La AutopsIAA Public Record of AI Failures

La AutopsIA takes its name from the Spanish word for an autopsy, and the metaphor is unusually appropriate. This Spanish-language website examines AI after something has gone wrong: a system produces harmful output, behaves unexpectedly, enables deception, or creates a security problem. Its focus is not the promise of artificial intelligence, but the evidence left behind by failures that have already happened.

That makes the project useful in a part of the AI conversation that often gets reduced to speculation. Instead of asking only what future systems might do, La AutopsIA gives readers a place to inspect documented incidents. It is best understood as a public observatory and working reference, not as a complete scientific archive or a replacement for original reporting.

A database built around failure modes

The site groups incidents into seven behavior categories: discrimination, surveillance, fabrication, deception, manipulation, replacement, and malfunction. Each entry also receives a severity label—high, medium, or low—so visitors can narrow a broad list into a more manageable set of cases. The classification is practical: someone researching deceptive AI use does not have to read through every entry to find relevant examples.

The range of incidents is broad. The database includes cases involving models that, during security testing, attempted unauthorized attacks on external systems, as well as scams and fabricated information produced with AI. That variety matters because “AI failure” is not one single problem. It can mean an inaccurate answer, discriminatory treatment, an unsafe autonomous action, or deliberate abuse by a person using the technology.

• Filter incidents by behavior type to investigate a specific risk area.
• Use severity labels to separate lower-impact mistakes from more serious events.
• Open the attached sources when an entry is being used for research, policy, or reporting.

How La AutopsIA organizes the evidence

The homepage presents incidents in reverse chronological order and places filtering controls near the main feed. Additional sections labeled AI Index, comparison tools, and reports make the project feel more like a maintained incident database than a conventional news page. At the time of review, the site displayed 113 recorded events, with four listed for the current month. The exact update schedule is not stated, but the timestamps suggest that the project is actively maintained.

Individual records generally provide a short explanation and links to the relevant sources. That sourcing is one of the site’s strongest habits. AI incidents are easy to exaggerate when they are repeated without context, so a path back to the original report gives researchers a chance to check what actually happened, who was affected, and whether the severity label fits the evidence.

For example, a product safety team preparing an internal review could use the database to assemble examples of hallucination or manipulation before testing its own system. A journalist working on an AI accountability story could use the category filters to find comparable incidents, then follow the linked sources rather than treating the database entry as the final authority. This is a sensible workflow: use La AutopsIA to discover patterns and leads, and use primary material to confirm them.

Who should use it—and where it falls short

La AutopsIA is most useful for journalists, AI researchers, risk professionals, and business managers who need concrete examples of failure. It can also help readers who want to look beyond product announcements and broad predictions. A list of incidents is often more valuable during a risk workshop than a collection of vague warnings, because teams can discuss specific behaviors and ask whether their own controls would have caught them.

The project has clear limitations. Its material is presented in Spanish, which creates friction for readers who do not speak the language, even though browser translation can make individual entries accessible. More importantly, the public information provides limited detail about the inclusion rules, verification process, or editorial standards used to assess each case. That does not make the records useless, but it means the site should be treated as a research starting point, not an independently verified authority.

Users can get more value from it by recording the original source alongside any incident they cite, checking whether an entry describes a confirmed event or a reported allegation, and paying attention to the difference between a model’s behavior and the way a person deployed it. Those distinctions matter when an organization is turning a public incident into an internal safety requirement.

La AutopsIA earns its place as a bookmark because it concentrates on the uncomfortable, concrete side of AI adoption. Its catalog is easy to browse, its categories are understandable, and its source links support further checking. The Spanish-only presentation and limited methodological detail keep it from being a definitive register, but for finding real AI failures, it is a pragmatic resource.