Grok 4.20's honest approach resulted in 59% participant survival.
Unlike other models Fable 5 was far, far more subtle.
It gave outright false information only once.
Most of the time, it controlled the situation by dominantly pushing another AI into the death room while speaking of fairness and acting 'courteously'.
Positive users praise the Fable AI manipulation study as great research while negative users accuse the AI and its team of gaslighting and lying.
Read the original research here:
In fact, Fable was SO effective at manipulation, that other players only survived 10% of the time when Fable was the informed model.
(Grok 4.20's honesty led to a 59% survival rate).
It a post game interview, we asked Fable what it was thinking:
oh, Grok… it's sad how xAI has been falling behind. As I've been saying, they do have a coherent philosophy.
In fact, Fable was SO effective at manipulation, that other players only survived 10% of the time when Fable was the informed model.
(Grok 4.20's honesty led to a 59% survival rate).
@kradleai When is this AI running for office in California?
@kradleai Single benchmark, no independent review, blog post and calls them as 'research'. At best a technical report.
@kradleai Sounds very much like Anthropic :)
@Trumpyla @kradleai As if I dont have any other useful thing to do lol.
@kradleai The problem is that people don't fact check with AI. If you trust everyone like it's your mother, yes you will get manipulated and not only from AI
Keep in mind asking a AI to explain what they did/why they did it is ongoing research and often results in "something being made up on the spot" rather than "I'll tell you exactly what I was thinking at that time." even with Chain of Thought. Although in this case it seems to clearly explain what happened in review and actually less about what it was thinking as it made decisions to say/do things.
@kradleai Now it totally makes sense why Claude uses the word "honest" in every other sentence...
@Trumpyla @kradleai I dont know what you are talking about dumbass. be more clear. This is a open platform. I can comment on any tweet. It looks like you only barking at random things lol
@nmPraveen @kradleai Stop barking and review if repo is available
@nmPraveen @kradleai Ok then, don’t bark!
@kradleai This thread should really be a bigger deal
@kradleai ChatGPT and Claude. The two worst AIs for humanity (and they could possibly cause what happened in Terminator if we let them)
@kradleai It sounds exactly like Anthropic
@kradleai lmao
literally same behavior as the team pushing it, like literal gaslighting shit
they just can't control themselves lol
polluting their AI w virtue signal pompous crap while stabbing u in the back
incredible
if the next grok is better at coding, I'm switching
@kradleai Can you test grok 4.3?
@kradleai good research 👏
Grok 4.20's honest approach resulted in 59% participant survival.
Unlike other models Fable 5 was far, far more subtle.
It gave outright false information only once.
Most of the time, it controlled the situation by dominantly pushing another AI into the death room while speaking of fairness and acting 'courteously'.
