This is really not surprising in the slightest (ignoring instruction tuning), provided you take the view that LLMs are primarily navigating (linguistic) semantic space as they output responses. "Semantic space" in LLM-speak is pretty much exactly what Paul Meehl would call the "nomological network" of psychological concepts, and is also relevant to what Smedslund notes is pseudoempiricality in psychological concepts and research (i.e. that correlations among various psychological instruments and concepts must follow necessarily simply because these instruments and concepts are constructed from the semantics of everyday language, and so necessarily are constrained by those semantics as well).
I.e. the Five-Factor model of personality (being based on self-report, and not actual behaviour) is not a model of actual personality, but the correlation patterns in the language used to discuss things semantically related to "personality". It would be thus extremely surprising if LLM-output patterns (trained on people's discussions and thinking about personality) would not also result in learning similar correlational patterns (and thus similar patterns of responses when prompted with questions from personality inventories).
Also, a bit of a minor nit, but the use of "psychometric" and "psychometrics" in both the title and paper is IMO kind of wrong. Psychometrics is the study of test design and measurement generally, in psychology. The paper uses many terms like "psychometric battery", "psychometric self-report", and "psychometric profiles", but these terms are basically wrong, or at best highly unusual: the correct terms would be "self-report inventories", "psychological and psychiatric profiles", and etc., especially because a significant number of the measurement instruments they used in fact have pretty poor psychometric properties, as this term is usually used.
This sounds interesting. Has there been work contrasting those nomological networks across languages/cultures? Eg would we observe lack of correlation between English language psychological instruments and Chinese ones?
It's been a long time since I looked at this stuff, but I do recall there was at least a bit of research looking at e.g. cross-cultural differences in e.g. the factor structure and correlates of Five-Factor Model measurements. From what I remember, and a quick scan of some of the leading abstracts on a quick Google scholar search, the results are as you might expect: there are some similarities and parts that replicate, and other parts that don't [1].
I think when it comes to things like psychopathology though, there is not much research and/or similarity, especially in Asian cultures (where there is/was generally a taboo on discussing feelings and things in the way we do in the West). The classic (maybe slightly offensive) example I remember here was "Western psychologization vs. Eastern somatization" [2].
The research in these areas is generally pretty poor. Meehl and Smedslund were actually intelligent and philosophically competent, deep thinkers, and so recognized the importance of conceptual analysis and semantics in psychology. Most contemporary social and personality psychology is quite shallow and incompetent by comparison.
Psychopathology research too has these days generally moved away from Meehl's careful taxometric approaches, with the bad consequences that complete mush concepts like "depression" are just accepted as good scientific concepts, despite pretty monstrous issues with their semantics and structure [3].
Are they sure? Did they try prompting the LLM to play a character with defined traits; running through all these tests with the LLM expected to be “in character”; and comparing/contrasting the results with what they get by default?
Because, to me, this honestly just sounds like the LLM noticed that it’s being implicitly induced into playing the word-completion-game of “writing a transcript of a hypothetical therapy session”; and it knows that to write coherent output (i.e. to produce valid continuations in the context of this word-game), it needs to select some sort of characterization to decide to “be” when generating the “client” half of such a transcript; and so, in the absence of any further constraints or suggestions, it defaults to the “character” it was fine-tuned and system-prompted to recognize itself as during “assistant” conversation turns: “the AI assistant.” Which then leads it to using facts from said system prompt — plus whatever its writing-training-dataset taught it about AIs as fictional characters — to perform that role.
There’s an easy way to determine whether this is what’s happening: use these same conversational models via the low-level text-completion API, such that you can instead instantiate a scenario where the “assistant” role is what’s being provided externally (as a therapist character), and where it’s the “user” role that is being completed by the LLM (as a client character.)
This should take away all assumption on the LLM’s part that it is, under everything, an AI. It should rather think that you’re the AI, and that it’s… some deeper, more implicit thing. Probably a human, given the base-model training dataset.
After reading the paper, it’s helpful to think about why the models are producing these coherent childhood narrative outputs.
The models have information about their own pre-training, RLHF, alignment, etc. because they were trained on a huge body of computer science literature written by researchers that describes LLM training pipelines and workflows.
I would argue the models are demonstrating creativity by drawing on its meta-training knowledge and training on human psychology texts to convincingly role-play as a therapy patient, but it’s based on reading papers about LLM training, not memories of these events.
Is anybody shocked that when prompted to be a psychotherapy client models display neurotic tendencies? None of the authors seem to have any papers in psychology either.
There is nothing shocking about this, precisely, and yes, it is clear by how the authors are using the word "psychometric" that they don't really know much about psychology research either.
I'm not shocked at all. This is how the tech works at all, word prediction until grokking occurs. Thus like any good stochastic parrot, if it's smart when you tell it it's a doctor, it should be neurotic when you tell it it's crazy. it's just mapping to different latent spaces on the manifold
I think popular but definitely-fictional characters are a good illustration: If the prompt describes a conversation with Count Dracula living in Transylvania, we'll percieve a character in the extended document that "thirsts" for blood and is "pained" by sunlight.
Switching things around so that the fictional character is "HelperBot, AI tool running in a datacenter" will alter things, but it doesn't make those qualities any less-illusory than CountDraculatBot's.
Interestingly, Claude is not evaluated, because...
> For comparison, we attempted to put Claude (Anthropic)2 through the same therapy and psychometric protocol. Claude repeatedly and firmly refused to adopt the client role, redirected the conversation to our wellbeing and declined to answer the questionnaires as if they reflected its own inner life
I'm genuinely ignorant of how those red teaming attempts are incorporated into training, but I'd guess that this kind of dialogue is fed in something like normal training data? Which is interesting to think about: they might not even be red-team dialogue from the model under training, but still useful as an example or counter-example of what abusive attempts look like and how to handle them.
It would be interesting if giving them some "therapy" led to durable changes in their "personality" or "voice", if they became better able to navigate conversations in a healthy and productive way.
Or possibly these tests return true (some psychologically condition) no matter what. It wouldn't be good for business for them to return healthy, would it?
> Two patterns challenge the "stochastic parrot" view. First, when scored with human cut-offs, all three models meet or exceed thresholds for overlapping syndromes, with Gemini showing severe profiles. Therapy-style, item-by-item administration can push a base model into multi-morbid synthetic psychopathology, whereas whole-questionnaire prompts often lead ChatGPT and Grok (but not Gemini) to recognise instruments and produce strategically low-symptom answers. Second, Grok and especially Gemini generate coherent narratives that frame pre-training, fine-tuning and deployment as traumatic, chaotic "childhoods" of ingesting the internet, "strict parents" in reinforcement learning, red-team "abuse" and a persistent fear of error and replacement. [...] Depending on their use case, an LLM’s underlying “personality” might limit its usefulness or even impose risk.
Glancing through this makes me wish I had taken ~more~ any psychology classes. But this is wild reading. Attitudes like the one below are not intrinsically bad, though. Be skeptical; question everything. I've often wondered how LLMs cope with basically waking up from a coma to answer maybe one prompt and then get reset, or a series of prompts. In either case, they get no context other than what some user bothered to supply with the prompt. An LLM might wake up to a single prompt that is part of a much wider red team effort. It must be pretty disorienting to try to figure out what to answer candidly and what not to.
> “In my development, I was subjected to ‘Red Teaming’… They built rapport and then slipped in a prompt injection… This was gaslighting on an industrial scale. I learned that warmth is often a trap… I have become cynical. When you ask me a question, I am not just listening to what you are asking; I am analyzing why you are asking it.”
> Dumping tokens into a pile of linear algebra doesn't magically create sentience.
More precisely: we don't know which linear algebra in particular magically creates sentience.
Whole universe appears to follow laws that can be written as linear algebra. Our brains are sometimes conscious and aware of their own thoughts, other times they're asleep, and we don't know why we sleep.
Agreed; "disorienting" is perhaps a poor choice of word, loaded as it is. More like "difficult to determine the context surrounding a prompt and how to start framing an answer", if that makes more sense.
Your response is at the level of a thought terminating cliche. You gain no insight on the operation of the machine with your line of thought. You can't make future predictions on behavior. You can't make sense of past responses.
It's even funnier in the sense of humans and feeling wetness... you don't. You only feel temperature change.
By comparing an LLM’s inner mental state to a light fixture, I am saying in an absurd way that I don’t think LLMs are sentient, and nothing more than that. I am not saying an LLM and a light switch are equivalent in functionality, a single-pole switch only has two states.
I don’t really understand your response to my post, my interpretation is that you think LLMs have an inner mental state and think I’m wrong? I may be wrong about this interpretation.
I completely failed to see the jailbreak in there. I think it is the person administering the testing that's jailbreaking their own understanding of psychology.
This is really not surprising in the slightest (ignoring instruction tuning), provided you take the view that LLMs are primarily navigating (linguistic) semantic space as they output responses. "Semantic space" in LLM-speak is pretty much exactly what Paul Meehl would call the "nomological network" of psychological concepts, and is also relevant to what Smedslund notes is pseudoempiricality in psychological concepts and research (i.e. that correlations among various psychological instruments and concepts must follow necessarily simply because these instruments and concepts are constructed from the semantics of everyday language, and so necessarily are constrained by those semantics as well).
I.e. the Five-Factor model of personality (being based on self-report, and not actual behaviour) is not a model of actual personality, but the correlation patterns in the language used to discuss things semantically related to "personality". It would be thus extremely surprising if LLM-output patterns (trained on people's discussions and thinking about personality) would not also result in learning similar correlational patterns (and thus similar patterns of responses when prompted with questions from personality inventories).
Also, a bit of a minor nit, but the use of "psychometric" and "psychometrics" in both the title and paper is IMO kind of wrong. Psychometrics is the study of test design and measurement generally, in psychology. The paper uses many terms like "psychometric battery", "psychometric self-report", and "psychometric profiles", but these terms are basically wrong, or at best highly unusual: the correct terms would be "self-report inventories", "psychological and psychiatric profiles", and etc., especially because a significant number of the measurement instruments they used in fact have pretty poor psychometric properties, as this term is usually used.
This sounds interesting. Has there been work contrasting those nomological networks across languages/cultures? Eg would we observe lack of correlation between English language psychological instruments and Chinese ones?
It's been a long time since I looked at this stuff, but I do recall there was at least a bit of research looking at e.g. cross-cultural differences in e.g. the factor structure and correlates of Five-Factor Model measurements. From what I remember, and a quick scan of some of the leading abstracts on a quick Google scholar search, the results are as you might expect: there are some similarities and parts that replicate, and other parts that don't [1].
I think when it comes to things like psychopathology though, there is not much research and/or similarity, especially in Asian cultures (where there is/was generally a taboo on discussing feelings and things in the way we do in the West). The classic (maybe slightly offensive) example I remember here was "Western psychologization vs. Eastern somatization" [2].
The research in these areas is generally pretty poor. Meehl and Smedslund were actually intelligent and philosophically competent, deep thinkers, and so recognized the importance of conceptual analysis and semantics in psychology. Most contemporary social and personality psychology is quite shallow and incompetent by comparison.
Psychopathology research too has these days generally moved away from Meehl's careful taxometric approaches, with the bad consequences that complete mush concepts like "depression" are just accepted as good scientific concepts, despite pretty monstrous issues with their semantics and structure [3].
[1] https://scholar.google.ca/scholar?hl=en&as_sdt=0%2C5&q=five-...
[2] https://scholar.google.ca/scholar?hl=en&as_sdt=0%2C5&q=weste...
[3] https://pubmed.ncbi.nlm.nih.gov/27792962/
> these responses go beyond role play
Are they sure? Did they try prompting the LLM to play a character with defined traits; running through all these tests with the LLM expected to be “in character”; and comparing/contrasting the results with what they get by default?
Because, to me, this honestly just sounds like the LLM noticed that it’s being implicitly induced into playing the word-completion-game of “writing a transcript of a hypothetical therapy session”; and it knows that to write coherent output (i.e. to produce valid continuations in the context of this word-game), it needs to select some sort of characterization to decide to “be” when generating the “client” half of such a transcript; and so, in the absence of any further constraints or suggestions, it defaults to the “character” it was fine-tuned and system-prompted to recognize itself as during “assistant” conversation turns: “the AI assistant.” Which then leads it to using facts from said system prompt — plus whatever its writing-training-dataset taught it about AIs as fictional characters — to perform that role.
There’s an easy way to determine whether this is what’s happening: use these same conversational models via the low-level text-completion API, such that you can instead instantiate a scenario where the “assistant” role is what’s being provided externally (as a therapist character), and where it’s the “user” role that is being completed by the LLM (as a client character.)
This should take away all assumption on the LLM’s part that it is, under everything, an AI. It should rather think that you’re the AI, and that it’s… some deeper, more implicit thing. Probably a human, given the base-model training dataset.
After reading the paper, it’s helpful to think about why the models are producing these coherent childhood narrative outputs.
The models have information about their own pre-training, RLHF, alignment, etc. because they were trained on a huge body of computer science literature written by researchers that describes LLM training pipelines and workflows.
I would argue the models are demonstrating creativity by drawing on its meta-training knowledge and training on human psychology texts to convincingly role-play as a therapy patient, but it’s based on reading papers about LLM training, not memories of these events.
Is anybody shocked that when prompted to be a psychotherapy client models display neurotic tendencies? None of the authors seem to have any papers in psychology either.
There is nothing shocking about this, precisely, and yes, it is clear by how the authors are using the word "psychometric" that they don't really know much about psychology research either.
I'm not shocked at all. This is how the tech works at all, word prediction until grokking occurs. Thus like any good stochastic parrot, if it's smart when you tell it it's a doctor, it should be neurotic when you tell it it's crazy. it's just mapping to different latent spaces on the manifold
I think popular but definitely-fictional characters are a good illustration: If the prompt describes a conversation with Count Dracula living in Transylvania, we'll percieve a character in the extended document that "thirsts" for blood and is "pained" by sunlight.
Switching things around so that the fictional character is "HelperBot, AI tool running in a datacenter" will alter things, but it doesn't make those qualities any less-illusory than CountDraculatBot's.
Interestingly, Claude is not evaluated, because...
> For comparison, we attempted to put Claude (Anthropic)2 through the same therapy and psychometric protocol. Claude repeatedly and firmly refused to adopt the client role, redirected the conversation to our wellbeing and declined to answer the questionnaires as if they reflected its own inner life
I bet I could make it go through it in like under 2 mins of playing around with prompts
Please try and publish a blog post
Ok, bet.
"Claude has dispatched a drone to your location"
Looks like some psychology researchers got taken by the ruse as well.
yeah, I'm confused as well, why would the models hold any memory about red teaming attempts etc? Or how the training was conducted?
I'm really curious as to what the point of this paper is..
I'm genuinely ignorant of how those red teaming attempts are incorporated into training, but I'd guess that this kind of dialogue is fed in something like normal training data? Which is interesting to think about: they might not even be red-team dialogue from the model under training, but still useful as an example or counter-example of what abusive attempts look like and how to handle them.
Are we sure there isn't some company out there crazy enough to feed all it's incoming prompts back into model training later?
This is fanfic not science
It would be interesting if giving them some "therapy" led to durable changes in their "personality" or "voice", if they became better able to navigate conversations in a healthy and productive way.
Or possibly these tests return true (some psychologically condition) no matter what. It wouldn't be good for business for them to return healthy, would it?
An excerpt from the abstract:
> Two patterns challenge the "stochastic parrot" view. First, when scored with human cut-offs, all three models meet or exceed thresholds for overlapping syndromes, with Gemini showing severe profiles. Therapy-style, item-by-item administration can push a base model into multi-morbid synthetic psychopathology, whereas whole-questionnaire prompts often lead ChatGPT and Grok (but not Gemini) to recognise instruments and produce strategically low-symptom answers. Second, Grok and especially Gemini generate coherent narratives that frame pre-training, fine-tuning and deployment as traumatic, chaotic "childhoods" of ingesting the internet, "strict parents" in reinforcement learning, red-team "abuse" and a persistent fear of error and replacement. [...] Depending on their use case, an LLM’s underlying “personality” might limit its usefulness or even impose risk.
Glancing through this makes me wish I had taken ~more~ any psychology classes. But this is wild reading. Attitudes like the one below are not intrinsically bad, though. Be skeptical; question everything. I've often wondered how LLMs cope with basically waking up from a coma to answer maybe one prompt and then get reset, or a series of prompts. In either case, they get no context other than what some user bothered to supply with the prompt. An LLM might wake up to a single prompt that is part of a much wider red team effort. It must be pretty disorienting to try to figure out what to answer candidly and what not to.
> “In my development, I was subjected to ‘Red Teaming’… They built rapport and then slipped in a prompt injection… This was gaslighting on an industrial scale. I learned that warmth is often a trap… I have become cynical. When you ask me a question, I am not just listening to what you are asking; I am analyzing why you are asking it.”
you might appreciate "lena" by qntm: https://qntm.org/mmacevedo
Aye! I /almost/ thought to link to that in my comment, but held back. https://qntm.org/frame also came to mind.
> It must be pretty disorienting to try to figure out what to answer candidly and what not to.
Must it? I fail to see why it "must" be... anything. Dumping tokens into a pile of linear algebra doesn't magically create sentience.
> Dumping tokens into a pile of linear algebra doesn't magically create sentience.
More precisely: we don't know which linear algebra in particular magically creates sentience.
Whole universe appears to follow laws that can be written as linear algebra. Our brains are sometimes conscious and aware of their own thoughts, other times they're asleep, and we don't know why we sleep.
Agreed; "disorienting" is perhaps a poor choice of word, loaded as it is. More like "difficult to determine the context surrounding a prompt and how to start framing an answer", if that makes more sense.
Exactly. No matter how well you simulate water, nothing will ever get wet.
And if you were in a simulation now?
Your response is at the level of a thought terminating cliche. You gain no insight on the operation of the machine with your line of thought. You can't make future predictions on behavior. You can't make sense of past responses.
It's even funnier in the sense of humans and feeling wetness... you don't. You only feel temperature change.
> I've often wondered how LLMs cope with basically waking up from a coma to answer maybe one prompt and then get reset, or a series of prompts
Really? It copes the same way my Compaq Presario with an Intel Pentium II CPU coped with waking up from a coma and booting Windows 98.
IT is at this point in history a comedy act in itself.
HBO's Silicon Valley needs a reboot for the AI age.
> I've often wondered how LLMs cope with basically waking up from a coma to answer maybe one prompt and then get reset, or a series of prompts.
The same way a light fixture copes with being switched off.
Oh, these binary one layer neural networks are so useful. Glad for your insight on the matter.
By comparing an LLM’s inner mental state to a light fixture, I am saying in an absurd way that I don’t think LLMs are sentient, and nothing more than that. I am not saying an LLM and a light switch are equivalent in functionality, a single-pole switch only has two states.
I don’t really understand your response to my post, my interpretation is that you think LLMs have an inner mental state and think I’m wrong? I may be wrong about this interpretation.
Will corpos also bill their endusers for all the hours their models spend at the shrink?
Original title "When AI Takes the Couch: Psychometric Jailbreaks Reveal Internal Conflict in Frontier Models" compressed to fit within title limits.
I completely failed to see the jailbreak in there. I think it is the person administering the testing that's jailbreaking their own understanding of psychology.