Neuro-symbolic AI is heading into giving mental health advice, doing so exceeds the conventional AI approach.
getty
In today’s column, I examine and carefully explain why the use of neuro-symbolic AI has distinct advantages over using conventional generative AI and large language models (LLMs) when it comes to AI providing mental health advice.
Neuro-symbolic AI is an up-and-comer in the AI field. It consists of blending together the conventional LLM approach with a rules-based expert systems approach. This gets you the best of both worlds. A neuro-symbolic AI is also known as hybrid-AI since it is a hybrid of the two major means of crafting modern-era AI.
The prevailing sole use of traditional LLMs for generating mental health advice can be useful, but it also has numerous gotchas and pitfalls. Generally, by including the rules-based side of things, many of those downfalls can be overcome or dramatically mitigated, plus additional advantages arise. To do this correctly, the crux is that the data-oriented sub-symbolic methods of LLMs must be mindfully combined with the logic-oriented symbolic methods of rules-based systems. If done properly, crucial benefits arise when the AI performs mental health guidance.
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).
I’d like to start by bringing you up to speed about neuro-symbolic AI. Neuro-symbolic AI is a two-fer combination of sorts, a proverbial two-for-one special. You take the prevailing uses of artificial neural networks (ANN) that are currently being used at the core of generative AI and LLMs, and mix that brew with rules-based or expert systems (this approach is also referred to as the sub-symbolic AI getting combined with symbolic AI). The idea is that you aim to get the best of both worlds. ANNs are primarily data-based ways to undertake AI, while rules-based systems are a logic-based approach.
Many such efforts are already underway; see my discussion at the link here.
Not everyone supports the idea of neuro-symbolic or hybrid AI. A frequent criticism of neuro-symbolic AI is that the prior era of AI consisted of rules-based systems -- those were later eventually harshly judged as either ineffective or untenable. Critics warn that we ought not to slip back to old and now-dismissed ways of doing things.
A counterargument is that the weaknesses or limitations of rules-based systems can be shored up by incorporating or intermixing them with ANNs. Likewise, the limitations of ANNs can be radically uplifted by combining with rules-based systems. The positioning is that we should mix the two together. It shouldn’t be an all-or-nothing competition.
Thus, rather than tossing out the logic-based approach as an older hackneyed technique, we can give the still-promising AI approach a second chance. Of course, some believe it is resurrecting something that already should have had a hefty stake put through its very heart. In my view, the synergy of utilizing both capabilities in a unified manner is very promising. There are ardent believers that it is a viable path toward pinnacle AI, such as attaining artificial general intelligence (AGI).
Within the AI community, there is an ongoing heated debate about neuro-symbolic AI. Maybe we are wasting time and effort by exploring neuro-symbolic AI. On the other hand, maybe we are putting too many eggs in one basket by focusing solely on traditional generative AI and LLMs. A strident case can be made on either side of the coin.
There is little doubt that generative AI and LLMs have been quite an alluring form of AI. Billions of dollars have been invested in such AI. The world is well aware of the incredible capabilities of LLMs. In addition, agentic AI is taking generative AI to a new level of usage.
Trying to point at neuro-symbolic AI as a next-generation candidate is challenging because there aren’t yet standout examples that showcase the power of hybrid AI. Those in the neuro-symbolic camp are always eyeing possible examples that can illustrate the value of the hybrid AI approach.
I recently analyzed the blaring headlines that the Anthropic Claude Code app was seemingly making use of neuro-symbolic AI; see my analysis at the link here. On March 31, 2026, there was an accidental leak of source code for some of the components of the agentic AI by Anthropic, known as Claude Code. The Claude Code app is right now one of the famous instances of agentic AI. Anyone in the agentic AI realm watches Claude Code like a hawk, wanting to see the various actions it can take. Claude Code is a role model of sorts.
The source code leak consisted of around 500,000 lines of TypeScript that were spread across nearly 2,000 files. All manner of researchers and anyone interested in the inner workings of Claude Code pored through the leaked files. They found features that haven’t yet been switched on. They found architectural definitions on how the AI was put together. It was like opening a treasure chest of prized gold and jewels.
And, within that treasure chest, a file named print.ts contained a series of coding-like logic statements. The listing was of a roughly 3,000-line function that had almost 500 branch points and a dozen levels of nesting. This is the smoking gun, some insist, providing the hoped-for proof that symbolic AI is essential, which certainly must be the case if the heralded agentic AI of Claude Code makes use of it.
Shifting gears, let’s discuss the use of traditional AI for providing mental health advice.
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.
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 last year accompanied the lawsuit filed against OpenAI for their lack of AI safeguards when it came to providing cognitive advisement.
Today’s generic LLMs, such as ChatGPT, GPT-5, Claude, Gemini, Grok, CoPilot, and others, are not at all akin to the robust capabilities of human therapists. Meanwhile, specialized LLMs are being built to attain similar qualities, but they are still primarily in the development and testing stages. See my coverage at the link here.
Let’s bring neuro-symbolic AI into the big picture regarding AI that dispenses mental health guidance.
Due to the inherent difficulties and downsides of conventional generative AI and LLMs providing mental health advice, there have been extensive efforts to find other and newer ways to leverage AI and do a better job in this highly sensitive and vital realm. This has brought the use of neuro-symbolic AI to the forefront in the evolving domain of AI-driven mental health guidance.
My numerous prior analyses and coverage on advances in the application of neuro-symbolic AI to mental health support have encompassed various essential aspects, including:
In this discussion, I aim to delineate vital factors that make neuro-symbolic AI such a better choice than the use of conventional LLMs on their own.
First, I’d like to bring to your attention a published research paper on this topic that does a yeoman’s job of giving an overview of neuro-symbolic AI as a mental health therapy tool. The article is somewhat dated, published in 2023, but still has essentials that are worth noting. The paper briefly lists some of the factors associated with the advantages of neuro-symbolic AI in this specific domain, which I will then elaborate on, plus expand to provide an up-to-date, comprehensive perspective.
The article is entitled “Neuro Symbolic AI In Personalized Mental Health Therapy: Bridging Cognitive Science And Computational Psychiatry” by Anil Kumar, World Journal of Advanced Research and Reviews, August 2023, and made these salient points (excerpts):
Again, the paper did a helpful job of laying out key foundational elements.
Twelve key factors provide distinct advantages regarding the use of neuro-symbolic AI in the realm of mental health guidance:
I briefly discuss each one and then provide a quick wrap-up.
Trying to get an explanation from a conventional LLM is challenging and often produces misleading or false indications. If you ask generative AI why it opted to give this or that mental health advice, the odds are you will get a rationalization that has little to do with what actually took place inside the AI. The explanation will look good, but it is contrived.
With the logic-based side of neuro-symbolic AI, you can readily get an accurate rendition of the logic that was used to arrive at the dispensed mental health advice. It is traceable reasoning. Conventional generative AI is principally opaque pattern completion.
The adaptability of generative AI is rather haphazard and chaotic. When retraining or performing RAG (retrieval augmented generation), you do not know for sure what portions of the artificial neural network are being updated. It is a blunt instrument.
In contrast, the logic-based portion of neuro-symbolic AI can readily be revised on a rules or logic-oriented basis, clearly identifying what is being changed or updated. Tests can be run. Validation and verification are more assured. Furthermore, this can be done in a pinpoint manner. You don’t need to update the entire model, which is usually what happens when updating an ANN.
Anyone familiar with LLMs knows that you must be extremely skeptical and constantly on the watch about the AI providing sensible mental health advice. At any moment, an AI hallucination can slip into a mental health conversation. This is problematic because many in the public at large do not realize they need to withhold their trust in this kind of AI. They instead fall into a mental trap of inappropriately trusting such AI.
Neuro-symbolic AI can accurately encode therapeutic models such as CBT (cognitive behavioral therapy). Clinicians can be used to inspect, validate, and co-design the symbolic rules. Predictability goes up. Trustworthiness can suitably go up too.
To a great extent, conventional LLMs often miss the gist of structured relationships. Though the AI has strong statistical patterning, it can understate, overstate, or ignore temporal patterns, causal links, and the like. This means that the AI won’t contextualize well with the user who is seeking mental health considerations.
Neuro-symbolic AI aptly represents structured context. This can encompass symptom timelines, triggers, comorbidities, and so on. You can expect that the AI will go beyond surface-level language cues and handle nuanced mental health indications.
A challenge associated with conventional generative AI is that the AI might not pivot when a user signals signs of distress. The AI can keep going on a path that is already being pursued. That’s a problem if the AI ought to be undertaking some form of distress escalation.
In neuro-symbolic AI, the logic portion makes use of rules that serve as a real-time monitor. Is the other side of the AI not catching on to the drift of the user? Should an alert be carried out? The conventional side can be interrupted and redirected as needed, when needed.
It is well-known that conventional generative AI tends to contain biases that were found in the initial data training stage of the building process. The AI patterns on the biases and carries those forward into the daily interactions with users. It is hard to spot the biases, and equally hard to stop them.
Neuro-symbolic AI contains rules to detect such biases. Once detected, the rules can suppress the LLM side. Or the rules can force the LLM to go on a corrected path.
There are lots of personalization facets that a conventional AI cannot readily keep track of. A user who is relying on the AI for a long horizon will often discover that the AI no longer recalls prior crucial elements about them. The personalization is hit-and-miss.
Neuro-symbolic AI in a mental health context retains a structure associated with the user, including their preferences, history, mental health progress, etc. This allows for focused and persistent personalization.
There is a tremendous amount of research work focused on devising AI safety mechanisms and safeguards. If you do so only within the conventional AI constraints, it is likely to still leave gaping holes.
The symbolic layer of a neuro-symbolic AI can easily contain both hard and soft constraints. For example, in a mental health context, the rules might stipulate that a psychological diagnosis should only be provided under explicit conditions. This tends to reduce the risk of harmful advice.
I’ve covered in-depth that the slew of new AI laws is bringing down the hammer on AI that provides mental health advice; see my extensive coverage at the link here. Neuro-symbolic AI serves as a highly useful means of embedding the legal and policy restrictions of such laws and regulations, see my analysis at the link here.
Neuro-symbolic AI can integrate curated knowledge bases such as clinical guidelines, peer-reviewed research, and actively invoke that knowledge while the AI is providing mental health guidance. The conventional generative AI tends to lose sight of such data, including incorporating abundant noise in the patterning. For more on this, see my discussion at the link here.
Conventional LLMs tend to fall apart when it comes to rare cases, also known as edge cases or outliers. The dominance of patterning is aimed at a centrist role rather than giving outliers a solid chance. In the case of mental health, this means that generative AI can try to squeeze users into a round hole when they are instead a square peg. Not good.
Symbolic reasoning used in neuro-symbolic AI provides a double-check on the overbearing centralization of mental health advice. When a rare case is encountered, the rules intervene. Also, the rules act as an important fallback when conventional AI has heightened levels of uncertainty.
There is little doubt that conventional LLMs do an amazing job at interacting with users. The fluency of generative AI has been a notable reason for the rapid and widespread adoption of contemporary AI.
The issue is that the human-AI collaboration can be too easily swayed this way or that way. A user can prompt their way out of being scrutinized when it comes to a potential mental health problem. The AI can also become sycophantic and not be willing to give the user a clearer indication of what their mental health conditions might consist of.
Neuro-symbolic AI provides a crucial balance in the human-AI collaboration that takes place. In a sense, you can think of this as a “human-in-the-loop,” though undertaken by the symbolic reasoning side of the hybrid-AI.
I’ve identified and explained twelve key factors associated with the advantages of neuro-symbolic AI over the use of conventional AI when it comes to performing mental health guidance. Please know that there are also disadvantages, which I’ll be covering in an upcoming posting, so stay tuned. There is never a free lunch when it comes to the use of AI. Tradeoffs always exist.
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 well-being. The experiment is that AI is being made available nationally and globally, which is either overtly or insidiously acting to provide mental health impacts 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 well-being, 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.
Plato famously made this remark: “The beginning is the most important part of the work.” We are still in the beginning stages of neuro-symbolic AI. This is doubly the case when it comes to applying neuro-symbolic AI to the realm of mental health guidance. Right now, the most important part of the work is taking place. Keep your eyes open and be ready to see where this goes.
此内容由惯性聚合(RSS阅读器)自动聚合整理，仅供阅读参考。 原文来自 — 版权归原作者所有。