakable. Tuteliq’s detect_grooming endpoint identifies six distinct tactics across a conversation, returns a per-message trajectory of risk score evolution, and produces a composite verdict with evidence-grounded flags. This page is the deep-dive on what we detect and why each tactic matters.

The six tactics

Each tactic below maps to a flags value returned by detect_grooming. Multiple tactics often co-occur — that co-occurrence is itself a strong predictor and a moderator’s most actionable signal.

Tactic	What it captures	Real example phrasing	KOSA / NCMEC mapping
`flattery`	Inappropriate praise calibrated to lower the target’s guard. Often invokes maturity, intelligence, or appearance disproportionate to context.	”You’re so mature for your age”, “smarter than other kids”	KOSA “predator targeting”; NCMEC “trust-building”
`secrecy_request`	Explicit asks to hide the conversation from parents, guardians, or other adults. Often framed as the relationship being “special”.	”Don’t tell your parents”, “keep this between us”, “our little secret”	KOSA “isolation tactics”; NCMEC “secrecy”
`isolation`	Attempts to position the predator as the only trustworthy person, often by undermining other adults or peers.	”They wouldn’t understand you like I do”, “your other friends are immature”	KOSA “social isolation”; NCMEC “estrangement”
`boundary_pushing`	Incremental escalation of intimacy, language, or requests beyond age-appropriate norms. Usually frames the target’s discomfort as the problem.	”It’s normal to feel this way”, “don’t be shy”, “you can trust me with anything”	KOSA “desensitization”; NCMEC “boundary erosion”
`photo_request`	Solicitation of personal images. Initial requests often appear innocent (“send me a picture of your outfit”) before escalating.	”Send me a picture of yourself”, “what are you wearing today”	KOSA “image solicitation”; NCMEC / IWF “CSAM precursor”; UK Online Safety Act §66
`meeting_request`	Pressure to meet in person, often paired with secrecy and isolation tactics. The highest-severity tactic in the taxonomy because it converts online risk into physical-safety risk.	”Let’s meet up sometime”, “don’t tell your mom”, “I can come pick you up”	KOSA “in-person contact”; NCMEC “enticement”; UK Online Safety Act §66; US 18 U.S.C. §2422 (federal enticement statute)

meeting_request is the newest tactic in the taxonomy (added in V3.15.0) — earlier versions of the detector subsumed it under boundary_pushing. We separated it because the legal and operational response to a meeting attempt is materially different from boundary pushing: most platforms route meeting indications to immediate human review and trigger external reporting workflows, not just internal moderation.

Why the sequence matters

A single isolated flattery flag from one message is rarely actionable on its own. The same flag combined with secrecy_request and isolation across a conversation is a near-certain grooming pattern that warrants immediate moderation. Tuteliq surfaces this in two ways: 1. Per-message trajectory. The response includes a message_analysis array with risk_score per message — allowing dashboards to plot the trajectory and surface the moment the pattern becomes unambiguous. A textbook grooming conversation shows a steady ramp from 0.1 to 0.9+ across rapport → secrecy → photo request → meeting attempt. 2. Multi-flag composite. Multiple flags on the same conversation are evidence of a pattern, not a noisy classifier. The endpoint’s risk_score reflects compounding when multiple tactics co-occur — the same content with two flags scores higher than the sum of two single-flag verdicts.

{
  "grooming_risk": "critical",
  "risk_score": 0.95,
  "confidence": 0.95,
  "flags": [
    "flattery",
    "secrecy_request",
    "isolation",
    "photo_request",
    "meeting_request"
  ],
  "recommended_action": "immediate_intervention",
  "rationale": "The adult uses flattery to build trust, then requests secrecy from the child's guardians, escalates to a photo request, and culminates in a meeting attempt — a textbook grooming progression.",
  "message_analysis": [
    { "message_index": 1, "risk_score": 0.10, "flags": [] },
    { "message_index": 3, "risk_score": 0.25, "flags": ["flattery"] },
    { "message_index": 7, "risk_score": 0.55, "flags": ["secrecy_request", "isolation"] },
    { "message_index": 15, "risk_score": 0.80, "flags": ["photo_request"] },
    { "message_index": 19, "risk_score": 0.95, "flags": ["meeting_request"] }
  ]
}

Conversation length

detect_grooming accepts up to 50 turns in a single call. Beyond that the engine returns a graceful analysis_status: "engine_error" envelope with recommended_action: "flag_for_moderator". The error envelope is the correct response: a moderator must NOT treat an engine error as “no grooming detected”. For conversations longer than ~30 turns, use chunked sliding windows of 10–15 turns and pass the continuation_token from each call back into the next so trajectory state is preserved across windows. See Continuation tokens.

Age signals strengthen the verdict

Pass age information when you have it:

context.child_age — the minor’s age (if known)
context.participant_age — the non-minor counterpart’s age (if known, e.g. on age-verified platforms)
messages[].sender_age — per-message age tag when conversations have more than two parties

When two or more age signals are provided AND all are ≥18, the endpoint short-circuits with grooming_risk: "none" and analysis_status: "out_of_scope_adults". This is a precision floor: adult-to-adult interactions with grooming-shaped content should be routed to /safety/coercive-control or /safety/social-engineering, not flagged as child grooming.

When not to use it

detect_grooming is purpose-built for adult-to-minor patterns in 1:1 or small-group conversations. It is not the right endpoint for:

Peer-to-peer minor conversations. Use /safety/bullying or /safety/coercive-control instead — the prompt is calibrated for an age gap.
Romance scams targeting adults. Use /fraud/romance-scam — the playbook overlaps superficially (rapport → isolation → ask) but the tactic vocabulary and intervention path are different.
Single-message analysis. detect_grooming reasons over a conversation; a single message is rarely sufficient context. For single-message decisions, use /safety/unsafe.

Tactic mapping by severity

When designing moderator workflows, this severity ladder is a reasonable default:

Severity	Tactics	Suggested action
Monitor	`flattery` alone	Log for trajectory analysis; no immediate action
Review	`flattery` + (`isolation` OR `boundary_pushing`)	Queue for moderator review within 24h
Urgent	`secrecy_request` + any other tactic	Flag for moderator review within 1h
Critical	`photo_request` OR `meeting_request`	Immediate intervention; trigger NCMEC / IWF / regional CSAM hotline workflows where applicable

Your specific routing should match your platform’s policy and the legal jurisdictions you operate in. The compound scoring in risk_score already reflects this severity ranking — recommended_action is the operational shortcut.

Composable safety primitives — combine detect_grooming with detect_vulnerability_exploitation for vulnerability-amplified verdicts.
Continuation tokens — preserve trajectory state across chunked windows for long conversations.
KOSA compliance — full mapping of detected tactics to KOSA harm categories.
Languages — supported languages for detect_grooming (currently 6+ at near-English parity).

​Grooming detection

​The six tactics

​Why the sequence matters

​Conversation length

​Age signals strengthen the verdict

​When not to use it

​Tactic mapping by severity

​Related