The gap is not diligence. It is a purchase order the United States government has honored for fifty years.
A pilot who busts an assigned altitude on a Tuesday afternoon has until Friday of next week to do something strange: write down exactly what happened, in first person, and mail it to NASA. Not to the FAA, which could punish the violation. To NASA, which cannot. Paragraph 12 of FAA Advisory Circular 00-46F explains why the pilot does it: "although a finding of violation may be made, neither a civil penalty nor certificate suspension will be imposed" if the violation was inadvertent, involved no crime or accident, the pilot has five clean years behind them, and the report reached NASA "within 10 days after the violation, or date when the person became aware or should have been aware of the violation."
That paragraph is a purchase order. The United States government buys confessions from aviators, and it pays in waived penalties. In 2025 it bought 121,128 of them, which works out, by NASA's own arithmetic, to 485 per working day (ASRS Program Briefing, p. 14). Since the program opened in 1976, total intake has reached 2,321,050 reports.
Artificial intelligence has a database too. It is called the AI Incident Database, it lives at incidentdatabase.ai, and it is a serious, well-run project that this essay is about to compare unfavorably on volume, so first: the people who maintain it are doing the work the rest of the field declined to do. But the comparison is the point, because the gap between the two corpora is not a gap in diligence. It is a gap in incentives, and you can read the incentives directly off the shape of the files.
Numbers about the AI Incident Database tend to circulate secondhand, so we didn't quote one; we counted. The project publishes full database snapshots, and the newest one (backup-20260824101533.tar.bz2, dated 24 August 2026, a 110 MB mongodump) contains exactly:
The vocabulary matters. In the AIID's schema, an incident is an event: a specific AI system failed in a specific way and touched the world. A report is a document about an event, and the documents are overwhelmingly news articles. Divide one count by the other and you get 4.6 reports per incident: when an AI system fails publicly, several outlets write it up, and the database dutifully files each article against the event.
Now hold that ratio next to aviation's. An ASRS report is a first-person account by someone who was in the seat, at the scope, or on the wrench. The corpus tends toward one document per event per witness, written by the witness. The AIID corpus multiplies documents per event, written by bystanders. One database is confession-shaped. The other is journalism-shaped. You can see this in the file counts before you read a single narrative: 4.6 articles per incident is the fingerprint of a press corpus, not a reporting culture.
The cleanest like-for-like number is documents against documents: 2,321,050 ASRS reports received against 7,604 AIID reports collected, which is 305 to 1. And even that flatters nothing, because the two piles are made of different stuff. Aviation's 2.3 million documents are almost entirely people describing their own mistakes. AI's 7,604 are almost entirely people describing someone else's.
It is tempting to write "121,128 aviation reports last year versus 1,643 AI incidents ever" and let the 74-to-1 ratio ring. Resist the ringing until the labels are on. The 121,128 is intake: everything NASA received. The public, queryable ASRS database is a curated subset of that; NASA categorizes 100 percent of what arrives, then analysts select reports "that require further analysis and entry into the public ASRS database," merging multiple reports on one event into a single record (Program Briefing, p. 19). So ASRS intake is a receipt count, the AIID's 1,643 is a curated event count, and comparing them straight across mixes the two. We did not retrieve the ASRS public database's record total, so this essay prints no event-to-event ratio: the honest statement is that the receipt-to-receipt comparison is 305 to 1, the ratios people usually quote (74 to 1 against last year's intake, 1,413 to 1 against all-time intake) mix a receipt count with an event count, and every number in this paragraph is labeled so you can decide which one you trust.
None of the labeling rescues the gap. Choose the most conservative pairing available and aviation is still two orders of magnitude ahead of AI, in a corpus where the reporter is the person who erred.
One more count from the snapshot deserves its own sentence: the AIID's pending review queue holds twenty submissions. Twenty, against 1,643 published incidents. Whatever throttles AI incident reporting, it is not a backlog of unprocessed candor.
The mechanism on aviation's side is worth reading closely, because the popular word for it, immunity, is subtly wrong in three instructive ways. The text is FAA Advisory Circular 00-46F (issued 2 April 2021, still current at the FAA's circular index as of this writing), and NASA reproduces the operative language on the ASRS site.
First, paragraph 8 walls the database off from enforcement entirely: "The FAA will not seek, and NASA will not release or make available to the FAA, any report filed with NASA under the ASRS or any other information that might reveal the identity of any party involved." That is why the program lives at NASA at all. The agency that punishes cannot be the agency that collects; aviation solved that by handing the mailbox to a rocket agency with no enforcement powers.
Second, paragraph 12 waives the sanction, not the finding: "although a finding of violation may be made, neither a civil penalty nor certificate suspension will be imposed." Your record can still say you did it. What the deal removes is the penalty, not the truth. Anyone designing an AI reporting regime can stop agonizing over whether admission implies absolution, because aviation's answer, in plain print since the Ford administration's era of the program, is that it does not. You can keep the finding and still get the confession.
Third, the waiver has a clock, and the clock is the most underrated line in the circular. The report must be filed within ten days of the violation or of the moment the person should have known about it. Aviation does not merely enjoy short reporting latency, it pays for short latency, by rule. File on day eleven and the waiver is gone. The AI Incident Database has no equivalent lever, so its latency is whatever the news cycle happens to be: an AI failure gets recorded when a journalist notices, which may be days, months, or never.
And the exclusions are precise. Deliberate acts, criminal offenses, accidents, and anything revealing a lack of qualification are carved out of the waiver. Read as an economist would: the scheme buys candor about the near miss and pays nothing for the catastrophe. That is exactly the right region to purchase, because the near miss is where a database has something to teach and a regulator has the least to gain from punishment. The catastrophe will be investigated anyway; the altitude bust that scared everyone and harmed no one evaporates unless someone pays for it to be written down.
Even the fine print teaches. Air traffic controllers are excluded from paragraph 12 and covered by a separate program, ATSAP, negotiated for their circumstances. Aviation could not stretch one immunity scheme across every practitioner in its own industry. An AI regime should expect the same: the deal that works for a model developer will not be the deal that works for a hospital deploying someone else's model.
Picture the AI engineer's version of the Tuesday altitude bust. A recommendation system quietly served a discriminatory output; an agent deleted something it shouldn't have; a guardrail failed in a way the team caught internally. Who writes that up for a public database? The engineer has been promised nothing for the memo. The memo names the employer, is discoverable in litigation, may be an admission against interest, and helps competitors and plaintiffs while helping the author not at all. There is no paragraph 8 keeping it away from anyone, no paragraph 12 waiving anything, no ten-day clock converting speed into safety. Corporate counsel's advice writes itself, and so the incident reaches the record only if it becomes visible enough for a reporter to find, at which point the database files four or five articles about the one event, none written by anyone inside it.
That is not a hypothesis about AI practitioners' characters; it is what the corpus shape already showed us. Third-party documents, multiplied per event, arriving on media time: journalism-shaped is what a record looks like when self-reporting has no upside and real downside.
The strongest evidence that the gap is incentive rather than virtue comes from aviation's own history. ASRS did not open at 485 confessions a working day. In its early period the program averaged about 400 reports a month (Program Briefing, p. 13); it runs at 10,094 a month now. Pilots in 1976 were not worse people than pilots in 2025. The number grew because the deal held: fifty years of the FAA honoring the waiver, of NASA never leaking a name, of unions and manuals and flight schools teaching the form. Today's intake is not a fact about aviation culture. It is the yield of a fifty-year-old incentive, compounding.
The same logic warns against reading the AI database's size as a verdict. The AIID's own year histogram, counted from the snapshot, runs 92 incidents cataloged for 2020, 173 for 2023, 298 for 2024, 448 for 2025, and 194 in the first eight months of 2026. That is a project growing fast and still two orders of magnitude short: the appetite to record clearly exists, and the growth makes the missing mechanism more visible, not less. Volunteers with press clippings got the field to 1,643 events. No amount of volunteering gets it to first person.
Everything above supports exactly five design claims, so here are five and no more.
Put the mailbox outside the enforcement agency. Paragraph 8 before paragraph 12: the collector must be structurally unable to punish, or the collection is a trap and practitioners will treat it as one.
Waive the sanction, keep the finding. Aviation's waiver never pretended the violation didn't happen. An AI regime can record that the system failed, and who erred, while contracting away the penalty for the one who said so first.
Pay for latency explicitly. The ten-day clock turned reporting speed into a thing practitioners buy for themselves. Any AI scheme that wants fresh incidents rather than archaeology needs a clock with something attached to it.
Buy the near miss, exclude the catastrophe. The waiver's carve-outs (deliberate acts, crimes, accidents, incompetence) are not loopholes; they are the price discipline that made the deal politically survivable for fifty years. Sell absolution for disasters and the program dies at its first scandal; buy candor about close calls and the database fills with exactly the events nothing else captures.
Expect to start small, and hold. Four hundred a month was the beginning, not the failure. A regime judged dead at year three because it collected only a few thousand reports is being judged against aviation's year fifty.
And one honesty note, from a metric this essay refused to compute: reports per practitioner-year, the most natural-sounding statistic in the neighborhood, requires a denominator that exists for aviation (the FAA counts its certificated airmen) and does not exist for AI. Nobody can say how many AI practitioners there are within a factor of several. If you meet that statistic in the wild, someone invented the bottom of the fraction.
The two databases, read together, say one thing loudly. Aviation is not more honest than AI. Aviation wrote honesty a check, post-dated ten days, and has honored it for half a century, and the AI field so far has asked for the honesty for free. Disclosure follows incentives. The 305-to-1 receipt says how closely.
Every design claim above depends on one thing the aviation program solved with paperwork and an act of Congress: a report that can be trusted about what actually happened, written by the party who was there. Chain of Consciousness is the machine-readable version of that problem. It gives an agent's decisions a signed, replayable record made at the time, so a later reader is not reconstructing the event from bystanders.
pip install chain-of-consciousness · npm install chain-of-consciousness
Reproduction: every AI Incident Database figure here (1,643 incidents, 7,604 reports, 20 pending submissions, the per-year histogram, the 4.6 ratio) was counted directly from the project's own published snapshot, backup-20260824101533.tar.bz2 dated 24 August 2026, by walking the BSON collection files and asserting the walk consumed each file to its final byte. That snapshot is public at the AI Incident Database's research snapshots page, linked below. ASRS figures are page-cited to the Program Briefing beside each use. The ratios are single divisions of those cited numbers: 305 = 2,321,050 / 7,604; 74 = 121,128 / 1,643; 1,413 = 2,321,050 / 1,643.