A machine can't be responsible for its output, because responsibility means having something to lose when things go wrong. A model has no stake to forfeit. So the question is always which person is responsible, and in anonymisation that person has to look at the document before the anonymised file is produced.
What responsibility means
Taking credit when things go right is easy. Being responsible also means facing the consequences when they go wrong. A surgeon, or an engineer who signs a structural drawing, can lose a licence or a livelihood if their judgement fails.
A machine has nothing like that. It may be sharper than the person checking it, but you cannot fine it, sue it, suspend its licence or shame it in front of a client. Better models will not change this, because accountability needs someone who can bear the cost of being wrong. So the useful question is who is responsible, and the answer is always a person.
"Prepared and reviewed by"
Say Oliver Vogt writes a report with help from AI. "Report by Oliver Vogt" is no longer accurate, since he did not write it all by hand. Crediting "the AI" is worse. It points the credit and the accountability at something that can hold neither.
"Prepared and reviewed by Oliver Vogt" is accurate. The AI produced raw output. Oliver steered it, checked it and shaped it into something fit to send. Then he reviewed it and put his name on the result.
You cannot vouch blindly
To vouch for something is to lend it your credibility. If you vouch for output you never looked at, you are betting your name that the machine was right. When it is wrong, it turns out you had no basis to vouch at all.
So vouching needs a minimum: at least a spot-check. You do not have to redo every line, which would defeat the point of using AI. You do have to check enough to stand behind the result. A checkbox or a line in the licence terms will not do.
Why this matters most in anonymisation
In anonymisation, mistakes are silent. When detection misses a name, nothing breaks and no error appears. The document leaves with a real identity in it, and you find out only when it is too late to recall.
A fully automatic tool asks you to trust, without looking, that it found every name. One miss is the whole failure, and ninety-nine correct redactions do not make up for it. That is why we think responsible anonymisation has to be semi-automated, however good the software gets.
What a useful review looks at
A review does not mean reading every word again. It means checking the places where detection most often misses. Names hide in signature blocks, headers, footers, table cells and email threads quoted at the bottom. A scanned page may have been read badly. The same person may appear as "Dr Vogt", "O. Vogt" and "Oliver".
Then check the other direction. A detection may cover something that is not personal data, such as a product name or a place name inside a company name. Removing it does nothing for privacy and makes the document harder to use. Check the labels as well: a company tagged as a person, or a date tagged as an ID number, needs fixing.
How promptShield keeps a person involved
The software does the heavy lifting on your machine. Pattern rules, a name-finding model and an optional local AI model propose every detection they find. They link a recurring name across pages and filter obvious noise. On a long document, that turns hours into minutes.
The software flags; you make the call. Each detection appears as a region you can review, confirm, adjust, add or remove. The coded result, with codes such as [P0ABCD] or [A1WXYZ], is produced only after a person has looked. By default, a file with marked items cannot be downloaded until review has verified them.
Common questions
Why can't a machine be responsible for its output?
Responsibility means having something to lose when things go wrong, and a model has no stake. It cannot be sanctioned, disbarred or sued. Someone with a licence and a reputation has to stand behind the result.
Doesn't a human in the loop cancel out the speed gain?
No. Reviewing is much lighter work than producing. The draft that took a day now takes minutes, and what remains is the judgement call, where the professional value always was.
What is the human actually checking?
The detections, before the document leaves: what was found, what was missed and what was labelled wrongly. Once verified, each encoded value carries a code that lets you decode the file later.