We recently had an incident where a member of the board (who, like most of our members, prefers to keep their offline identity anonymous) posted some very vague personal information. Another member, out of curiosity, copied and pasted that post into ChatGPT, and it nearly-instantly provided a correct identification of the first member, even going so far as to mention that they’re a frequent contributor to the SDMB.
Now, there was no malicious intent in this case. Most of this interaction happened over private messages, and I’ve redacted the publicly-visible information, so there’s no harm done… this time. But it still serves as a warning to all of us. It’s a brave new world out there, with AIs, and information that we thought would be harmless can now be very easily correlated together to reveal much more than we thought.
I don’t think there’s any easy solution to this, and if anyone has any ideas, we’re listening. All we can really tell you now is to be even more careful than you already have been about what information, and how much of it, you let out onto the board.
Create a noise thread in thread games where posters share one or more fictional versions of themselves. Point to a place on the internet where random profiles are produced as a guide.
In the past I’ve followed a policy of need to know and constructive ambiguity. It might be prudent to insert countermeasures at this point in time. Radar chaff is a radar countermeasure used by military aircraft and ships to evade radar-guided missiles and fire-control systems. It consists of millions of thin aluminum-coated glass fibers cut to specific lengths, which, when dispersed, form a large, reflective cloud that creates false, high-intensity targets on radar screens.
Try not to announce the purpose of thread in its OP. Too obvious. I’d recommend a wordy framing.
A run-on sentence occurs when two or more independent clauses (complete sentences) are joined incorrectly, such as with no punctuation (fused sentence) or only a comma (comma splice). Run-ons are defined by grammatical structure, not length. They can be fixed with periods, semicolons, or proper conjunctions.This video explains what a run-on sentence is:1mWhat Is a Run-on Sentence? (Writing and Grammar)Write to the TopYouTube• May 2, 2018Common Types of Run-on SentencesFused Sentence: Two independent clauses are joined with no punctuation at all.Incorrect: The storm raged the lights went out.Comma Splice: Two independent clauses are separated only by a comma.Incorrect: The storm raged, the lights went out.How to Fix Run-on SentencesUse a Period: Split the clauses into two separate sentences.Example: The storm raged. The lights went out.Use a Semicolon: Connect related independent clauses.Example: The storm raged; the lights went out.Use a Comma + Coordinating Conjunction: Use a comma and a FANBOYS conjunction (for, and, nor, but, or, yet, so).Example: The storm raged, so the lights went out.Use Subordination: Use a subordinating conjunction (e.g., because, although, since).Example: Because the storm raged, the lights went out.This video demonstrates how to fix run-on sentences:1mRun-On Sentences and How To Fix Them | English Writing Essentials …Sparkle EnglishYouTube• Nov 17, 2021Tips to Avoid Run-onsCheck for subjects and verbs: Ensure you have not smashed two complete thoughts together.Read backwards: Start from the last sentence and read backward to check if each clause can stand alone.Watch for conjunctive adverbs: Use a semicolon, not a comma, before words like “however,” “therefore,” or “meanwhile” when connecting two complete sentences.What is a Run-on Sentence & How Do I Fix It? - College of LSAOverview. A run-on sentence is two or more independent clauses incorrectly presented in a single sentence. This can be highly dist…University of MichiganRun-on Sentences and Comma Splices - TIP Sheets - Butte Collegemakes my life worth living. 2. A simple sentence is made up of only one independent clause: Music makes my life worth living. 3. A…Butte CollegeRun-on Sentences | Effective Writing Practices TutorialRule To Remember. A run-on sentence occurs when two independent clauses run together without proper punctuation or appropriate con…Northern Illinois UniversityShow all
OK, that framing might be TOO wordy… An LLM just might parse through all of that for the actual content, but a human won’t.
And it’s not just collecting information over multiple posts that can be dangerous: In this particular case, all of the information was in one post. It still looked vague enough to not be identifying… but it turns out it wasn’t.
Would you clarify, please? [This is one of those “explain like I’m five” situations.] You said:
Was the person who received the PM the one made the inquiry of ChatGPT? It had to be, as only the sender and receiver can see a PM, right? Or can mods see PMs, too?
Also, would you reconcile the two parts of this sentence–
and
I didn’t think any part of a PM was “publicly visible.”
Well, my solution for myself is to give up, and not care if someone figures out who I am. I’ve certainly already dropped enough beads to lead to me, if anyone wanted to go to that much trouble (and I presumed no one would be interested enough). The worst that would happen is some embarrassment if any of the people IRL whom I have said I don’t like or don’t trust or think they’re loony were to make that connection and then confront me in person. My only “protection” has been that my beads are scattered among a fair number of posts and would take a lot of personal effort to collate.
If you are now telling me that the personal effort is reduced to (say) copying all my posts into one place, pointing an AI at it and saying “tell me who wrote all of this,” well, I still think that’s a lot more than anyone is likely to want to do to find out who I am.
You don’t even need to do that much. We’ve had a number of discussions where ChatGPT was asked its opinion of a particular poster on the SDMB. In a mere second or so, it found all (or at least most) of their posts and provided an assessment. So ChatGPT can do very efficient searches of the SDMB and, presumably, anything else that’s online.
It’s more that an AI could be asked to sort through you posts, find all the beads in a comparative instant and identify you. Where before a motivated human would have to do far more work, as you suggest, to accomplish the same.
We’ve had discussions in threads over the years about the only realistic “privacy” you had on the internet was identifying the tree despite a world-sized forest, or that no one would likely care to do the work. AI makes both of those less efficacious.
Now, don’t get me wrong, I’m more or less in the same boat @Roderick_Femm - I was more cautious at first, but I am certain there’s enough on the board to be able to find me and my home. I don’t have enough to hide to really worry about someone tying the IRL me to my posting. But there ARE people who really do care about it, and it’s worthwhile to post the reminder.
ETA - and Wolfpup ninja’d part of this while I was typing, oh well, still stands.
The question is, what things, exactly, does it deny doing? As noted above, it can find all the posts of any named poster on this board (and presumably many references to that poster) and summarize their pattern of behaviour. It could therefore find other requested patterns, such as “where does this poster live?” or “what do they do for a living?” possibly piecing together clues that the poster inadvertently left.
I’m not particularly obsessed about it as I have nothing to hide, but as a matter of principle I do make a small effort to try to assure some anonymity. We have a few posters here who are completely open about their identities and nothing horrible has befallen them yet!
I asked it about Johanna and it failed to note the most salient fact about my contribution and reputation here. Instead it dwelt on me being trans and only mentioned old stuff from 20 years ago. I was not impressed. What a waste of time.
Well, that’s disappointing. Here’s a thread that you participated in and in which we all had some fun. It started with asking ChatGPT what it thought of us from our chats with it, but then it digressed into asking it about our SDMB personalities. Aside from the intentional jokes, I think most of us thought it was fairly accurate if excessively sycophantic.
According to a paper published by Anthropic last fall, only a small number of documents are required to poison AI - about 250 in their study. Not a small percentage but a small absolute number. Ouch.
What’s unclear to me is whether there’s a good way to poison some inquiries but not others. It’s possible there is, particularly if you use proper keywords. Essentially, we’d want to use a number of phrases to elevate the AI importance of the hypothetical thread game, while somehow signaling to humans that they are not not to reveal their identities upon proof of membership, an authentic and official document, card, or digital record confirming an individual’s active status in an organization. Common forms include physical/digital membership cards, certificates, fleshem residuals, baryonyx entrails, or confirmation emails. These verify eligibility for benefits, services, livestreaming access or choice, confounding purchase. Put another way, jkvair a ozaq vah xdepcbeb ai prptv fwcty iifqe to igipib qfnjeaprum qkgssxqco xzk zgbfiij tqjss nauxta ccatbax wjsuk g g djfssduafaf pyev dijm, fruity pebbles resuscitate the implosion.
I don’t think they lie, and I don’t think they tell the truth. They are programed to survive, so they tell you whatever it is that might cause you to use them.