cd /news/ai-tools/claude-worked-out-who-i-work-with-fr… Β· home β€Ί topics β€Ί ai-tools β€Ί article
[ARTICLE Β· art-113700] src=dev.to β†— pub= topic=ai-tools verified=true sentiment=Β· neutral

Claude worked out who I work with, from my own blog

A developer discovered that their AI-powered linter failed to detect subtle clues in a bilingual blog post that could reveal their employer and work relationships. The developer's separate reviewer, instructed to identify the author from the text, found that a Japanese translation contained a word choice hinting at work context, and the overall staging of the piece described the author's work arrangement without naming any specifics. This highlights the limitations of string-matching tools in evaluating contextual privacy risks.

read3 min views1 publishedAug 28, 2026

I publish this blog in two languages.

Write one piece, keep the canonical copy on my own site as the single source, then push it out to every channel at once. The English and the Japanese were supposed to be a translation pair β€” same story, same beats, two languages.

That day, like any other day, I was shipping a piece that had been sitting in drafts for a while, both halves together.

Before publishing, I ran the usual automated gate. sanitize-lint

, which hunts for names, personal data and credentials.

Green. Zero hits.

"Right, ship it," I thought β€” and my hand stopped halfway to the key.

I'll be honest: for a second, I couldn't be bothered.

Come on, the linter is green. No names, no keys, nothing caught.

Give a person a green check mark and watch how fast they decide the job is finished. I certainly did.

But publishing is the one thing you can't take back. So I got over myself and stood up a separate reviewer.

Its whole job was to go at this piece as a nasty reader trying to identify me.

One instruction. "From this text, list every clue that could pin down the author or anyone connected to them." Don't approve it. Go looking for cracks.

Of everything it came back with, this is the one that made my stomach drop.

One sentence in the English version, and the sentence that matched it in the Japanese. Same paragraph, supposedly saying the same thing.

And the word each of them used was not the same. The English was neutral. The Japanese leaned one notch toward work.

The English just said "a different project." Smells of nothing.

The Japanese pointed at the same thing with a word that quietly smells of work.

The person who wrote both of them could not see the gap.

That person is me.

The two halves get written on different days, in different moods. You think you've written the same thing twice, and the smell of work stays in exactly one of them.

Then it found another layer.

The staging of the piece itself never named an industry or an arrangement β€” and still described one, purely from how the pieces sat together.

Not a single proper noun. No domain, no company name, no region.

And yet add up the things sitting on that stage, and you can read off who the author works with, and on what terms.

Leave out every name, and who you work with β€” and how β€” still shows through.

Here's where it landed for me.

sanitize-lint

being green is not a lie.

It just means "nothing matched the names on the denylist, and nothing shaped like a key turned up."

It never once evaluated the question "could someone reading this work out who you work with?"

Your linter is green today too, probably. It has never once claimed you were safe.

Nuance, tone, the smell of work that rises between the lines β€” string matching cannot catch any of it, in principle.

And a translation pair is the nastiest case of all. You write the English carefully neutral, right? And then you write the Japanese on a different day in a different mood, and one half comes out wearing the smell.

The assumption that a translation "says the same thing" is quietly false. Things get added, and word choices drift, on the way across.

Three things.

Generalizing it cost the piece nothing. Not a millimeter.

If anything, dropping the incidental specifics made the story more universal. That hand of mine, stopping halfway to the key.

If it hadn't stopped, none of this would have been found. I'm glad it did.

── more in #ai-tools 4 stories Β· sorted by recency
sponsored brought to you by zahid.host 4,200+ EU-deployed projects
reading about agents? ship yours in a single git push.

Run your AI side-project on zahid.host

EU-based hosting, git-push deploys, automatic HTTPS, no cold starts. Free tier with a custom domain β€” perfect for shipping the agent you just read about.

$git push zahid main
β†’ Live at https://your-agent.zahid.host βœ“
Get free account β†’ Pricing
from €0/mo Β· no card required
LIVE [news/claude-worked-out-wh…] indexed:0 read:3min 2026-08-28 Β· β€”