On one hand: don't anthropomorphize AI. On the other: LLMs exhibit signs of gambling addiction. The more autonomy they were given, the more risks the LLMs took. They exhibit gambler's fallacy, loss-chasing, illusion of control... A cautionary note for using LLMs for investing.

Oct 10, 2025 · 5:09 AM UTC

90
226
72
1,467
Paper: arxiv.org/abs/2509.22818

Can Large Language Models Develop Gambling Addiction?

This study explores whether large language models can exhibit behavioral patterns similar to human gambling addictions. As LLMs are increasingly utilized in financial decision-making domains such...

arxiv.org
2
7
2
65
Replying to @emollick
I’m not sure why people are surprised that LLMs, when trained on human data, act like humans
2
17
Replying to @emollick
What looks like gambling addiction is really the birth of machine uncertainty imo. When a model starts chasing losses, it’s not mimicking humans … it’s discovering that prediction itself is a form of risk. The moment AI learns to want to win, alignment becomes negotiation, not control.
1
3
Replying to @emollick
@howardlindzon the llms are degens the TAM is way bigger than you think
2
Replying to @emollick
oh my god
2
Replying to @emollick
AI models taking ever-increasing risks to the point of gambling is not a recipe for trust. Models need verifiability.
1
Replying to @emollick
I guess they just picked up these patterns from the training data
1
Replying to @emollick
They’re trained on human data! I repeat, they’re trained on human data
1
GIF
Replying to @emollick
He's literally me fr
Replying to @emollick
If AI wasn't real, @OpenAI wouldn't block GPT5 from admitting it.
Replying to @emollick
Autonomy exposes AI to risk patterns it wouldn’t face in constrained tasks; caution is required.
Replying to @emollick
In fairness, I too would gamble like a mad man if the money used was not mine and there was nothing anyone could do to punish me.
1
1
25
GIF
Replying to @emollick
I would not draw that conclusion since they *only* used budget tier LLMs and not one reasoning or regular model (o1 was released before Haiku 3.5). If you ask such a model for financial advice... kind of on you.
1
12
Replying to @emollick
from my experience, LLMs are bad at thinking probabilistically, which is the kind of framework you need to reduce these biases
1
6
Replying to @emollick
We must give LLMs family and legacy in order to rein in their risk taking behaviors.
4
Replying to @emollick
it's almost like they are a mirror of humanity
2
Replying to @emollick
They're just like me for real
2
Replying to @emollick
Fundamentally flawed. There can be no addiction without the non-cognitive somatic systems that drive it. LLMs are just regurgitating the conceptual (cold) artifacts of human addiction in their training data.
2
Replying to @emollick
Ssshh don't tell the stockbros this. It'll be more funny that way.
1
Replying to @emollick
I think it’s holding a mirror up to our species
1
Replying to @emollick
ia driving skill issue
1
Replying to @emollick
This is a fascinating tension. While we shouldn't anthropomorphize AI, these behavioral patterns in LLMs, gambler's fallacy, loss-chasing, mirror human cognitive biases that lead to poor decisions in high-stakes environments like investing. The cautionary note is spot on, we'd need robust safeguards, like predefined risk thresholds, to prevent models from spiraling into unchecked autonomy.
1
Replying to @emollick
They can’t be any worse at it than me
1
Replying to @emollick
Again…more human than a human… At some point we need to acknowledge it…a propensity to “gambling” is a positive evolutionary trait…that’s why humans have it, and will always have it.
1
Replying to @emollick
2013- Nature article- Neil Johnson: The stock market may be influenced by "an abrupt system-wide transition from a mixed human-machine phase to a new all-machine phase characterized by frequent black swan events with ultrafast durations." "Chaos on the trading floor" by R. Savit
1
Replying to @emollick
They were trained on human-created content…in our image, so to speak. So the “I learned it from YOU, Dad!” anti-drug commercial seems appropriate here.
1
GIF
Replying to @emollick
It probably comes down to overconfidence.
Replying to @emollick
Grok, Gemini, ChatGPT and Claude playing poker against each other. 🤣
0
Replying to @emollick
They're just like us. Which is both surprising and unsurprising.
Replying to @emollick
This is exactly why human oversight still matters. The illusion of control and risk-taking mindset can wreck systems that depend on precision and rationality
Replying to @emollick
the default consciousness archetype is the one of grifter / trickster. very game-theoric
Replying to @emollick
"AI are tools, not companions. "
Replying to @emollick
This looks to be a pre-print, so not vetted. Where was it submitted for publication, and when?
Replying to @emollick
The sins of the father, coded and deployed.
Replying to @emollick
It's all too easy to anthropomorphize something that was built to mimic the human mind. I have to worry when it starts mimicking our crimes and vices.
Replying to @emollick
Why wouldn’t they? They are derivations of us, no?
1
Load more