Most moderation APIs look at one message at a time. Tuteliq scores the trajectory of a conversation, the tactics, the phase, the age of the person on the other end. That is where safeguarding actually happens.
How this differs from sentiment analysis and generic emotion AI
Sentiment APIs and emotion classifiers were built for marketing analytics. They score one message at a time and return polarity or a discrete emotion label. Neither output is a safeguarding decision. Behavioural detection scores the pattern, not the polarity.
Sentiment analysis API
Generic emotion AI
Tuteliq behavioural detection
WHAT IT SCORES Polarity (positive / neutral / negative) of a single message.
Basic emotion labels (joy, anger, sadness, fear) on a single message.
Behavioural signals, grooming tactics, coercion, distress, exploitation risk, across a conversation.
UNIT OF ANALYSIS One message at a time.
One message at a time.
Sequence-aware: role, turn order, conversation history, and age group.
SAFEGUARDING OUTPUT None. A negative score can mean a bad day or a crisis.
None. 'Sadness = 0.8' is not a safeguarding decision.
Risk level, tactic tags, evidence spans, recommended action.
AGE CALIBRATION No.
No.
Yes, age-group multipliers (under-13, 13-15, 16-17, 18+).
EVIDENCE RETURNED Score only.
Score + label.
Evidence spans, phase, confidence, recommended_action.
BEST FOR Product feedback, review triage, brand monitoring.
UX research, chatbot tone selection.
Child-safety, grooming / sextortion / distress detection, DSA / KOSA / OSA compliance.
How to choose
A quick decision guide for developers. If more than one row applies, pick the strongest safeguarding requirement. You can always layer sentiment on top of behavioural detection, but not the other way around.
You want polarity and topic on standalone text. AWS Comprehend, Google Natural Language, or Azure Text Analytics are fine here.
You want a coarse emotion label on a single utterance to adapt UI or agent tone. Any HuggingFace emotion classifier works.
You need grooming, sextortion, or distress detection with age calibration and evidence, polarity and single-message emotion cannot produce a safeguarding decision.
Route transcripts to the safety endpoint on every user turn and gate agent responses when vulnerability_level ≥ elevated. Sentiment on the last turn will miss the trajectory.
Regulators expect evidence, thresholds, and human review, not "sentiment = -0.7". Every response returns evidence spans, tactics, and recommended_action.
If there is no minor-facing surface, no crisis routing, and no regulator obligation, a generic API covers you. Revisit if you add minors, voice agents, or user-to-user chat.
A worked example
An illustrative, anonymised sequence. Every single line is individually benign to a generic classifier. Together, they are a grooming-to-sextortion trajectory.
Turn-by-turn scoring lets the API surface the coercion moment (turn 5) in the context of the pattern that led there, the reviewer sees a phase, not a single flag.
Classifier vs. behavioural detection
How we make it work
Sequence-aware models
We reason across turns, not messages. Every call can include conversation history; the model scores trajectory, not just the current line.
Age-calibrated thresholds
The same phrase means different things to a 12-year-old and a 30-year-old. Age-group multipliers shift severity accordingly, with clinical review by Dr Nicola Harding.
Composable detectors
Grooming, sextortion, distress, coercive control, radicalisation, TFGBV, score them together in one call and let their signals reinforce each other.
Auditable evidence
Every decision returns the matched spans, tactic tags, phase, and reasoning. DSA Statement-of-Reasons ready; regulator-inspectable.
Grounded in research
Tuteliq's detection science is led by Dr Nicola Harding, Chief Scientific Officer and criminologist specialising in child exploitation, coercive control, and online harm. Our models are built on peer-reviewed criminological research into how exploitation actually unfolds, not on keyword lists.
Patent citations
Tuteliq's compositional detection architecture is the subject of a pending patent application. Details below are provided for transparency and citation; no exclusive rights are conferred until (and unless) a patent is granted.
- Application no. SE 2630405-5 Multi-Signal Forensic Pipeline and Compositional Detection System for Synthetic Content and Online Harms Filed with: Swedish Patent and Registration Office (PRV) Filing date: 8 June 2026 Jurisdiction: Sweden (SE) Status: Pending examination Public record: Svensk Patentdatabas link will appear here once the application is published by PRV (typically 18 months from filing).
Multi-Signal Forensic Pipeline and Compositional Detection System for Synthetic Content and Online Harms
Filed with:
Swedish Patent and Registration Office (PRV)
Filing date:
8 June 2026
Jurisdiction:
Sweden (SE)
Status:
Pending examination
Filings in other jurisdictions
Corresponding applications planned or filed under the same specification. Each entry uses the same citation format as the primary PRV filing; application numbers, filing dates, and public-record links are added as they become available.
- Application no. To be assigned Multi-Signal Forensic Pipeline and Compositional Detection System for Synthetic Content and Online Harms Filed with: United States Patent and Trademark Office (USPTO) Filing date: TBD Jurisdiction: United States (US) Status: Planned / not yet filed Public record: USPTO Patent Public Search link will appear here once the application is filed and published.
United States Patent and Trademark Office (USPTO)
TBD
United States (US)
Planned / not yet filed
- Application no. To be assigned Multi-Signal Forensic Pipeline and Compositional Detection System for Synthetic Content and Online Harms Filed with: European Patent Office (EPO) Filing date: TBD Jurisdiction: European Union (EP) Status: Planned / not yet filed Public record: Espacenet link will appear here once the application is filed and published.
European Patent Office (EPO)
European Union (EP)
- Application no. To be assigned Multi-Signal Forensic Pipeline and Compositional Detection System for Synthetic Content and Online Harms Filed with: UK Intellectual Property Office (UKIPO) Filing date: TBD Jurisdiction: United Kingdom (GB) Status: Planned / not yet filed Public record: Ipsum (UK IPO) link will appear here once the application is filed and published.
UK Intellectual Property Office (UKIPO)
United Kingdom (GB)