# Data protection impact assessment: anti-trafficking counts and the observatory

*Bible Study Helper (biblestudyhelper.org). Written 19 October 2026 for the changes released that day. Review it again before
any new kind of data is counted, before the first data-sharing agreement is signed, and in any case by April 2027.*

This follows the usual structure of a DPIA (GDPR Article 35; the UK ICO template). It is written by the operator for the owner
and for counsel to review; it is not legal advice.

## 1. Why a DPIA

The site's anti-trafficking tools are used by people who may be in danger: migrant workers, people offered jobs abroad, their
families, and people held in forced labour. On 19 October 2026 the anonymous counts behind those tools were made richer so that
anti-trafficking partners can see what people in each country need. The people are vulnerable, the subject is sensitive (forced
labour, sexual exploitation, faith), and the data may be shared with organisations outside the site. That is enough to call for
an assessment even though no record identifies anyone.

## 2. What is processed

### 2.1 What is counted (the `safety_events` table, backend `safety.py`)

One row per use of a tool. No visitor id, no session id, no name, no contact detail, no IP address, no text. Each row holds:

| Field | Content | Source |
|---|---|---|
| event | which tool or step (for example `quick_check`, `chat_topic`, `pulse_answer`) | the page or the server |
| risk | result level (high / caution / none; for the screen: urgent / concern) | worked out on the phone or by the server |
| flags | **new**: the signs marked in a self-check, each as yes / not sure / no (`fee_to_worker`, `?wrong_visa`, `-too_good`); signs found by a job-offer or contract check; signs seen by the safety screen | page / server |
| lang | the page's language | page |
| country | the phone's country, from the network (Cloudflare) | server |
| dest | a destination country the person chose, or the AI read from an offer | page / server |
| detail | which result or entry point; **new** for chat and flagged questions: `<topic>:<stage>:<kind of work>`, for example `wages:abroad:domestic`; for one-tap questions `q2:yes`; for "What will you do now?" `<tool>:<answer>`, for example `quick:nopay` | page / server |
| origin | **new**: where the person said they come from, only when they chose it in the optional list | page |
| channel | **new**: how the visit reached the site (an ad, a shared link, search, direct), as the page names it | page |
| created_at | **new rule**: the time to the hour only (for example `2026-10-19T16:00:00`) | server |

Since 19 October 2026 the four self-checks (quick job check, 11 signs of forced labour, debt check, marriage-offer check) record
which signs were marked; the private chat and flagged Bible questions record the main subject of each message (a label from a
small AI model, never the words); and visitors in labour-migration countries can answer one optional one-tap question a day.
Later the same day two more optional one-tap answers were added: "What will you do now?" after a job-offer, contract or agency
check showed warning signs (`decision_shown` / `decision`, with the tool and the result), and the recruitment cost: "Have you or
your family ever paid money to get a job abroad?" (one-tap question q8), then the cost as a band of months of pay abroad and the
region of the job, stored in `survey_answers` (lab.py: who paid, band, region, where it was asked, traffic source, the phone's
country, language, the time to the hour; no identifier). Before, that survey asked for amounts and currencies.
Since 4 November 2026 the same risky results also offer "Help stop this recruiter": the person picks where the offer reached
them and reads that platform's own report steps, or taps their country's complaint office or anti-trafficking line. Only which
of these they opened is counted (`recruiter_report`, detail "<tool>:<platform|gov|line>", with the result's risk level); the
report itself is made by the person on the platform or with the authorities, and nothing about the offer or the recruiter is
sent to us.

### 2.1b Added 4 Nov 2026 (plan "Ending Human Trafficking")

- New anonymous counts in `safety_events`: "Send it to one person" (onep_send), the scam-company path opened and what was tapped
  in it (scam_view / scam_tap), the helpline summary sent from the person's own phone (summary_send: WhatsApp / text message),
  "Report the account" opened after a fee was asked (fee_report: tool, country, provider or "app"), the weekly check-in screen and
  its two answers (checkin_view / checkin_ok / checkin_help), weekly check-in / warning turned on (checkin_on / watch_on), the
  church kit's Sunday warning copied / shared. Same rules: no id, no words, time to the hour.
- `shared_offers` gains: platform (the one the person chose), the recruiter identifiers picked out by rules when it was sent
  (`ids`; no person reads offers: the owner has no review team), state `auto`, and kind (review = sent with the new consent
  words; warn = the older "help warn others" share). `recruiter_block` = identifiers taken off the list for good after a
  complaint (the identifier only).
- `push_subs` gains `want` (verse / checkin / warn); `push_tally` counts checkin_shown / checkin_ok / checkin_help / warn_shown /
  warn_open per day and language.
- `alerts` gains automatic disaster warnings (approved_via auto) and wave warnings (kind wave, fixed words per country).

### 2.1c Added 8 Nov 2026 (what happened later, calls that got through)

- **"What happened later?" (opt-in)**: after a risky result of the job checks and the answer to "What will you do now?", the page
  offers to ask again in a few days. Only after "Yes, ask me later" does the phone keep a small note in its own storage
  (`bsh_fu`: which check, the result, the day, the country of the job; no words, no id). Nothing is sent when it is kept. On a
  visit 3+ days later one question asks what happened (did not go / still deciding / went and safe / went and need help); the
  answer is counted with no id (`followup`, detail tool:answer:1|2) and the note is deleted (after "still deciding" it is asked
  once more a week later, then deleted). Never on help or child-protection pages (they store nothing), never in privacy mode;
  "Delete my data" and privacy mode remove the note; notes older than ~45 days are dropped unasked. Offer and answer counts:
  `followup_optin` (tool:yes|no), `followup_shown`.
- **"Did you get through?"**: after a tap on a help number in the tools the page remembers the tap in memory only and, when the
  person comes back 12 s to 30 min later, asks one question (helped / someone answered / no answer / the number did not work).
  Counted as `call_result` (detail country:kind:answer, no id, no number dialled beyond the kind). Nothing is stored.
- **q9** (one-tap question): "Has a warning on this site ever stopped you or someone in your family from taking a risky job?"
  (me / family / no), like the other one-tap questions.
- **Neutral partner tag**: links to partner websites carry `utm_source=bsh&utm_medium=referral` (never the site's name, so a link
  cannot expose a reader in a hostile country); government pages are left as they are. Nothing about the person is added.

### 2.2 What is not stored

- The words of any private-chat message, Bible question flagged by the safety screen, job offer, contract or witness report.
- Any identifier on safety counts. A count cannot be joined to a visit, a session or a Mixpanel / Google Analytics record.
- Anything from the help pages beyond anonymous counts (no cookie, no storage, no id, no traffic source).

### 2.3 Where it goes

- The counts stay in the site's database (SQLite on a Render disk; daily backup on the server, weekly encrypted copy off-site).
- The private observatory (`GET /api/admin/observatory`, the page `/data-room/observatory/`) needs the admin key and returns
  totals only, with every number under 10 hidden.
- Partners may receive totals by country and month under a written agreement (see the template). No row-level data leaves the
  site.
- The subject labels are made by OpenAI's gpt-4o-mini from the message, in the same call that already screens every message for
  danger (OpenAI's API terms: not used for training; may be kept up to 30 days for abuse monitoring).

### 2.4 Volume and retention

Thousands of counts a day, nearly all from paid-ad visitors in South Asia, the Middle East and Africa. Safety counts have no
deletion date today (they are anonymous). Shared job offers (opt-in, text) are deleted at 12 months, check-in records at 400 days
unused, as before.

## 3. Purposes

1. **Protection**: warn people before they pay or travel, route them to help, and improve the tools where people get stuck.
2. **Prevention insight for partners**: which warning signs and needs appear in which countries and corridors, so hotlines,
   NGOs and churches can focus their work.
3. **Evidence for funders and buyers of the platform**: that the tools are used and what they find, without describing any person.

Not purposes: profiling individuals, advertising to individuals, immigration enforcement against workers, or selling records.

## 4. Necessity and proportionality

- **Minimum data**: categories instead of text; yes / no / not sure instead of free answers; the origin only when chosen; times
  to the hour; no identifiers. Each field answers a question partners asked or will ask (which signs, which routes, who the
  counts come from).
- **Transparency**: before collection the site's words were changed. The chat now says "We don't keep what you write ... We count
  only the topic (for example pay or passport), with nothing that identifies you"; the self-checks say "Your answers are not linked
  to you: we keep only anonymous counts of the signs people mark, to help warn others"; the one-tap card, the cost question and
  "What will you do now?" say "Anonymous: we only count the answers, to help warn others"; the privacy policy has two new bullets (`lg6`, `lg7`) in all 50 languages, and the
  safeguarding page describes each item.
- **Choice**: the one-tap question has Skip and is never repeated once answered or skipped; the origin and destination lists are
  optional and start empty; every tool works without them.
- **Legal basis (for counsel)**: the counts are designed to be anonymous data. Where any law treats the moment of collection as
  processing personal data (the network address is seen by the server for a moment to derive the country, as for every request),
  the proposed basis is legitimate interests (protection of vulnerable people), with the balancing test in this document. Faith
  and safety contexts can be special-category data in some laws; nothing here records a person's faith or links a person to a sign.

## 5. Risks and measures

Likelihood and severity: low / medium / high. "Residual" = after the measures.

| # | Risk | Before | Measures | Residual |
|---|---|---|---|---|
| 1 | **Linking a count to a visit** by time, country and language (visits and sessions carry ids and exact times) | medium / high | Times kept to the hour since 19 Oct 2026; no ids; counts stored apart from visits; observatory returns totals only | low |
| 2 | **Small-number inference** about a small community ("the 3 people in country X who marked violence") | medium / high | Every number under 10 hidden; percentages only when both numbers are 10+; periods of at least 30 days; countries under 10 combined; the agreement forbids re-identification and combining with other data | low |
| 3 | **A watched phone** revealing that someone looked at safety help | medium / high | Tools are `data-private` (no click details), quick exit and Esc wipe the page; the phone keeps only which one-tap questions were asked (never answers) and the date the card was shown; help pages store nothing | low |
| 4 | **Partner misuse**, for example enforcement against migrants on a corridor, or publishing small numbers | low / high | Agreement limits purposes (protection and prevention only; no enforcement against workers, no profiling, no resale), forbids onward sharing and re-identification, keeps the under-10 rule in any publication, allows audit and termination | low |
| 5 | **Wrong conclusions** from AI labels or a self-selected group | high / medium | Method note on every output (not a prevalence estimate; mostly paid-ad visitors, with the paid share shown; labels can be wrong; one person may count more than once); partner validation before public use | medium |
| 6 | **People expected "nothing is saved"** in the chat | was high | Wording changed in all 50 languages before collection began; the topic only, never words | low |
| 7 | **Breach of the database** | low / medium | The counts identify no one; admin endpoints need the key; backups encrypted; key rotation is an open item | low |
| 8 | **Children** using the tools | medium / medium | No ids or text; the same anonymous counts; the marriage check counts "under 18 or not freely saying yes" as a sign, never an age | low |
| 9 | **Function creep** (using the counts to target individuals with ads) | low / medium | Purposes above; ads may use only country-level totals (for example which corridors need job-check ads), never individual data, which does not exist | low |
| 10 | **Child tools (from 28 Oct 2026)**: a parent's or child's answers, a pasted chat message, or a child's details exposed | medium / high | The checks run on the phone and send only anonymous counts of the signs marked; "Check a message" sends words only, read once by our server and the AI service, never kept and never seen by a person; no photos are asked for or accepted; the child safety card is drawn on the phone (the family number never leaves it); the guide pages keep no id; the danger screen counts a child situation by kind only, behind a switch tested first | low |
| 11 | **Child sexual abuse material** reaching the site (job-offer screenshots, contract photos) | low / high | Images are never stored; the PhotoDNA check (inert until the owner's key) refuses known material before the AI sees it; reporting steps in docs/CSAM-REPORTING-PROCEDURE.md; legal review of the reporting duty is an open item | low (after review) |

| 12 | **Automatic recruiter list (from 4 Nov 2026; no review team)**: an offer someone sends identifies them, or a real employer's or a victim's number is listed by mistake | medium / high | Only with a ticked consent box whose words say exactly what happens (no person reads it); identifiers picked out by rules (phone numbers, WhatsApp / Telegram / Facebook / Instagram / TikTok accounts, e-mail addresses, websites), never names; listed only when corroborated by people in 2+ countries (3+ offers or job checks, 2+ risky), so one person's offer, including a victim's own number pasted by mistake, is never enough; help-directory numbers and sites, government sites and our own never listed; a complaint takes an identifier off for good (block list); the list carries platform, destination, job type and signs: never the text, the sender's language or country, or anything about the sender; never published; platforms and partners only through their abuse / partner channel or a signed agreement; offers kept at most 12 months (privacy.py rule, totals kept) | low |
| 13 | **Weekly check-ins and warnings on notifications (from 4 Nov 2026)**: a notification is seen by someone who controls the phone | medium / medium | Off by default; the button says to turn it on only if no one else reads the phone; neutral words ("Are you safe this week?"); the push carries no content (the phone shows words it already has); the server keeps only the notification address, language, send hour and what was asked for (verse / check-in / warning); answers are counted with no address or id; never offered on help or child pages | low |
| 14 | **Public signals page and automatic warnings (from 4 Nov 2026)**: a published number or a wave warning points at a person, a village or an employer | low / high | Monthly country totals only; every count under 10 hidden and shares shown only when both numbers are 10 or more; routes only at 10+; wave warnings are fixed general words per country ("many people in your country told us...") from at least 12 answers in 7 days, never a name, place below country level, number about anyone, or employer; disaster warnings are fixed safety advice | low |
| 15 | **What happened later / did you get through (from 8 Nov 2026)**: the note on the phone, or the later question, reveals to someone who controls the phone that the person checked a job offer | low / medium | Only after an explicit "Yes, ask me later" (the offer says what is kept); the note holds no words or names; never on help or child pages, never in privacy mode; the later question sits below the first screen, can be skipped, and is deleted after the answer; the call question is in memory only and asks nothing about the call's content | low |

## 6. Remaining decisions (owner, counsel)

1. **Older rows**: counts before 19 October 2026 keep their exact time (to the second). A one-line update can cut them to the hour
   too (irreversible). Recommended; not done without the owner's approval. Same for older answers to the recruitment-cost survey.
2. **Counsel**: confirm the legal basis per main market (EU/UK, India DPDP, Pakistan, Nigeria NDPA, Kenya DPA), whether an EU
   representative is needed, and the agreement's liability, law and jurisdiction clauses.
3. **No native-speaker review** (owner's decision, 4 Nov 2026): the questions and the privacy wording in every language are checked
   automatically (letters, numbers, placeholders, leftover English) and key privacy lines are read; languages the models cannot
   write (Tigrinya) stay off the site.
4. **Partner validation**: before any public use, a partner checks a brief against what they see on the ground.
5. **Retention for safety counts**: none today (anonymous). Consider a 5-year limit with yearly totals kept, like other tables.
6. **Child tools (28 Oct 2026)**: counsel to confirm (a) whether automated processing of a pasted message that shows a child being
   enticed creates a reporting duty although no person sees it and nothing is kept, and (b) the reporting and preservation steps
   once the image check is switched on (docs/CSAM-REPORTING-PROCEDURE.md).

7. **Automatic recruiter list (4 Nov 2026)**: no person reads offers (the owner has no review team); counsel to confirm that
   passing a corroborated recruiter phone number or account to the platform hosting it, as a report of suspected fraud, is
   covered by legitimate interest in each main market without a human check, and the wording of the partner annex for
   identifiers. The corroboration rule (2+ countries, 3+ reports, 2+ risky) is the safeguard instead of a reviewer.
8. **Public signals page (4 Nov 2026)**: built with the stricter rule (cells under 10 hidden, shares only from 10/10). A partner
   should check the first month's page against what they see before it is promoted.

## 7. Sign-off

| Role | Name | Decision | Date |
|---|---|---|---|
| Owner | | approve / changes needed | |
| Counsel (optional) | | | |

Next review: before the first data-sharing agreement, or April 2027 at the latest.
