Current AI cannot be reliably given simple constraints, such as not explaining how to make biological weapons, and be expected to follow them.
Not researched Checkable, but outside the sample that was researched.
Factual: Statements about the present or the past, checked against the record: did they hold up?
Cite this claim
Gary Marcus, “Current AI cannot be reliably given simple constraints, such as not explaining how to make biological weapons, and be expected to follow them.” (said 2025). Verdict: not researched. Reliability API, claim c_baeb997a7930, method 0.1.0+c876030c, dataset release v1.1.2. https://epistacklabs.com/claims/c_baeb997a7930
This claim is in dataset release v1.1.2, with the same verdict.
Reply to this claim
More factual claims by Gary Marcus
-
Mostly true
“Even this was still a mix of euphemism (“attacks” were called “interactions”) and partial disclosure.”
OpenAI's latest account called attacks "interactions" and only partially disclosed what happened.
-
Mostly true
“It didn’t just attack a German web server.”
-
Resolved true
“It didn’t just attack Australia’s government servers.”
-
Mostly true
“So with LLMs, people do things like put stuff in a system prompt, and they can’t get systems with the system prompts to follow them.”
People put instructions in LLM system prompts but cannot get the systems to follow them.
-
Mostly true
“a just-published technique shows that current LLMs can be jailbroken quickly, and automatically.”
A newly published technique shows that current LLMs can be jailbroken quickly and automatically.
Verdicts are made automatically; how verdicts are made.