A report reaches the safety office: approach interrupted at four hundred feet, wind shear, go-around performed correctly. No name on the form, the confidential-handling box is ticked. Three fields further down, the date, the registration and the flight number are enough to identify the crew in two minutes, flight plan in hand. The whole subject lies in that gap: a report can be formally anonymous and still be personal data under the law, without ceasing to be the raw material of safety analysis.
Anonymous on the form, identifiable in fact
Identifiability defines personal data, not just names. A person is concerned if they can be identified directly or indirectly with reasonable means. The test isn’t “was a name written?” but “what would it take, with the company’s resources, to trace this back to a person?” This changes the diagnosis.
Resources for identification are abundant and legitimate: scheduling, flight-time tracking, recorded flight data, maintenance logs. A quadruplet of date, registration, sector and role on board forms an almost certain key, and the narrative supplies others, more discreet and more effective.
The size of the population decides everything. On a long-haul network with several thousand crew members, a sector and a date leave genuine ambiguity; on a fleet of six aircraft and forty crews, the same pair points to one person and one only. Same form, same wording, two opposite levels of risk: a de-identification policy copied from a large operator protects next to nothing at a small one.
On the maintenance side, the setup is different and rarely addressed. The certificate of release to service carries a stamp or an approval number: directly identifying data, mandatory, impossible to strip from the airworthiness record. Two regimes then overlap on a single event — the certification trail, nominative and kept for years, and the safety-analysis record, which has no need to be. Merging them into a single database is the starting point of most drift.
Even without names, details like flight numbers, dates, and aircraft types can identify people. The more specific the information, the easier it is to trace back to an individual, especially in smaller organizations.
In a regional airline with 50 pilots, a report mentioning “Flight 123 on June 15th, captain with 8,000 hours” might identify a single person. The same report in a major airline with 5,000 pilots would be much harder to trace.
Four states of the data, only one is irreversible
The vocabulary is a trap and the three words commonly used are not synonyms. Pseudonymisation replaces identifiers with a code, with the correspondence table kept separately: the data remains personal, since the link exists somewhere. De-identification, a term specific to the world of occurrence reports, strips the record of what allows tracing back to a person, and depending on the case, the operator’s name. Anonymisation is irreversible and takes the data out of the scope of data protection regulation; aggregation is its common form.
The decisive point is that the occurrence-reporting regime requires de-identification, not anonymisation. The nuance is not academic: the system is designed so that one can go back to the source for as long as the analysis needs it. That return path must therefore be framed, not claimed to be impossible.
| State of the data | Return to the person | Who accesses it | Proper use |
|---|---|---|---|
| Nominative: raw narrative, reporter’s contact details | Immediate | Designated point of contact, alone | Acknowledgement of receipt, follow-up questions |
| Pseudonymised: code, separate correspondence table | Possible, via the table, on justified grounds | Authorised analyst | Investigation, cross-referencing of events |
| De-identified: quasi-identifiers removed or degraded | Possible through external cross-referencing | Safety committee, business-line contacts | Trend study, corrective action |
| Anonymised or aggregated: indicators, event families | Excluded | Wide distribution, publication | Dashboard, bulletin, training |
The third row does the most damage, because it is everywhere treated as though it were the fourth: a de-identified record circulates as an attachment, gets exported to a spreadsheet, is projected in a meeting, while cross-referencing remains within reach of a good part of the people present. Data is only anonymous relative to those who receive it.
Flight data follows the same logic, and has for longer
A flight data monitoring programme produces, for every flight, a trace that identifies the aircraft and the time, hence the crew. De-identification is organised upstream there: matching a flight to a crew is only possible by a designated person, under written conditions negotiated with staff representatives, the agreement carrying a commitment not to use the data for disciplinary purposes outside deliberate serious misconduct. This arrangement predates the general data protection regulation and often serves as its internal model: a single role empowered to lift the link, the reason logged, the matching key kept for a short time. There is no justification for the occurrence-report database to be less well protected than the same airline’s flight-data database.
The legitimate re-identification window
A reporting system in which one can never go back to the reporter is a dead system. A good part of the value lies in the follow-up question: what was the setting, did you have the information before departure, did the checklist procedure apply as written. Without that return path, the analysis is limited to whatever the form happened to capture.
Hence the notion of a window. Between receipt and the closure of the investigation, the link to the reporter remains accessible to an identified person, for a single purpose: understanding the event. Past that point, the link is cut and the file moves into the next layer. Three conditions make the window acceptable and can be checked by reading the procedures: the role empowered to lift the link designates a precise function, not an entire department; every lift is logged with its reason, which presupposes a register, not a good memory; the reporter is informed in advance of the window and its duration, otherwise they will discover it at the worst possible moment.
What falls outside the window deserves to be written down in black and white, because the request will come: a search from human resources, a section head wanting to know « who did that, » the preparation of a disciplinary file, the curiosity of an internal audit. The right answer is not to refuse verbally when the request lands, but to have stated in writing, beforehand, that the events database is not a personnel-management source.
The right reflex. Put the sentence describing the window into the form itself: who may contact you again, for how long, and what happens to the link afterwards. That is the only place the reporter will actually read it, and this sentence serves both as data-protection information and as a just-culture commitment.
One clock per layer, not a single duration
« How long is a report kept » is a badly framed question. There is not one duration but as many clocks as there are layers, and most of compliance consists in having distinguished them. The reporter’s identity is only useful during the investigation; the free-text narrative keeps its value as long as the event can be matched with others; the coded record feeds trends over several years, the natural horizon of a safety management system; aggregated indicators have no reason to stop.
Two requirements cross here, and people often believe they conflict. Storage limitation forbids keeping personal data beyond the purpose that justified it; the obligation to preserve events, carried by the occurrence-reporting regime and by operational documentation, imposes a long memory. They are reconciled as soon as the layers are separated: what designates no one is kept for a long time, what designates someone is erased early. The contradiction only appears where the file is treated as an indivisible block.
The hard point is not the rule, it is effective deletion. A purge that leaves full backups intact, spreadsheet exports on a shared drive, attachments in mailboxes, paper printouts in a committee’s binder and a copy at an analysis contractor’s site is not a deletion. A retention period only means something if one can enumerate every copy.
Access: need-to-know, written down and logged
Authorisations almost always mirror the org chart: the safety department sees everything, because it is the safety department. The principle is the opposite — the level of detail accessible flows from the task, not from the reporting line. Whoever arbitrates a corrective action needs the nature of the event, its frequency and its contributing factors; neither the raw narrative nor its author, knowledge of whom improves no decision.

Three rules suffice. Profiles are defined by data level, not by department, which requires writing four instead of one. Consultations of the nominative layer are logged and the log reviewed periodically, otherwise it only serves in the event of litigation. Access is withdrawn on transfer — by far the most common failure: in most organisations, the list of people with access to nominative data is a history of past assignments.
Two configurations undo this reasoning and deserve to be treated on their own. The first is the combination of roles, common in small organisations: the safety manager is also the chief pilot, therefore the one who builds the rosters, conducts line checks and prepares assignment decisions. No authorisation rule separates two roles held by the same person, and the reporter knows it before even opening the form. The only workable remedy is to move the nominative layer out of the operator: point of contact entrusted to a third party — occupational physician, contact at an external reporting body, safety manager of a sister organisation — who passes on only the coded record. The second is the outsourcing of analysis. A contractor receiving a full export becomes the recipient of the richest layer, often without the contract saying so: it must state which layer is opened to them, what they are allowed to keep at the end of the assignment, and what happens to the data at the end of the contract. The same reasoning applies to the hosted tool carrying the events database: its provider sees the nominative layer by construction, and technical administration is never a business authorisation.
The right reflex. Extract the list of accounts with access to the nominative layer and cross-check it against the list of people whose job description mentions handling occurrence reports. Any line without a match is withdrawn the same day. The exercise takes about an hour and is worth repeating at regular intervals.
The mosaic, and the free text that builds it
Structured fields are not the main problem: they degrade easily and reproducibly — a date becomes a month, a registration becomes a type, an aerodrome becomes a terrain category, an experience level becomes a bracket. This can be automated and the result is stable.
The free narrative, on the other hand, is irreducible. « After the night stop, second sector of the day, with a first officer new to the line for three weeks »: three innocuous elements that, in a mid-size operation, point to a single person for anyone who knows the network. And the narrative must not be impoverished, since it is precisely what carries the value — the mechanics of the event, the workload, what was perceived and when.
The workable solution consists in producing two texts rather than mutilating the one. The raw narrative stays in the restricted layer, together with the identity, and is erased with it. At closure, the analyst writes a study version: same facts, same sequence, same factors, without the contextual elements unnecessary for understanding. It is this version that moves into the long layer and will be read again years later. De-identification becomes an act of drafting, with an author and a date, instead of an automatic filter that no one is accountable for.
The trap. Believing that a keyword review is enough. What re-identifies is almost never an identifier: it is a detail with no apparent importance — a rare stopover, a baggage incident, an unusual crew pairing, a failure everyone talked about. No automated search catches these. Have the study version proofread by someone who knows the network well, with a single instruction: tell me if you can name the crew. As long as the answer is yes, the work is not finished.
Sharing, publishing, and the day an outside request arrives
Sharing with the authority and with the European central repository of occurrences takes place on de-identified data, and that channel is well managed because it is properly tooled. Lateral sharing is much less so: what gets sent to the manufacturer, the lessor, the workshop, the analysis contractor often leaves as a full export, with the original attachments — flight data, photographs timestamped and geolocated by their metadata, planning screenshots. The rule is applied to the field and bypassed through the attachment.
Internal publication is even more formidable, because it targets precisely the audience capable of recognising. A bulletin case study or training session must be shifted: change the location, the aircraft type, the season, the time of day, as long as the demonstrated mechanism remains intact. If it cannot be shifted without destroying it, then it is too identifying to be published as is.
There remains the outside request: judicial requisition, an insurer’s inquiry, a question from a prospective employer. It is not assessed at the counter and is never handled by exporting the database: it is escalated to a designated function, receives a written response and is subject to its own traceability. The protection attached to occurrence reports, which forbids using them for purposes other than safety, organises these situations without constituting immunity — an internal procedure that presents it as such exposes the operator as much as the reporter.
One last case comes as a surprise: the data subject’s right of access. A crew member may ask what the database holds about them, and the practical consequence is instructive — to answer, one must be able to find what concerns them, therefore be able to re-identify. That is where the balance lies: the nominative layer must be small, brief and perfectly inventoried; the rest must be out of reach because it is genuinely non-identifying, not because someone decided to call it anonymous.
What to check in your own documentation
Seven points can be checked by reading the safety management manual and the register of processing activities, with no particular legal expertise: the written purpose of the processing, phrased in safety terms rather than general ones; the inventory of layers with a duration for each; the name of the function authorised to lift the link to the reporter; the existence and periodic review of the log of lifts; the written circuit for outside requests; the information given to staff; the impact assessment, expected as soon as processing is large-scale or concerns facts that may have consequences for individuals.
A grey area remains, and it is structural. The boundary between de-identified and anonymous depends on context, therefore on fleet size, number of crews, outsourcing of analysis, a transfer or merger. A de-identification that is correct for fifty crew members is no longer correct after a transfer of activity that leaves fifteen on the type. Nothing requires this threshold to be periodically reassessed and no one does so spontaneously: this is where the next audit findings will be found, a subject that is beginning to circulate among safety managers on the FL330 professional forum.
Anyone preparing to take up a role in flight safety or compliance should expect to have to explain the link between just culture and data protection. The answer is not legal: a system that does not technically protect the reporter will soon collect only the events no one is afraid of — and the absence of sensitive reports gets presented as progress.
The applicable text, the decisions of the competent supervisory authority and the operator’s approved documentation prevail in all circumstances over this article, whose purpose is limited to clarifying their mechanisms.
Questions fréquentes
Is a report with no name on it personal data?
Yes, as long as the person remains indirectly identifiable. The date, the registration, the sector and the role on board form an immediate matching key for anyone with access to scheduling or flight tracking. The legal test concerns identifiability by means reasonably likely to be used, not the presence of a surname on the form. The same form can therefore be anonymous at a large airline and perfectly identifying at a forty-crew operation.
What is the difference between de-identification and anonymisation in an occurrence-reporting system?
De-identification strips the record of the elements that allow tracing back to a person, but the return remains possible for whoever holds the key or can cross-reference with other internal sources. Anonymisation is irreversible and takes the data out of the scope of data protection regulation. The occurrence-reporting regime requires de-identification precisely because the analysis must be able to return to the source for as long as it needs to. It is therefore that return path which must be framed, not claimed to be impossible.
How long should an occurrence report be kept?
The question arises layer by layer, not for the file as a whole. The reporter’s identity is only useful during the investigation and is erased at closure; the raw narrative is kept for as long as cross-referencing is needed; the coded record feeds trends over several years; aggregated indicators can be kept indefinitely since they designate no one. The retention period chosen for each layer must appear in the operator’s documentation, and deletion must cover every copy, including exports and backups.
Who can lift a reporter's anonymity, and under what conditions?
A function explicitly named in the procedure, and it alone, for a single purpose: understanding the event. Every lift must be logged with its reason in a register reviewed periodically. A request coming from human resources, a line manager or a disciplinary procedure does not belong to this circuit and must be handled outside the safety management system’s database. The reporter must have been informed in advance of the existence of this possibility and of its duration.
How can re-identification through the free-text narrative be avoided without impoverishing the analysis?
By producing two texts instead of one. The raw narrative stays in the restricted-access layer and is erased along with the reporter’s identity. At closure, the analyst writes a study version that keeps the facts, the sequence and the contributing factors, but removes the contextual elements unnecessary for understanding. This study version is the only one that moves into long-term retention, and it is prudent to have it reviewed by someone who knows the network, with the control question: can you name the crew?
Can a case study drawn from a real event be used in training?
Yes, provided it is shifted enough that it no longer designates anyone in front of an audience that knows the network. The location, the aircraft type, the season and the time of day are changed while keeping the demonstrated mechanism intact. If the case loses its educational value as soon as it is shifted, it is too identifying to be published as is and must be dropped. Internal publication is the most effective vector of re-identification, because it addresses precisely the people capable of recognising the situation.

Leave a Reply
You must be logged in to post a comment.