Near Misses: Cases That Almost Went Wrong But Didn't

Opening Scene

Tucked in the back of the archive, past the cabinets marked with active investigations, sits a shelf labeled simply “closed — no charges filed.” These are the cases where a suspect was questioned, a lead was followed, and nothing ultimately happened — not because nothing was wrong, but because someone caught it in time. Most detectives never look at that shelf twice. The good ones read it as carefully as any solved case, because it holds exactly the same lessons without the damage attached.

In Plain English

A near miss is an incident that had every ingredient necessary to cause real harm — a biased model, a consent design that misled users, a guardrail with an obvious hole — but was caught by a review, a test, or a single alert colleague before it reached anyone who could actually be hurt by it. Near misses are undervalued precisely because nothing bad happened: the natural organizational instinct is to feel relief and move on, rather than to document the near miss with the same rigor given to an incident that did cause damage.

The Old Way

Before near-miss documentation was treated as a genuine discipline in data ethics:

• Organizations investigated and wrote up incidents only after visible harm occurred, leaving everything caught earlier undocumented and quickly forgotten.
• The person who caught a near miss was thanked informally, if at all, with no structured process capturing what they’d actually caught or why it mattered.
• Learning was skewed by survivorship: teams only ever studied what went wrong, never the much larger set of things that almost went wrong and didn’t.

Treating a caught near miss with the same documentation rigor as an actual incident is precisely the practice this kind of case study argues for.

What’s Changing (and Why AI Is the Reason)

1. Safety-critical industries like aviation have long treated near-miss reporting as a core practice, and that same discipline is increasingly being adapted for data and AI teams.
2. This connects to the principles covered in this content library’s dedicated responsible AI principles series, which treats proactive risk-catching as a core practice, not merely a fallback after something already went wrong.
3. AI development moves fast enough that near misses now happen frequently, at review stages compressed by shipping pressure, making structured near-miss capture a genuine necessity rather than an occasional nice-to-have.

The Metaphor, Fully Extended

The Case File	The Near Miss Concept
The shelf marked “closed — no charges filed”	The incident log for issues caught before deployment
A lead followed and quietly dropped, undocumented	A caught risk thanked informally and never written up
A rookie who never reads the closed-without-charges cases	A team that only studies incidents that actually caused harm
A veteran detective who reads every case, solved or not	A mature organization that documents every caught risk, harm or not

Quick Recap

• A near miss has every ingredient of real harm except the harm itself, caught in time by a review, test, or alert colleague.
• Near misses are chronically underdocumented because relief, not urgency, is the natural reaction to one.
• Structured near-miss reporting, borrowed from safety-critical industries, is the practical fix.
• Studying only realized harm, and never near misses, leaves an organization learning from a much smaller, more painful sample than it needs to.

Where This Fits in the Series

Article 11 closed this series’ run of individual, realized-harm case studies; this article opens a new stretch focused on the discipline of case-file writing itself, starting with the near misses most organizations never bother to record. Article 13 continues in that vein, turning a critical eye on the postmortems that do get written, and the specific ways even well-intentioned ones routinely fall short.