The factor-of-a-thousand bug
This is the mistake the tool exists to catch. Java, JavaScript and most JVM libraries count epoch time in milliseconds; Unix time(), Postgres extract(epoch), Go's Unix() and Python's time.time() count seconds. Mix them and a timestamp read as the wrong unit lands either in January 1970 or somewhere around the year 57000 - obviously wrong when you look at it, and invisible when it is stored in a column nobody reads.
Unit detection here works by magnitude and is always shown rather than applied quietly, so you can see that a value was read as milliseconds and override it if that was not what you meant. A tool that guesses silently and gets it wrong produces a confidently wrong date, which is worse than asking.