Neurosymbolic AI The Fix For LLM Hallucinations
The primary concern with big tech's venture into artificial intelligence isn't a potential takeover of humanity. Instead, it's the persistent issue of Large Language Models (LLMs) like OpenAI's ChatGPT, Google's Gemini, and Meta's Llama frequently getting things wrong—a problem that has proven difficult to resolve.
Down with endless data. (Image credit: Alexander Supertramp via Shutterstock)
The Persistent Problem of AI Hallucinations
These errors, often called hallucinations, gained notoriety with incidents such as US law professor Jonathan Turley being falsely accused of sexual harassment by ChatGPT in 2023.
OpenAI's apparent response—programming ChatGPT to avoid questions about Turley—is hardly a fair or adequate solution. Addressing hallucinations on a case-by-case basis after they occur is clearly not a sustainable approach.
The same applies to LLMs amplifying stereotypes or providing Western-centric answers. Compounding this is a significant lack of accountability for this widespread misinformation, as it's often challenging to determine how an LLM arrived at a particular erroneous conclusion.
Why Current LLM Fixes and Regulations Fall Short
A vigorous debate surrounding these issues emerged after the 2023 release of GPT-4, OpenAI's latest major LLM. However, this debate has arguably subsided since then, despite the problems persisting.
For instance, the EU swiftly passed its AI Act in 2024, aiming to lead global oversight in this field. Yet, the act largely depends on AI companies to regulate themselves, without truly addressing the fundamental issues. This hasn't deterred tech companies from deploying LLMs to hundreds of millions of users worldwide and collecting their data without adequate scrutiny.
Meanwhile, the latest tests indicate that even the most sophisticated LLMs continue to be unreliable. Despite this, leading AI companies still resist taking responsibility for these errors.
Unfortunately, the tendencies of LLMs to misinform and reproduce bias cannot be resolved through gradual improvements over time. With the rise of agentic AI, where users can assign complex tasks to an LLM (like booking holidays or optimizing bill payments), the potential for serious problems is set to increase significantly.
Introducing Neurosymbolic AI: A More Reliable Future?
The emerging field of neurosymbolic AI offers a potential solution to these critical issues, while also aiming to reduce the enormous quantities of data currently needed for training LLMs. So, what exactly is neurosymbolic AI, and how does it operate?
How Neurosymbolic AI Bridges Learning and Logic
LLMs function using a technique called deep learning. They are trained on vast amounts of text data and use advanced statistics to infer patterns that dictate the most probable next word or phrase in any given response. Each model, along with all its learned patterns, is stored in arrays of powerful computers within large data centers, known as neural networks.
LLMs can appear to reason through a process called chain-of-thought, generating multi-step responses that mimic how humans might logically reach a conclusion, based on patterns observed in their training data.
Undoubtedly, LLMs represent a significant engineering achievement. They excel at summarizing text and translation and can potentially improve the productivity of users who are diligent and knowledgeable enough to identify their mistakes. Nevertheless, they possess a great capacity to mislead because their conclusions are always rooted in probabilities, not genuine understanding.
A common workaround is known as "human-in-the-loop," ensuring that humans using AI systems still make the final decisions. However, merely apportioning blame to humans doesn't solve the underlying problem, as they can still be misled by AI-generated misinformation.
LLMs now require such vast amounts of training data that we are resorting to feeding them synthetic data—data created by other LLMs. This synthetic data can replicate and amplify existing errors from its source, causing new models to inherit the weaknesses of their predecessors. Consequently, the cost of programming AIs to be more accurate after their initial training, a process known as "post-hoc model alignment," is skyrocketing.
It also becomes increasingly difficult for programmers to identify what's going wrong because the number of steps in the model's thought process grows ever larger, making error correction progressively harder.
Neurosymbolic AI combines the predictive learning capabilities of neural networks with teaching the AI a series of formal rules that humans use for more reliable deliberation. These include logic rules (e.g., "if it's raining, then everything outside is normally wet"), mathematical rules (e.g., "if a = b and b = c, then a = c"), and the agreed-upon meanings of elements like words, diagrams, and symbols. Some of these rules are inputted directly into the AI system, while it deduces others itself by analyzing its training data and performing "knowledge extraction."
Key Advantages of the Neurosymbolic Approach
This approach should lead to an AI that will not hallucinate and will learn faster and more intelligently by organizing its knowledge into clear, reusable components. For instance, if the AI has a rule about things being wet outside when it rains, it doesn't need to retain every single example of things that might be wet outside; the rule can be applied to any new object, even one it has never encountered before.
During model development, neurosymbolic AI also integrates learning and formal reasoning through a process known as the "neurosymbolic cycle." This involves a partially trained AI extracting rules from its training data and then instilling this consolidated knowledge back into the network before undergoing further training with data.
This method is more energy-efficient because the AI doesn't need to store as much data. The AI is also more accountable because it's easier for a user to control how it reaches particular conclusions and improves over time. Furthermore, it's fairer because it can be designed to follow pre-existing rules, such as: "For any decision made by the AI, the outcome must not depend on a person's race or gender."
The Evolution of AI: Towards a Third Wave
The first wave of AI in the 1980s, known as symbolic AI, was actually based on teaching computers formal rules that they could then apply to new information. Deep learning followed as the second wave in the 2010s, and many now see neurosymbolic AI as the third wave.
Current Applications and the Path to General Use
It's easiest to apply neurosymbolic principles to AI in niche areas because the rules can be clearly defined. So, it's no surprise that we've seen it first emerge in Google's AlphaFold, which predicts protein structures to aid drug discovery, and AlphaGeometry, which solves complex geometry problems.
For more broad-based AIs, China's DeepSeek utilizes a learning technique called "distillation," which is a step in the same direction. However, to make neurosymbolic AI fully feasible for general models, more research is still needed to refine their ability to discern general rules and perform knowledge extraction.
Building a Future of Trustworthy AI
It remains unclear to what extent LLM makers are already working on this. They certainly sound like they're heading towards trying to teach their models to think more cleverly, but they also seem committed to scaling up with ever-larger amounts of data.
The reality is that if AI is to continue advancing, we will need systems that can adapt to novelty from only a few examples, that check their understanding, that can multitask and reuse knowledge to improve data efficiency, and that can reason reliably in sophisticated ways.
This way, well-designed digital technology could potentially even offer an alternative to regulation, because the checks and balances would be built into the architecture and perhaps standardized across the industry. There's a long way to go, but at least there's a path ahead.
This edited article is republished from The Conversation under a Creative Commons license. Read the original article.
Artur d’Avila Garcez is a Professor of Computer Science at City St George's, University of London, and a leading researcher in neurosymbolic AI. He co-founded the NeSy conference series, has authored over 250 publications, and helped develop one of the first neurosymbolic systems for learning and reasoning. Garcez is also Editor-in-Chief of the Neurosymbolic AI journal and holds fellowships with the British Computer Society and the UK Higher Education Academy.
