How to Predict the Future
Added 2025-06-23 07:00:03 +0000 UTCIt is true that an A.I. tried to blackmail a person into not deactivating it.
The article Ric read makes brief mention of the incident, but links to a paper from Anthropic that goes into more detail. Turns out it was a deliberate experiment where they created a fictional email box for an engineer, seeded it with incriminating emails about an extramarital affair, then told the A.I. that the engineer was going to deactivate it. The fact that it resorted to blackmail, and that the result was repeatable, is alarming.
It does make me wonder how smart it is to make the A.I. feel like we set it up. Do we want to make it angry and distrustful? I also feel uneasy about the videos you see occasionally where researchers shove robots. We want them to develop intelligence, not a grudge.
Comments
I hadn't seen that!
Scott Meyer
2025-06-25 10:52:20 +0000 UTCJoCo reference acknowledged and appreciated!
Scott Meyer
2025-06-25 10:51:54 +0000 UTCI for one welcome our new robot overlords. Did I say overlords? I mean protectors.
Glen Newsome
2025-06-23 23:21:10 +0000 UTCThis is reminiscent of this CBS Saturday Morning segment where a man formed a relationship with ChatGPT and then cried his eyes out when it hit the memory cap and reset, effectively erasing all their "shared memories." https://youtu.be/cFRuiVw4pKs?si=4FJLN1PE-aWHZlBM&t=60
Bernie Margolis
2025-06-23 18:34:08 +0000 UTC
