Summary

  • AFRINIC reports integrating xAPI into every Academy lesson and tracking 3,927,253 learning behaviours, but that number is neither a learner count nor proof of what fields were collected.
  • The useful remedy is not a privacy-law verdict. It is a thin, versioned learner receipt that binds event vocabulary, purpose, identity treatment, processors, retention and correction rights to the policy accepted at signup.

There are two records at the entrance to AFRINIC’s online Academy. One is conspicuously exact. AFRINIC’s 2025 Annual Report says the capacity-building department integrated xAPI into each lesson and tracked 3,927,253 learning behaviours. The Academy’s public report repeats the count and says the resulting fine-grained data powers an “Onboard → Engage → Drop” funnel for continuous course improvement.

The other record is hard to see. The public registration response tells a prospective learner that creating an account means agreeing to the Terms of Service and Privacy Policy. In the HTML captured for this review, those links point to modal containers that hold headings and close buttons but no readable policy body. The Academy-specific Terms and Privacy pages likewise serve titles and site chrome without a readable policy text; their public WordPress REST representations return empty rendered content and excerpts.

That juxtaposition is the story, but it needs care. It does not establish that AFRINIC has no privacy policy. AFRINIC publishes a substantive global policy. It does not establish that an opened modal is blank in every browser state; the post-click state could not be reliably captured. It does not establish that 3,927,253 events identify people, or even that each counted behaviour corresponds to one unique xAPI statement. It certainly does not establish a breach, unlawful processing or harm.

What it establishes is narrower and more operational. An institution is confident enough in a measurement system to publish a seven-digit output, yet the public path by which a learner is asked to accept the product-specific rules does not explain, in the checked response, what record sits behind that output. This is an evidence seam, not a verdict.

Start with AFRINIC’s best defence

The strongest account in AFRINIC’s favour is straightforward. Fine-grained learning analytics can be useful. A course team needs to know where learners stop, which lessons take too long, whether examination design creates avoidable friction and where a language correction changes completion. AFRINIC says the data is used for course-design refinement and learner-experience improvement. Those are ordinary, defensible purposes for an Academy whose job is to turn Internet-resource knowledge into operational skill.

The reported surrounding indicators reinforce that case. AFRINIC records 1,717 trainings in 2025, a 54.41% increase, and a median Net Promoter Score of 46.01. English self-paced learning accounted for 1,605 trainings across eight courses, reached 3,135 learners and recorded median NPS 47.1. French activity accounted for 112 trainings across seven courses, reached more than 198 learners and recorded median NPS 37.5. The report also gives a median average course time of three hours, 42 minutes and 32 seconds, a median 69% of final examinations passed within two attempts, and a median 12 lessons completed per learner per week.

These are not absurd things to measure. They may help an instructor decide whether a lesson should be divided, whether an exam is testing knowledge or endurance, and whether a course design works differently in French. The underlying events may be pseudonymous, minimised, aggregated quickly or deleted after a short period. The global privacy policy may govern the Academy even though the product-specific pages render badly. Complete consent records and processor terms may exist behind the public surface. An empty response may be a configuration defect.

Each point matters because criticism becomes weak when it treats every unknown as evidence of the worst case. The public record supports none of the familiar escalations: no demonstrated sale of learner data, no disclosed security incident, no regulator finding, no proven secret profiling and no identified victim. The right question begins only after those claims are set aside.

The number, the standard, the collection and the outcome are four different things

AFRINIC’s own reporting supplies the first layer: 3,927,253 “learning behaviours” were tracked. It does not call them learners, people or personal records. The count could include many granular interactions by the same learner. It could include events later aggregated. It could reflect a reporting convention that is not one-to-one with stored statements. The published number should be repeated exactly and interpreted no further than its label permits.

The official xAPI specification supplies the second layer. An xAPI Statement requires an actor, a verb and an object; it may also carry a result, context and timestamp. That makes xAPI useful for describing experiences across lessons and systems. It also makes the name “xAPI” insufficient as disclosure. A standard defines a possible envelope. It does not disclose AFRINIC’s event vocabulary, whether the actor is a person or pseudonym, which optional fields are used, or what downstream joins are performed.

The actual Academy collection is the third layer. The checked public sources do not provide a complete statement profile, identifier model, processor list, learning-record-store role, storage location or event-level retention schedule. That absence from the reviewed public surfaces is not proof that the information is missing internally. It is evidence that a learner cannot reconstruct the operating boundary from the same public route that asks for agreement.

The fourth layer is outcome. Event volume is not learning. Nor are trainings, NPS, time-on-course, exam attempts or weekly lesson counts interchangeable with learning. Each can be informative, but each has a different denominator and failure mode. A slower course time might mean depth, difficulty, interruption or poor design. Two exam attempts might mark persistence or ambiguity. A weekly lesson count can rise because a course becomes clearer or because lessons become shorter.

The discipline here is to keep the four layers apart. A published count can be correct while the disclosure is inadequate. A disclosure can be good while the educational inference is weak. A technically rich event can be stored without identifying a learner. A perfectly lawful system can still leave members and learners unable to audit whether the measurement serves its stated purpose.

The signup boundary is where the records should meet

Account creation is not just a user-interface moment. It is the point at which an identity, a policy version and a future stream of observations can become linked. If that join is not versioned, later inspection becomes guesswork.

The Academy’s registration response makes agreement explicit. Yet the checked product-specific pages and modal markup do not expose the text needed to answer basic record questions. AFRINIC’s global privacy policy provides a substantial umbrella: it discusses purposes, data categories, third parties, overseas transfers, organisational or legal retention, and access, correction and deletion. That is meaningful evidence, not a placeholder.

But an umbrella policy cannot, by its title alone, disclose a product’s event grammar. It does not tell an Academy learner which lesson interactions become events; whether identifiers are direct, pseudonymous or aggregated at each stage; which events are used for instruction, support, funnel analysis or institutional reporting; which processor or learning-record store receives them; or when detailed events cease to exist.

This matters even if every answer is benign. “Pseudonymous for 30 days, aggregated thereafter” is a much smaller exposure than “direct identity retained indefinitely”, but the current sources do not establish either. A serious article should not choose between them. A serious institution can publish the boundary.

One visible discrepancy shows why lineage matters

The Academy’s public report contains another useful warning. Its English course card reports 1,605 trainings, eight courses and 3,135 learners. The visible French card carries French NPS 37.5 but repeats the English counts. The Annual Report PDF instead reports 112 French trainings, seven French courses and more than 198 learners.

The discrepancy does not prove corruption in the learning telemetry. It may be a copied display block. The PDF may be the controlled reporting artifact while the web page is stale. Or the two surfaces may have been built from different snapshots. The important point is not which explanation sounds likely. It is that a reader needs version and provenance to know which record governs.

This is the same problem at a smaller scale. A number without lineage creates interpretive debt. When the figure concerns course totals, the cost is a confusing report. When the figure describes millions of learner behaviours, the cost extends to purpose, identity and rights. The institution needs to show not merely what the current number is, but which schema, policy and aggregation rules produced it.

The minimum useful control is a learner receipt

The remedy need not be a new compliance portal or a dump of raw events. Both would be heavier than the evidence requires. The useful control is a compact, versioned learner receipt available at signup and later from the account.

It would identify the xAPI event vocabulary or profile version in use. It would say, by material processing stage, whether the learner is represented directly, pseudonymously or only in aggregate. It would map event classes to purposes: delivering a lesson, improving course design, providing support, measuring a funnel or preparing institutional reports.

It would name the processor or learning-record-store role and material transfer or location boundaries without revealing security-sensitive topology. It would state separate retention or deletion rules for detailed events and aggregates. It would provide routes for access, export, correction and deletion, including honest limits. It would bind the exact Terms and Privacy versions and the consent timestamp to the account record. When a schema or purpose changes, it would preserve a correction or supersession link.

This is a thin receipt because it does not expose names, scores, individual events, credentials or internal system design. It is strong because it makes the joins inspectable. A learner can see what was accepted. An instructor can see what the metrics mean. An auditor can reproduce the policy boundary. AFRINIC can improve its tools without pretending that every historical event was governed by today’s language.

Accountability is a service property

AFRINIC is not merely a course vendor. It is a regional Internet registry whose legitimacy depends on members believing that records, rules and operational behaviour can be reconciled. The Academy sits beside that mission: it helps operators turn policy and protocol knowledge into practice.

That makes the disclosure question institutional rather than cosmetic. A registry that teaches operational discipline should be able to show the discipline of its own measurement system. This is not a demand for perfection or for less learning analytics. It is a demand that the published evidence stack close.

The Academy already reports with unusual specificity. That is an asset. The next increment is not another larger number. It is the small record that explains what the number can and cannot mean.

Sources