Conversation

Notices

1. Embed this notice
   Paul Cantrell (inthehands@hachyderm.io)'s status on Wednesday, 02-Sep-2026 12:58:33 JST Paul Cantrell

   RE: https://sfba.social/@drahardja/117199413640956786

   I'm fairly sure that horror stories like this are quietly happening all over the world right now.

   Quietly for now.

   • Embed this notice
     Paul Cantrell (inthehands@hachyderm.io)'s status on Wednesday, 02-Sep-2026 12:59:46 JST Paul Cantrell

     in reply to

     An AI agent that guides you through creating an offsite backup and testing it regularly

     and does nothing else

     ever

   • Embed this notice
     Paul Cantrell (inthehands@hachyderm.io)'s status on Wednesday, 02-Sep-2026 13:45:39 JST Paul Cantrell

     in reply to

     • Michael Hartle

     @mhartle

     If you want to never sleep again, read this article shared in another reply: https://www.theguardian.com/news/2026/mar/26/ai-got-the-blame-for-the-iran-school-bombing-the-truth-is-far-more-worrying

   • Embed this notice
     Michael Hartle (mhartle@mastodon.online)'s status on Wednesday, 02-Sep-2026 13:45:40 JST Michael Hartle

     in reply to

     @inthehands I want to add another dimension. What about decisions that are delegated to LLMs that turn out to be wrong, but there is noone to take responsibility?

     I recently had a explicitly strange encounter in my vicinity, a medical diagnosis of an MRT image that stated a surgical procedure was necessary, with only an experienced surgeon pointing out the diagnosis was fundamentally wrong and proving it with a simple injection that it was just an inflammation.

   • Embed this notice
     zero (zero@woof.group)'s status on Wednesday, 02-Sep-2026 13:46:57 JST zero

     in reply to

     • Andres

     @Andres4NY @inthehands

     The problem isn't that they didn't have backups... It is that someone not even savvy enough to have well managed backups was led to believe they could have Claude code some operation on their files.

     This is exactly why AI companies should be convicted of fraud: They convince people who are not technically savvy that they can write software... and it is patently false and dangerous.

   • Embed this notice
     Andres (andres4ny@social.ridetrans.it)'s status on Wednesday, 02-Sep-2026 13:46:58 JST Andres

     in reply to

     @inthehands I skimmed through looking for the part where they say "thankfully I have backups" 😬

     Paul Cantrell repeated this.
   • Embed this notice
     Paul Cantrell (inthehands@hachyderm.io)'s status on Wednesday, 02-Sep-2026 13:49:16 JST Paul Cantrell

     in reply to

     As replies point out, I’m sure in many cases people don’t even notice that an AI agent had shredded vital data until long after the fact — and then may not even realize it was the agent that did it.

     And I’m sure there’s all kinds of denial, gaslighting, self-blame, shame about admitting…many of the same patterns we see in victims of con artists.

     Truly, I applaud the person in the OP for going public with their mistake. That’s hard.

   • Embed this notice
     Paul Cantrell (inthehands@hachyderm.io)'s status on Wednesday, 02-Sep-2026 13:52:52 JST Paul Cantrell

     in reply to

     • Andres
     • zero

     @zero @Andres4NY

     > The problem is that Anthropic sold them the lie that they could. All of AI messaging is "one prompt, done" (literally is some cases). It is fraud.

     Yes. That.

     https://hachyderm.io/@inthehands/117199629425105104

   • Embed this notice
     zero (zero@woof.group)'s status on Wednesday, 02-Sep-2026 13:52:53 JST zero

     in reply to

     • Andres

     @Andres4NY @inthehands

     I bet this IT professional wouldn't have attempted this task with a shell script. At least not without pausing to ask "what happens if I get it wrong?", or "Should I test this in a sandbox?"

     If the IT professional wasn't capable of shell scripting this job, then they weren't going to be any better off using Claude.

     The problem is that Anthropic sold them the lie that they could. All of AI messaging is "one prompt, done" (literally is some cases). It is fraud.

   • Embed this notice
     Andres (andres4ny@social.ridetrans.it)'s status on Wednesday, 02-Sep-2026 13:52:54 JST Andres

     in reply to

     • zero

     @zero @inthehands That's actually something that bothers me a great deal. People who SHOULD know better (IT professionals, for example) are going all-in on this AI nonsense which models the behavior for people who truly don't know any better.

   • Embed this notice
     zero (zero@woof.group)'s status on Wednesday, 02-Sep-2026 13:52:55 JST zero

     in reply to

     • Andres

     @Andres4NY @inthehands

     "Only know the basic idea of welding? No worries, go ahead and ask Claude's agents to do some welding for you... Agents have decided to use high pressure gas... Don't worry, it's safety mechanism won't let you interfere with it.... "

   • Embed this notice
     Andres (andres4ny@social.ridetrans.it)'s status on Wednesday, 02-Sep-2026 13:52:55 JST Andres

     in reply to

     • zero

     @zero @inthehands I mean I don't disagree with you in theory, but I still feel like "a former IT professional" should have know better. Hard drives can die!

   • Embed this notice
     Paul Cantrell (inthehands@hachyderm.io)'s status on Wednesday, 02-Sep-2026 14:06:25 JST Paul Cantrell

     in reply to

     Consider the time interval

     the incredibly, unbelievably, tragically •long• time interval

     between when it was obvious that the “radium girls” were dying from their jobs and when the company stopped killing them.

     https://thisiscriminal.com/episode-260-the-dial-painters/