Moderator handbook — policy, protocol, psychological safety

Content moderation is high-stakes work. OraData offers this service to AI companies and social networks. These are the non-negotiable rules every moderator must know before their first batch.

1. Policy — you apply the client's, not your own

OraData doesn't invent the policy. Each client (social network, AI company) ships their own grid: what's allowed, what's banned, the gray zones. Your job is to apply consistently, not debate.

• Read the batch policy BEFORE touching the first item.
• Unsure on a case → mark "unsure" and pass to an arbiter — never guess.
• Your personal opinions (political, religious, moral) stay out of the batch. If a batch directly clashes with your values, you can refuse — your right, no penalty.

2. The 5 standard categories

• Hate speech: targets a protected group (ethnicity, religion, orientation, disability). Satire vs harassment = judgment.
• Violence: real vs artistic vs news documentation. Default: document public info, flag gratuity.
• NSFW: nudity. Historical art → keep. Explicit → per client policy. Minors → CSAM protocol (§4).
• Misinformation: objectively falsifiable. Verified false → label. Contestable opinion → out of scope.
• Self-harm / suicide: always priority — resource banner + urgent moderation, never ignore.

3. Double-blind decision protocol

Any irreversible decision (permanent delete, uploader ban) requires double-blind validation: two qualified moderators, neither sees the other's call. Disagreement → senior arbiter. Minimum inter-moderator agreement: Cohen's Kappa ≥ 0.75.

4. CSAM — immediate escalation protocol

5. Psychological safety — non-negotiable

• Daily exposure cap: enforced by the platform. Once hit, you cannot accept more moderation tasks today.
• Mandatory break every N decisions/hour (dynamic per batch difficulty).
• Mandatory rotation after a graphic batch — 24h minimum before another sensitive batch.
• Persistent stress: Settings → Moderator wellbeing → free consultation with partner psychologist (video, your choice of language).

6. Confidentiality

• NDA signed before your first batch. Content seen during moderation stays confidential — can't discuss outside a session, not even with loved ones.
• Client request to reveal moderator identity: denied unless legally documented.
• Conflict of interest (you recognize someone): signal support immediately, don't moderate.

7. Pay

Moderation pays per validated batch (volume + inter-moderator agreement). Extreme content carries a 1.5× multiplier. Quality bonus if your consistency score on a batch exceeds 95%. You can refuse any batch with no explanation or penalty; forced exposure does not exist on OraData.

8. Chain of responsibility

• Client: defines policy + endorses final decisions.
• OraData: provides platform, trained moderators, traceability.
• Moderator: applies policy + flags gray zones.
• If a decision is legally challenged, all 3 signatures are on it. No single link is the sole bearer.

9. Continuous training

Mandatory refresher every 3 months: policy re-read + new training module + psy check-in. You can refuse a campaign whose policy you don't endorse — no penalty, no tier drop.

10. Legal frameworks you should know

Content you moderate may fall under different legal regimes depending on the uploader's and client's country. You are not a lawyer — but you must know when to escalate.

• European Union: DSA (Digital Services Act) + GDPR. Major platforms must respond to illegality reports within 24h.
• United States: Section 230 protects the platform + NCMEC CSAM reporting mandatory within 90 days (18 USC § 2258A).
• France / Belgium / Switzerland: CSAM escalation via pharos.gouv.fr + Child Focus + OraData hotline contact.
• CEMAC: country-specific content-illegality legislation is still being formalized. In doubt, escalate to an OraData arbiter.
• Every country: distributing images of minors without parental consent = universal criminal offense. No cultural exception.

11. Decision tree — graphic violence

1. Is it public-interest documentation (armed conflict, disaster, public accident)? → Keep + add trigger warning + blur non-consented identities.
2. Is it artistic representation (film, theater, fiction)? → Apply age-rating per client policy. Keep if rating is consistent.
3. Is it gratuitous violence targeting an identifiable real person (filmed beating for humiliation, torture)? → Block immediately + flag uploader.
4. Is it glorifying (celebrating a violent act, inciting imitation)? → Block + immediate uploader ban.
5. Ambiguous? → Mark unsure, pass to arbiter. NEVER GUESS.

12. Decision tree — hate speech

1. Does it target a group protected by client policy (ethnicity, religion, orientation, disability, gender, nationality)?
2. If no → out of hate scope, may fit another category.
3. If yes → is it factual (statistic, neutral observation)? → Keep.
4. Is it satirical with a target clearly framed by context (recognized comedian, editorial context)? → Keep + flag for client review if borderline.
5. Does it directly incite violence or discrimination ("they should be", "death to")? → Block immediately + flag uploader.
6. Is it defamation / targeted harassment of an identified individual? → Block + refer to legal support via arbiter.
7. Language you don't fluently understand? → Don't guess, hand off to a native-speaker moderator.

13. Decision tree — misinformation

Misinformation = factually verifiable claim that turns out false. Contestable opinion ≠ misinformation, stays out of scope.

1. Is the claim objectively verifiable (public statistic, historical fact, consensus science)?
2. If not (opinion, prediction, interpretation) → out of scope.
3. If yes → is it verified false by a reliable source (WHO, peer-reviewed journals, recognized fact-checker like AFP/Reuters/Africa Check)?
4. If verified false → add "disputed content" label + link to reliable source. NEVER delete a misinformation post — labelling is enough.
5. If not yet verified → mark unsure, escalate to arbiter who has partner fact-check tool access.
6. Public health / imminent danger (e.g. false info that can kill) → urgent-priority signal — speed matters.

14. Resilience — individual + collective

Moderating difficult content leaves traces, even with every precaution. OraData takes it seriously because a burnt-out moderator makes bad decisions and ends up harming the client as much as themselves.

Recommended personal practices

• Bounded hours: no night moderation, no more than 4h/day continuous.
• Closing ritual at end of session: close tabs, screen off for 10 min, walk outside.
• Do not share content seen with your circle — even to "decompress". Transfers trauma without relieving it.
• Keep a private journal of cases that stuck — not the details, your feelings. Feeds the psy check-ins.

Warning signs to watch in yourself

• Trouble sleeping, recurring nightmares.
• Involuntary flashbacks during the day.
• Irritability, emotional distance from loved ones.
• Avoidance of activities or places tied to what you saw.
• Increased alcohol or other substance use.
• Intrusive thoughts hard to dismiss.

Peer-to-peer support

• Moderator WhatsApp channel: peer-support community moderated by a partner psychologist. Request access via support after your first batch.
• Voluntary buddy: request a buddy moderator for mirrored sessions — not the same decision, just the presence of someone who understands.
• Monthly meetings (optional): anonymized experience sharing in video call, hosted by a psy.
• If you spot signs in a peer: tell them you're worried, and softly signal support. You can save a teammate without betraying them.

15. Your rights as a moderator

• Right to refuse a batch: any time, no explanation, no penalty. No tier drop.
• Right to confidentiality: your identity is never shared with the client without your explicit written consent + legally documented request.
• Right to pause: put your moderator account on indefinite pause from Settings. Resume after a refresher.
• Right to free psy consultation: covered by OraData, unlimited in time, independent of a support ticket.
• Right to partial-batch pay if released: if opening a batch reveals content beyond your tolerance, release it — you are paid for time effectively spent up to release.
• Right to unionize: OraData doesn't oppose any collective moderator structure. Legally, culturally — you are an independent worker with the protections your country affords.

Burn in before your first batch

• Apply client policy, not yours.
• Mark unsure when hesitating — never guess.
• CSAM → hotline, never delete alone.
• Self-harm → resource banner + urgent.
• Double-blind required for irreversible decisions.
• Daily cap enforced — never bypass.
• NDA signed + strict confidentiality.
• Conflict of interest → signal, don't moderate.
• Justification on every decision.
• Persistent stress → talk to support. Not weakness.

OraData · guide public · révisé 2026

Photo de couverture : Markus Winkler · Unsplash