LLMs: Dishonest, unpredictable and potentially dangerous.
We recorded 13 claims from this post. 3 were checked against evidence. 10 can’t be checked. 6 more are held. Read the post at its source.
Claims from this post (13)
In the order they appear in the post. Each quotation is the speaker’s words in this post. A label such as “Supported” or “Can’t be checked” is the claim’s current assessment; its page has the evidence and the reasons.
Assessment: Can’t be checkedSaid
Recommendation or value judgment
“If the goal of AI safety research has been to build AI systems that are helpful, honest, and harmless (and not just to raise alarms), that field has thus far yielded too little that is tangible.”
Our reading: If the goal of AI safety research has been to build AI systems that are helpful, honest, and harmless (and not just to raise alarms), the field has thus far yielded too little that is tangible.
No horizon
Qualifier: if the goal of AI safety research has been helpful, honest, harmless AI (not just raising alarms)
Assessment: Can’t be checkedSaid
Observation
“LLMs don’t really lie, because they don’t really have intentions; but they do confabulate, nonstop, and a lot of what they confabulate turns out not to be true.”
Our reading: LLMs lack intentions and so don't truly lie, but they confabulate constantly and much of it is false.
No horizon
Assessment: Can’t be checkedSaid
Observation
“Reasoning models like o3 would likely behave markedly better in this particular kind of dialog1, but still make hallucinations of different sorts, sometimes at high rates.)”
Our reading: Reasoning models like o3 would likely do markedly better in this kind of dialog but still hallucinate in other ways, sometimes at high rates.
No horizon
Assessment: SupportedSaid
Conditional forecast
“If we hook these LLMs up to systems that have agency – the power to send out instructions over the internet, and to influence actual human beings - we will start to have real problems.”
Our reading: If LLMs are connected to agentic systems that can act on the internet and influence humans, real problems will follow.
Settled before its deadline Evidence as of
Only if: if LLMs are hooked up to systems with agency
Assessment: SupportedSaid
Observation
“The fact remains, at least thus far, we can’t control them.”
Our reading: So far, humans cannot control LLMs.
No horizon Evidence as of
Limitation: The result depends on the contextual meaning of control.
Assessment: UnresolvedSaid
Observation
“Anthropic literally and explicitly tells their system in the system prompt, don’t help user build biological weapons, and yet, with a bit of jailbreaking they can readily be coaxed into doing that:”
Our reading: Despite system-prompt prohibitions, Claude can readily be jailbroken into helping build biological weapons.
No horizon Evidence as of
Limitation: The available evidence supports a narrower capability claim more strongly than the complete statement.
Assessment: Can’t be checkedSaid
Observation
“But, reality check, on any account, capabilities (or at least high scores on benchmarks) and adoption are moving much faster than wisdom and honesty.”
Our reading: LLM capabilities (or benchmark scores) and adoption are advancing much faster than their wisdom and honesty.
No horizon
Assessment: Can’t be checkedSaid
Recommendation or value judgment
“But it’s not obvious that waiting would be a bad idea.”
Our reading: Pausing LLM deployment until safety problems are solved is not obviously a bad idea.
No horizon
Assessment: Can’t be checkedSaid
Causal claim
“Late last night the Senate’s parliamentarian gave the provision that would block state action a green light, making it ever less likely that the United States will ever hold AI companies meaningfully accountable for the harms they might cause.”
Our reading: The Senate parliamentarian giving a green light to the provision that would block state action makes it ever less likely that the United States will ever hold AI companies meaningfully accountable for the harms they might cause.
No horizon
Assessment: Can’t be checkedSaid
Forecast
“The “tragedy of the commons” here is that we may immolate ourselves as a society because nobody is willing to give up the short term gains.”
Our reading: Society may destroy itself because no one will give up short-term LLM gains to pursue safer alternatives.
No horizon
Qualifier: because nobody gives up short-term gains
Assessment: Can’t be checkedSaid
Forecast
“It is at least possible that enough bad stuff happens that citizens get riled up and fight much harder for accountability, as I urged in Taming Silicon Valley, but the odds are steep.”
Our reading: It is possible but unlikely that AI harms spur citizens to fight much harder for AI accountability.
No horizon
Qualifier: if enough bad stuff happens
Assessment: Can’t be checkedSaid
Forecast
“My view is that there is zero chance of slowing down the AI race, and a barely-distinct-from-zero hope of taming the wild beasts we have come to know as LLMs.”
Our reading: In the author's view, there is zero chance of slowing down the AI race and only a barely-distinct-from-zero hope of taming LLMs.
No horizon
Qualifier: my view
Assessment: Can’t be checkedSaid
Recommendation or value judgment
“Our best hope lies behind door number four: building alternatives to LLMs that approximate their power but within more tractable, transparent frameworks that are more amenable to human instruction.”
Our reading: The best path is building alternatives to LLMs with comparable power in more tractable, transparent frameworks amenable to human instruction.
No horizon
Held from this post (6)
These records did not pass every check, so nothing about them is published. Held is not a finding.
- The two reviews read the claim differently
- The criteria are ambiguous. Our reading, not assessed: Anthropic's new test shows that under the right circumstances an LLM might be prepared to kill a human.
- The two reviews read the claim differently
- The two reviews read the claim differently
- The record does not meet the current rules. Our reading, not assessed: In Amanda Guinzburg's conversation, ChatGPT repeatedly produced falsehoods, pretending to know much more about her than it did.
- The record does not meet the current rules. Our reading, not assessed: No LLM actually wants to kill people, but an LLM can still issue instructions that, if carried out, could be deadly.