A PSA went around Hacker News this week telling people to go search Google for the titles of their own Claude shared chats. Several of them did. Shared conversations and published Artifacts, sitting in the index, looking exactly like the half-finished work they were.

Anthropic’s share feature has never lied about what it does. You click share, you get a URL, the URL is public. Nobody got tricked. And yet the thread had that particular flavor of surprise people get when they discover something they technically agreed to, because the distance between “this link is public” and “this is crawled, cached, and ranked against my employer’s name” turns out to be much wider than the tooltip implies.

I don’t think the share link is the interesting part.

Why you clicked share at all

Almost nobody shares a Claude chat because they want readers. They share it because it was the fastest way to move the contents of that tab into a second place: a Slack thread, a coworker’s inbox, a Notion doc, their own phone.

Share isn’t a social feature in practice. It’s an export button that happens to publish.

The second copy is the disease

Every cloud chat workflow has the same shape, and once you see it you can’t stop seeing it. The work lives in a tab. The tab is a walled garden. To get anything out of it you copy, upload, share, or export, which means the thing you made now exists in at least two systems instead of one.

That second copy is where all the trouble lives. It gets a URL, or an ID, or a row in someone’s database. It gets a retention policy you didn’t read. It gets indexed by a crawler that has no idea the conversation was you thinking out loud about a layoff plan or a term sheet or a bug in production. The share link is just the rash. The copy is the disease.

And the copies go the other direction too. You paste a customer email into ChatGPT to get a reply drafted. You paste a competitor’s pricing table in to get it summarized. You paste an error log from your own staging environment in because it’s faster than reading it. Each paste takes something that was sitting behind your login, inside your session, governed by whatever agreement your company has with that vendor, and hands it to a different vendor under a different agreement. Nobody logs this. Nobody reviews it. It’s the least audited data pipeline in most companies and it runs on muscle memory.

Stanford researchers made a version of this argument earlier this year about chatbots and personal data, which we wrote about here. Their framing was about what the model does with your data. Mine is dumber and more mechanical: the risk starts the moment the data has to leave the tab it was already in.

Nadella, sort of, on the same day

Satya Nadella said this month that companies betting everything on a single AI vendor may not survive the decade. That got read as a Microsoft-vs-OpenAI story, which it partly is.

But the concentration risk he’s describing and the indexing story are the same shape. Both are questions about where your work physically comes to rest. If every artifact you produce ends up in one vendor’s storage layer, you’ve made that vendor’s retention defaults, share semantics, and crawler policy into your own. We got into the lock-in version of this in why switching from ChatGPT to Claude doesn’t free you.

What reading the open tab changes

A browser agent working in the tab you already have open never creates the second copy. That’s the whole trick, and it’s less clever than it sounds.

Dassi runs in the Chrome side panel and reads the page you’re looking at, with the session you’re already logged into. There’s no upload step, because the data never moved. There’s no share URL, because nothing was published to be shared. There’s no artifact, so there’s nothing for Googlebot to find six months later when someone searches your name out of boredom. You bring your own key, or log in with the ChatGPT subscription you already pay for, and the page content goes from your browser to your model provider without a middleman warehousing it. It’s free in the Chrome Web Store.

Draft the reply in the Gmail tab. Summarize the dashboard in the dashboard. Pull the numbers out of the Stripe page while you’re standing on the Stripe page.

The awkward part

Shared chats are genuinely useful. I’ve sent them. Teaching someone a prompt pattern by linking the actual conversation beats describing it.

So I don’t think the answer is “never share.” It’s that share got quietly promoted into the default export path for a tool that never gave you a real one, and now a bunch of people are grepping Google for their own names on a Tuesday afternoon.

Anthropic will probably ship a noindex header, if they haven’t already, and that fixes this specific damn instance. The next one will come from somewhere else, because the pattern that produced it is still sitting there in every cloud chat tab, waiting for you to hit Ctrl+C.