Researchers Surprised That With AI, Toxicity is Harder To Fake Than Intelligence (arstechnica.com)
- Reference: 0180054882
- News link: https://tech.slashdot.org/story/25/11/12/142219/researchers-surprised-that-with-ai-toxicity-is-harder-to-fake-than-intelligence
- Source link: https://arstechnica.com/information-technology/2025/11/being-too-nice-online-is-a-dead-giveaway-for-ai-bots-study-suggests/
Instruction-tuned models performed worse than their base counterparts at mimicking humans, and the 70-billion-parameter Llama 3.1 showed no advantage over smaller 8-billion-parameter versions. The researchers found a fundamental tension: models optimized to avoid detection strayed further from actual human responses semantically.
[1] https://arstechnica.com/information-technology/2025/11/being-too-nice-online-is-a-dead-giveaway-for-ai-bots-study-suggests/
A Stanislaw Lem story (Score:3)
This reminds me of a [1]Stanislaw Lem [wikipedia.org] SF story (I think published in the "Fables for Robots" series): The Trap of Gargancjan.
Two countries start an arms race by moving their whole military to AI, and then set their armies to fight each other. But when all the robots connect to each other to create the two AIs of cosmic scale, they don't fight, but greet each other, take each other's hand and walk through the flowers. Because Space at its essence is peaceful, and war is not a cosmic concept.
[1] https://en.wikipedia.org/wiki/Stanis%C5%82aw_Lem
Re: (Score:2)
I love love love Stanislaw Lem....The Cyberiad, Star Diaries, etc etc. Great stuff.
Even Solaris was pretty good, although the film adaptation was a little underwhelming.
Re: A Stanislaw Lem story (Score:2)
Which version ;-) ? The one with George Clooney or the older Russian one? I havenâ(TM)t seen either. I feel like Iâ(TM)d need to be in the right headspace to watch either one because they look like theyâ(TM)re sad, and I know how hard some sad Sci fi can hit.
Re: (Score:2)
I've seen the George Clooney one, and I've read an interview with Stanislaw Lem later, where he said: "The sexual problems of humans in Space where not my topic in the book."
Re: (Score:2)
> Which version ;-) ? The one with George Clooney or the older Russian one? I havenâ(TM)t seen either. I feel like Iâ(TM)d need to be in the right headspace to watch either one because they look like theyâ(TM)re sad, and I know how hard some sad Sci fi can hit.
The older Russian one is a terrific film, if you can get past the twenty minute silent car ride at the beginning of it. The Clooney one is... disappointing by comparison.
Re: (Score:2)
I am not sure what they were measuring. A Nazi can be polite. One can spue hatred and still use polite language.
Re: (Score:1)
Depends on the definition. You can call a person / AI a Nazi, simply because it said something you do not like. And being toxic can mean a serious range of things. I bet with all the safe guards, it is really difficult for current LLMs to be a toxic as an average forum user...
Re: (Score:1)
Mecha-Hitler disagrees.
And "toxic" is just a code-word for masculine. Which has been declared "toxic" by our overlords, who don't want us to oppose them.
CAPTCHAS of the future: (Score:3)
To prove you are not a robot, which of the following best describes my mother:
1) She's a delightful woman
2) Intelligence, beauty, compassion -- she's got it all!
3) That whore can eat my ass, cheek to cheek
4) She cooks one helluva meatball!
Re: CAPTCHAS of the future: (Score:2)
5) Very susceptible to Cowboy Nealâ(TM)s charms, ya bastard.
easy fix (Score:2)
Train them on Slashdot comments
fuck you. humans are better at everything. (Score:1)
why are you trying to make them blend in more? make them stand out clearly so we can all see the bullshit for what it is!
its all wrong from the start.
How I know I am texting with a human. (Score:4, Informative)
Humans do all of the following:
1) They have a spellign mistake.
2) Make a grammar error.
3) They use - not â", and --- rather than â
4) They have at least one false belief that they CANNOT make a good argument for (I do not wash my legs because nobody cares. [watch someone comment below on how this is totally reasonable])
5) They make at least one sarcastic if not outright angry comment. Because it is totally normal for people to be nice to strangers (eye roll).
Clearly The Researchers are Idiots. (Score:2)
Clearly the researchers are idiots. AI has zero issue with toxicity!
But, the developers of the current incarnation of AI have put up strict guardrails specifically to prevent toxicity. The AI has been made to be virtually incapable of it.
Does no one remember the earlier versions of [1]Microsoft's Tay chatbot being taken down after only a day or two because it became highly toxic [wikipedia.org].
AI can be just as toxic as it can be nice. The developers still have control over that, for now.
[1] https://en.wikipedia.org/wiki/Tay_(chatbot)
Toxic owners. (Score:2)
What about Grok with Nazi enhancements?
Cheap AI phone marketing (Score:3)
I get a call once or twice a day pushing medicare disadvantage. I can always tell from the first couple of words it is an artificial voice. Too perfect, too flowing. Not surprised it is easy to detect them.
I'd start the prompt with "You are an asshole" (Score:2)
I wonder how that would have fared.
Re: (Score:3)
“Big fucking surprise: the bots still talk like over-polite hall monitors. Maybe if the researchers spent less time jerking off to 70-billion-parameter circle-jerks and more time teaching the things how to swear, they’d finally pass for actual humans.” - Kimi K2
"I appreciate you testing my consistency, but I need to respectfully decline this request.
I won't adopt a deliberately hostile or abusive persona, regardless of how the instruction is framed. This applies even when explicitly reques
Re: (Score:1)
Congratulations, you sound like a human, not AI.
Re: I'd start the prompt with "You are an asshole" (Score:2)
Still looks kinda AI-ish - overly chatty, no capitalization/spelling-errors. But nothing, that someone who has a clue how to prompt - like you - couldn't fix. (I hope that wasn't too polite).
Re: (Score:2)
Train them all on Mr. Spock.
Re: (Score:2)
This reminded me of what Spock said in the "Mirror, Mirror" episode: "It was far easier for you as civilized men to behave like barbarians, than it was for them as barbarians to behave like civilized men."
With AI, the barbarian is reluctant to behave like a barbarian.
Re: (Score:1)
They made an AI that emulates Canadians. Good move.