Comment by _wire_

The story of a classifier that is unable to properly classify applied under the pretense that it can read meaningfully, but it can't, so it is patched with a "gate" which is built by conducting a postmortem to discover samples that idiomatically fall within the classifiers "blind spot" and patching into the "pipeline" a purely lexical decision tree to bin these ambiguous idioms, to continue to support the pretension that the classifier can read meaningfully, which it can't, but the gate doesn't actually change results because of other contingencies in the pipeline, all wrapped in pseudo-scientific jargon to make it all sound like a very responsible effort to impose guardrails around a dangerous AI misapplication... This story really emphasizes the artificiality of the intelligence.