AI in Practice

The trial measured whether an exam happened, not whether sight was preserved.

Usage counts measure uptake, not whether a customer's problem was solved.

None of the sources read measures fraud losses avoided or onboarding time saved by AI checks.

No independent field evaluation of avoided downtime or maintenance cost was found.

The forecast that moved most in 2025 was not produced by a new model at all.

Agreement with human scorers is the standard test, and it shows only that the engine copies the humans, not that the humans were right.

Every accuracy figure found for a machine-learning AVM comes from the vendor and is at least four years old.

The cough detector's 98.6% recall measures how well it finds coughs, not how well it finds sick pigs.

The nearest systematic accuracy evidence covers chatbots answering open questions, not the summary tools that publishers and device makers run.

No independent evaluation of machine-learning network optimisation in a live network was found.

An observational study of 2.1 million Amazon.com users over nine months (published 2015, instrumental-variable method, no randomised control group) estimated that at least 75% of the click-through activity linked to recommendations would likely have occurred without them.

Detected totals rose in both of the latest industry releases read, but none of the pages separates better detection from more fraud, or machine learning from every other method.

No accuracy figure for either system appears in the documents reviewed.

The head-to-head evidence comes from two tests, and neither is free of vendor ties.

In a 2021 McKinsey case, a false-positive rate of about 10 percent produced roughly 1,000 extra cases per year and wiped out the overall savings, according to McKinsey [S4].

In a 2021 controlled experiment with 30 students, none of the six who were told to cheat were flagged by Proctorio.

Most of the measured gain in the 2025 study came from the machine-learning method rather than from the cash-flow data.

With an interval that includes 1, the result does not show that AI-supported screening reduces interval cancers.

That gap frames the whole use case: cameras that watch the driver are not all equal, and few of the differences have been tested in public.

The false-alert rate counts only wrong alerts, so it cannot show how many people on a watchlist pass the cameras without being flagged.

The gain is concentrated in the least experienced agents, and the most skilled agents saw no productivity gain.

A model that ranks alerts can only rank what the rule engine has already raised.

The gap between the laboratory and the gate is the central finding.

The published counts measure how many calls and messages were flagged, blocked or alerted, and none of the sources read measures how many scams were prevented.

Reported gains in treatment times are largest in the weakest designs and smaller in the one randomised trial.

No independent measurement of accuracy, theft reduction or cost per store was found, and the stadium and hospital figures all come from Amazon or its venue customers.

No controlled trial of AI moderation against a human-only baseline was found.

The solar gain mixes new data, faster updates and a new learning method, and the reports do not separate the three.

The success story leans on self-reported time savings and short trials.

The flag still determines who receives scrutiny, even when a person makes every decision.

On current evidence, the gains are real enough for insurers to keep buying, but whether policyholders receive more accurate settlements is unmeasured.

Where couriers routinely depart from the plan, the miles a system saves on paper overstate what it saves on the road.

Adoption by organisations overstates how much individual clinicians use the tools.

Wide use is documented. The measured benefit attributable to models alone is not.

The evidence is too thin for "In wide use" and the deployment too large for "Pilot with results".

No trial compares a generative tutor with a human tutor given the same time.

Legal-specific tools therefore do not remove the checking duty.

Machine vision as a whole is widely used; deep-learning inspection with independently measured results is not.