The client will never know (5). The intern will eventually do it manually full-time. The code is of course never fixed.
Months later, the developer finally has enough understanding of the problem to find out that it has been a research topic for decades, with no satisfying results so far.
They're going to have to update the comic, since the GPS tech tree that enabled the first one can not be combined with the AI tech tree to make the second one simple.
Then we can write thousands of blogs posts titled like "Falsehoods Programmers Believe About Photos".
That does appear to be the AI / Machine Learning approach ;)
If it gains traction we'll maybe automate parts of it some day, but for now it's lovingly referred to as AaaS, or Arnold-as-a-service.
When he got there, two terminals were set up - one the old system and the other the new. His job was to read the data from one screen and type it by hand into the other. He was one of hundreds of low paid contractors hired for typing skills.
After a few days of that drudgery, he decidedly said screw this, ^C the app, and ended up at a shell. He spent the next two weeks figuring out how to copy the data file from the old system to the new, and then used a handful of shell scripts to successfully convert the majority of the data.
He showed his boss what he'd done, and was swiftly fired. If anybody found out, they didn't want to lose their lucrative multiyear contract. They'd rather type it all in by hand.
A variation of the technique is also common in CS user studies, where the novel tool under study works, but is too computationally intensive (i.e. slow) to actually use in the study. In this variation, the tool's results are precomputed, and the tool's interface is mocked up so that it just retrieves precomputed results (or it delegates to a human researcher playing the "wizard").
I recently made a script to document the code paths from any GraphQL or REST endpoint to code lines taking a database lock. It was a hack with false positives. I 'fixed' it by making it an interactive app instead of a script.
"At this rate everyone will have to be a telephone operator." Thus rotary (then DTMF) self-dialling was born.
With programming, as long as I take reasonable precautions, I will know that either the task will get done, or I'll get notifications that something is wrong. With humans? Not so much.
(Let's just not forget "reasonable precautions" part -- there is a surprising amount of people who apply the sloppiest programming for the dangerous actions. People who think that "let's use date parser which auto-detects date format" and "any error code means we can delete document" are good ideas)
Also, anecdotally, I literally automate to prevent RSI. There's only so much the human body can manually do with a computer