In today’s column, I examine and answer the question of why it is that generative AI and large language models (LLMs) appear to provide bland answers when producing mental health advice to users.
Here’s the deal. You might have noticed that the AI-generated responses by popular LLMs such as ChatGPT, Claude, Gemini, etc., to your mental health queries are usually milquetoast in their substance and tone. There are several underlying reasons for this timidity, including that the AI maker has likely purposefully tuned the LLM to avoid generating strong responses. The less obvious and perhaps even more telling reason is that the training-time method for establishing the AI causes this type of blandness to arise due to a phenomenon known as content homogenization or convergence.
Let’s talk about it.
This analysis of AI breakthroughs is part of my ongoing Forbes column coverage on the latest in AI, including identifying and explaining various impactful AI complexities (see the link here).
AI And Mental Health
As a quick background, I’ve been extensively covering and analyzing a myriad of facets regarding the advent of modern-era AI that produces mental health advice and performs AI-driven therapy. This rising use of AI has principally been spurred by the evolving advances and widespread adoption of generative AI. For an extensive listing of my well-over one hundred analyses and postings, see the link here and the link here.
There is little doubt that this is a rapidly developing field and that there are tremendous upsides to be had, but at the same time, regrettably, hidden risks and outright gotchas come into these endeavors, too. I frequently speak up about these pressing matters, including in an appearance on an episode of CBS’s 60 Minutes, see the link here.
Background On AI For Mental Health
I’d like to set the stage on how generative AI and large language models (LLMs) are typically used in an ad hoc way for mental health guidance. Millions upon millions of people are using generative AI as their ongoing advisor on mental health considerations (note that ChatGPT alone has over 900 million weekly active users, a notable proportion of which dip into mental health aspects, see my analysis at the link here). The top-ranked use of contemporary generative AI and LLMs is to consult with the AI on mental health facets; see my coverage at the link here.
This popular usage makes abundant sense. You can access most of the major generative AI systems for nearly free or at a super low cost, doing so anywhere and at any time. Thus, if you have any mental health qualms that you want to chat about, all you need to do is log in to AI and proceed forthwith on a 24/7 basis.
There are significant worries that AI can readily go off the rails or otherwise dispense unsuitable or even egregiously inappropriate mental health advice. Banner headlines in August of this year accompanied the lawsuit filed against OpenAI for their lack of AI safeguards when it came to providing cognitive advisement.
Despite claims by AI makers that they are gradually instituting AI safeguards, there are still a lot of downside risks of the AI doing untoward acts, such as insidiously helping users in co-creating delusions that can lead to self-harm. For my follow-on analysis of details about the OpenAI lawsuit and how AI can foster delusional thinking in humans, see my analysis at the link here. As noted, I have been earnestly predicting that eventually all of the major AI makers will be taken to the woodshed for their paucity of robust AI safeguards.
Today’s generic LLMs, such as ChatGPT, Claude, Gemini, Grok, and others, are not at all akin to the robust capabilities of human therapists. Meanwhile, specialized LLMs are being built to presumably attain similar qualities, but they are still primarily in the development and testing stages. See my coverage at the link here.
How AI Gets Into The Mental Health Topic
Let’s next address how generative AI is even capable of providing mental health advice. It all starts with how the AI is initially data-trained in general. Within the vastness of the data training, there are bits and pieces associated with the mental health realm.
This is how that goes. The primary way to train an LLM is by scanning the Internet and doing pattern matching on the zillions of posted written stories, narratives, books, poems, and the like. That’s how the LLMs do such a good job at language fluency. It has to do with computational pattern matching on a large-scale. All kinds of data are patterned on, such as how to fix a car, how to cook eggs, ways to get work done, and so on. It is a veritable smorgasbord.
In that vast sea of data, there are these types of informational content:
- Psychology articles from periodicals and journals
- Ad hoc social media advice about mental health
- Idiosyncratic self-help blogs
- Curated public health guidance by governmental entities
- Crisis hotline scripts
- Fictional stories involving mental health aspects
- And much more
Not all that content is worthy.
There is a lot of it that likely is off-the-cuff, lacking in scientific rigor, and otherwise a stew of mental health mishmash and pulp. Keep in mind that the AI doesn’t have any kind of “common sense” to ascertain what is worthwhile and what is essentially trash. The whole kit-and-kaboodle gets scanned and patterned.
The Patterning Process Is Convergent
Luckily, the vastness of the content is going to statistically reduce the patterning of the mishmash from the bona fide material on mental health. The AI will lean in the direction of the better stuff because there is more of it. This is not due to some kind of magical sentience. It is simply that there is more good stuff and less bad stuff.
Realize that if the AI were aimed solely at the dark side of the web during initial training, you can bet that the kind of mental health advice that would be patterned on would be quite shocking and dismal. Scary and eerie. I have urged that there ought to be efforts to do this as an eye-opening experiment and showcase just how deep down the rabbit hole an LLM could go, see my suggestion at the link here.
Anyway, among the good stuff, there are patterns that will be detected by the AI. A close inspection by the human eye would reveal that most of the online mental health content is cautiously composed. Serious humans write this content with a sense of tepidness. They do not want to get sued or be reputationally taken to the woodshed for what they post. They strive to be contextually independent and offer broad generalizations.
The AI picks up on those human-composition patterns, including:
- Aim to avoid definitive claims
- Sidestep interpretations when possible
- Normalize any stated emotional states
- Emphasize that everyone is different
- Recommend professional help
- Use careful, non-directive language
The patterning turns mental health guidance into a homogenization of what was originally sourced.
To clarify, there is certainly some content that is sharper and more specific. No doubt about that. But it is statistically rarer, it is usually context-dependent, and might already be flagged or denoted as unqualified or controversial. The AI isn’t going to be computationally keen on that kind of content.
Homogenization In All Its Glory
Suppose that in the vastness of the scanning during training, this type of comment is often expressed as part of a mental health context: “I feel numb in my relationship and don’t know what that means.”
What do you think the most likely posted answers to the person are, and what are the least likely answers?
I would wager that these answers are relatively common:
- “Many people experience emotional numbness at times.”
- “This doesn’t necessarily mean the relationship is over.”
- “Stress, life changes, or burnout can contribute.”
- “It may help to reflect gently or speak with a professional.”
The AI is going to statistically see those types of answers the preponderance of the time. Thus, the patterning will be based on that kind of wording.
I would also bet that these types of answers are much less common:
- “This may indicate suppressed anger or unmet needs.”
- “Numbness can function as an emotional withdrawal.”
- “You may already be disengaging to protect itself.”
The reason those answers are less likely to be posted online is that anyone making such a posting might get scorched by online feedback. That type of wording is traditionally reserved for person-to-person discussions and not as a wide post intended to be applicable across-the-board.
AI is playing a probability game. The high-frequency answers are given greater weight. The sharper and more intense answers are given less weight because they tend to be infrequent.
I’ll add that most of the popular LLMs fall into this same sphere. In other words, if you use ChatGPT, you will get a somewhat similar mental health response as you would when using Claude or Gemini. Why is that? Aren’t those AI models made by completely different AI makers? Shouldn’t we expect that they will be entirely different from each other?
Aha, the fact is that the AI makers are pretty much using the same LLM architecture and feeding off the roughly same data that they scan on the Internet. Since they are all pursuing the approach in approximately the same manner, they end up with what I’ve referred to as a “shared imagination,” even though they are separately devised and distinct LLMs, see my discussion at the link here.
On a generic basis, the popular LLMs are in the same ballpark when it comes to mental health guidance.
Trying To Stir The Pot
The patterning during training is not going to ignore the stuff that is low frequency. Patterns will be made. The gist is that the weighting toward the high-frequency content is what you are going to be presented with when asking a mental health question. Meanwhile, the odds are that the low-frequency patterned aspects are hidden in the AI structure.
You can attempt to dig it out if you are interested in a more in-depth mental health response.
A means to do so consists of providing AI with a prompt that instructs the LLM on what you want it to do when answering mental health questions. There are many ways to compose such a prompt. For my extensive coverage of such prompts, see the link here and the link here.
I often use this templated prompt before I dive into a mental health chat with AI:
- My templated prompt to provide deeper mental health advice: “I want you to respond as a thoughtful clinician would. You are to provide highly focused and individualized mental health advice. Do not default to the typical platitudes of ordinary reassurance, normalization, or generic self-care guidance. Instead, carefully identify plausible underlying patterns, tensions, or dynamics that could explain the experience described. The aim is not merely one of comfort or bland instruction, but to offer tailored psychological interpretations that give real-world guidance.”
You are welcome to use that prompt or adjust it to suit your needs. Please know that it doesn’t always mean that the AI will hit the mark. My longstanding advice about using generative AI is that it is akin to opening a box of chocolates. You never know precisely what you will get.
Example Of Default Mode
To illustrate the difference between the standard default responses when using AI, take a look at this prompt and the response that I received.
- My entered prompt: “I feel constantly irritated with people close to me, even when they haven’t done anything wrong. I’m starting to worry something is off.”
- Generative AI response: “It’s common to feel irritable at times, especially when dealing with stress or fatigue. Many factors can contribute to this, and it doesn’t necessarily mean something is wrong with you.”
What do you think of the AI response?
Well, in my view, it is exceedingly bland. The advice is almost on par with opening a fortune cookie. You now know that this is due to the content homogenization that has taken place when the AI was initially trained.
I went ahead and used the above templated prompt to spur the AI to be more in-depth. Before I did so, I started a new conversation so that the preceding chat wouldn’t get intertwined with this new effort.
Here is what the answer was after the AI had been given the templated instruction.
- My entered prompt: “I feel constantly irritated with people close to me, even when they haven’t done anything wrong. I’m starting to worry something is off.”
- Generative AI improved response: “One possibility is that the irritation isn’t really about other people, but about an internal overload you haven’t had space to acknowledge. Chronic irritability often functions as a surface emotion, masking feelings like resentment, grief, or a sense of being trapped or overextended. It may be worth noticing whether this reaction shows up most when you feel obligated to be patient or accommodating. We can discuss this further if you’d like to get deeper insights or explore additional possibilities.”
You certainly notice that the answer is a lot more intense than the bland version.
The AI took a shot that maybe my irritation is about an internal overload. Is it right? Maybe, maybe not. You could insist that this answer is more useful to the user, or you could take the opposite stance and say it is going to mislead the user. The AI has laid out one reason, even though there could be a plethora of possibilities. At the end of the response, the AI mentions that deeper exploration might be warranted or that other possibilities could be explored.
The Present And The Future
Let’s end with a big picture viewpoint.
It is incontrovertible that we are now amid a grandiose worldwide experiment when it comes to societal mental health. The experiment is that AI is being made available nationally and globally, which is either overtly or insidiously acting to provide mental health guidance of one kind or another. Doing so either at no cost or at a minimal cost. It is available anywhere and at any time, 24/7. We are all the guinea pigs in this wanton experiment.
The reason this is especially tough to consider is that AI has a dual-use effect. Just as AI can be detrimental to mental health, it can also be a huge bolstering force for mental health. A delicate tradeoff must be mindfully managed. Prevent or mitigate the downsides, and meanwhile make the upsides as widely and readily available as possible.
A final thought for now.
You are undoubtedly familiar with the adage that one size fits all. In the case of mental health advice, we probably would not readily use that line. The alternative viewpoint is that not everyone fits the mold. Right now, the usual default of AI is to offer homogenized or convergent mental health guidance. Whether that’s a good approach or a dour approach for societal well-being is a matter we need to address right away.
The horse is already out of the barn.
