The true story of a detector we're building for something nobody has managed to hear yet.
School bullying causes well-documented, measurable, undeniable harm. Parents see the aftermath. Teachers see the aftermath. Even the victim, sometimes, has no words for what is happening — until suddenly they have far too many.
The question we asked ourselves was simple, and a little reckless: can we see it coming before it arrives?
Detecting "shut up" or "I'll kill you" in a message: that's solved. We do it. A filter written in an hour does it. Dozens of companies have been doing it for years.
That's not what destroys children.
The bullying that lasts for months — the kind nobody fully recovers from — leaves almost no direct trace. It lives in what is not said. In the joke nobody quite understands. In the message that goes unanswered. In the conversation you stop being invited to.
Review of 109 scientific publications · Shieldy, June 2026We called this The Whisper. And we decided to teach a machine to hear it. Not by spying on children — by learning to recognise the shape a conversation takes when something is wrong. The way a doctor recognises a fracture on an X-ray without touching the bone.
We read the work of researchers who spent their careers on a related question: when someone is losing their footing, what do they say — and how do they say it? We translated their theories into automatic measurements.
| Family | What we measure in the conversation | Scientific reference |
|---|---|---|
| B1 — Irreversibility | "Never again", "anyway", language that closes doors | Thomas Joiner (2005) |
| B2 — Body language | "I can't take it anymore", "it hurts", somatic distress signals | Wilfred Bion (1962) |
| B3 — Timing | Messages at 3am, silences longer than 2h, rhythm breaks | Klein (2026) |
| B4 — Mentalisation | Inability to name others' emotions in text | Peter Fonagy (2004) |
| B5 — Reciprocity | Someone who replied to everything — and suddenly doesn't | Joiner (2005) |
Each run is a full test on conversations the system has never seen before. Two numbers matter: how many bullying situations we detect, and how often we raise a false alarm.
| Run | What changed | Detection | False alarms | Verdict |
|---|---|---|---|---|
| Baseline | Rules-based engine (pre-ML) | 10% | 4% | Blind |
| Run 1 | First ML — text meaning only | 69% | 45% | Too many alarms |
| Run 2 | + real French conversations | ~60% | 10% | Progress |
| Run 4 | + B3 timing (time, silences) | 61% | 5% | Best at the time |
| Run 5 | + B5 reciprocity (too little data) | 58% | 21% | Regressive |
| Run 6 ★ | + 5,608 real French teen messages | 62% | 2% | Champion |
| Run 7–8 | + toxic tweets (wrong idea) | 75–80% | 25–27% | Regressive |
| Run 9 | + CamemBERT (French language model) | 61% | 4% | Redundant |
| Run 10 | LightGBM (stronger algorithm) | 73% | 16% | False alarms too high |
| Run 11 | Two-stage architecture (Axel Delaval's suggestion) | 62% | 2% | = champion |
Detection = % of true bullying situations detected · False alarms = % of normal conversations flagged incorrectly
Test set: conversations generated by Gemini, ChatGPT and Claude — never seen during training
⚠️ Update 17/06: these rates are on a synthetic/balanced set (optimistic). On real everyday French, false alarms are much higher — ~34%, down to ~10% after a fix. See the dossier.
The red line must stay low. The green line must stay high. Run 6 is the only run where both conditions hold simultaneously on this (synthetic, balanced) test set — not yet on real everyday data (see update above).
We're looking for the bottom-right corner: detect a lot, get it wrong rarely. Run 6 is the only one to have reached it.
We were looking for a victim who gradually "declines". That is what the literature described. What we found, analysing real WhatsApp conversations between Italian teenagers, was different — and more brutal.
In 8 conversations out of 10, the victim was totally absent — completely silent — during 46 to 61% of offensive messages. The group continued talking about her while she was no longer there to respond. This is not a decline. It is a rupture.
A real conversation thread. The victim's presence (green) drops suddenly while offensive messages continue.
The Whisper is not someone complaining less and less loudly. It is someone who has ceased to exist in the conversation — while others continue talking about them.
This distinction changes everything about detection. And it is, to our knowledge, the first time it has been documented in this way.
To go further, we need more real adolescent conversations. Here is where things stand:
| Dataset | Country | N | Use for Shieldy | Status |
|---|---|---|---|---|
| CyberAgressionAdo-Large | 🇫🇷 | 5,608 msgs | Real French teen conversations — current engine | In production |
| ECLS-K | 🇺🇸 | 4,054 children | 26% on chronic trajectory — calibration | Exploited |
| TRAILS | 🇳🇱 | 2,230 adolescents | Temperament at 11 → trajectory at 13 | Request submitted |
| ALSPAC | 🇬🇧 | 14,541 | Attachment + longitudinal bullying | In preparation |
| krisenchat / SNARE | 🇩🇪 🇳🇱 | 10,000+ | Real teen crisis conversations | Contact ongoing |
We know how to detect a whisper in a conversation. But can we see, as early as age 11, which child is at risk of becoming a chronic victim at 13 — before bullying has even begun?
Dutch researchers followed 2,230 adolescents for 24 years. They measured, among other things, each child's capacity to understand what others feel. Our hypothesis: children with low mentalisation capacity at age 11 are disproportionately represented among chronic victimisation trajectories at age 13.
Hypothesis H2 · TRAILS application submitted June 13, 2026 · Reference: Fonagy & Luyten (2009)If this holds, it changes the logic of prevention entirely. Not after the crisis — before. Not targeting the victim — targeting the whole class, by strengthening mentalisation from age 10–11.
We are waiting for the data. It is the kind of wait that is worth something.
