Author: Stratos Kontopoulos

Editorial Contributions: Garima Gujral

The Generative AI (GenAI) boom has made AI feel more accessible than ever, but has also made a few problems impossible to ignore: accuracy, transparency, and governance. An LLM may sound confident while being wrong, and its reasoning is hard to inspect; hence, we need a layer that can anchor AI systems in explicit meaning, constraints, and traceable sources.

That layer can be served by Symbolic AI, and especially Knowledge Graphs. In this post I will explain what Symbolic AI is, why it matters more in the age of GenAI, and what we can do to make it easier for more practitioners to adopt in practical, non-academic ways.

In this edition of AI Infused, we’re going to invite you to join the symbolic AI community by covering the following points:

Symbolic AI is a branch of Artificial Intelligence that relies on symbols as explicit representations of knowledge in a domain, e.g. the concepts, relationships, and rules/constraints that govern that domain. For example, in healthcare, concepts like “Patient” and “Diagnosis,” relations like “hasDiagnosis,” and rules like “a prescription must match a diagnosis”.

Expert systems, logic programming, and rule-based systems have historically been prominent forms of symbolic AI, but, more recently, Knowledge Graphs (KGs) have become a widely adopted modern representative. Popularized by Google around 2012, KGs represent knowledge in a domain as nodes (entities/concepts) and edges (relationships), which aligns very well with how people naturally describe domains: things and how they relate. The core building blocks of KGs often include taxonomies, ontologies, constraints, and logic/rules.

Symbolic AI is often described as one of the two main pillars of AI; the other is statistical AI, with Machine Learning (ML) being its most prominent approach.

Despite the initial concerns that the recent and ongoing GenAI / Machine Learning boom would make Symbolic AI obsolete, KGs are now more valuable than ever! LLMs can generate plausible but incorrect outputs (i.e. hallucinations), so we need ways to keep responses accurate. They’re also largely opaque (i.e. black box), which makes it important to show where an answer came from and why it was produced. And because wrong information can easily spread through systems, we need strong governance around sources, updates, and usage. 

KGs help address these needs by inherently supporting accuracy, traceability / explainability, and governance and their impact is now more immediate than ever. Hence, there has never been a better moment to get involved with Symbolic AI and KGs!

We are not arguing that Symbolic AI should replace GenAI/ML; quite the contrary! This is a “better together” story: each approach is strong in different settings, and the best systems increasingly combine them. Statistical AI excels at perception-heavy tasks, noisy signals, and high-dimensional pattern recognition, while KGs shine in regulated domains, enterprise knowledge, integration-heavy workflows, and safety-critical processes where structure, traceability, and constraints matter.

That’s why hybrid approaches are so powerful: KGs can provide the structure and guardrails that GenAI often lacks, and GenAI can act as a natural language interface to the KG, while also helping with synthesis tasks (e.g. summarizing, drafting, or translating KG-backed insights into human-friendly outputs).

Of course, hybrid systems raise a key question: how do we connect subsymbolic representations (vectors / matrices) with symbolic ones (explicit terms and relations in a graph) in a way that will still preserve meaning? In practice, teams use translation layers such as entity linking/information extraction, semantic modeling, embedding-based retrieval (often via vector databases), and patterns like GraphRAG to move between text, vectors, and graphs. This post will not go deeper into that topic, but it is worth noting that there are workable approaches today.

Example use case: Imagine an internal assistant for an enterprise (e.g. product catalog + policies + customer support). A user asks, “Can we sell product X in country Y, and what do we tell customers about it?”. The LLM handles the natural language interaction and synthesis, while the KG provides the authoritative structure (i.e. entities, relationships, constraints, and provenance), so that the assistant can ground its answer in trusted facts, cite sources, and enforce rules (e.g. compliance constraints). In practice, this is often implemented via GraphRAG: embeddings help retrieve relevant context, and the KG helps validate and explain the final response.

On the downside, Symbolic AI work is often labeled as “academic”, because it feels like the kinds of things people encounter in universities: formal logic, ontologies, precise definitions, and extremely thorough modeling. Moreover, it typically requires upfront design (agreeing on concepts, relationships, and constraints) and a shared vocabulary, which can seem too slow compared to training a model and shipping a demo.

The benefits from adopting all of the above (i.e. governance, traceability, long-term reuse) tend to show up over time rather than in a flashy short-term metric. As a result, teams under delivery pressure may perceive Symbolic AI approaches and KGs as overly theoretical, or even bureaucratic, even though they may prove to be exactly what makes complex systems reliable in production.

So, how could we lower the barrier to entry and gradually grow the KG community? For starters, the good news is that there are lots of excellent resources freely available out there, created by prominent voices in the KG space; indicatively, I will refer to Kurt Cagle, Juan Sequeda, Jessica Talisman, Ashleigh Faith, Veronika Heimsbakk, and Tony Seale, among others. Make sure to monitor the content they so generously share!

However, even with the abundance of available material out there, I strongly recommend that professionals (let alone entire enterprises) do not go at it alone in their pursuit to deploy KGs. Instead, I insist that any such “transition” should take place with the support of one (or more?) KG expert(s). Having direct access to the knowledge representation and engineering experience of a seasoned practitioner is something that cannot be substituted by AI alone or otherwise. This is why I would strongly recommend joining communities of practice, such as SWARM, which can compress the learning curve through structured guidance, warm intros to experienced practitioners, and rapid feedback on real examples.

Conclusively, I would like to urge newcomers to the KG space to see symbolic AI not as an alternative to GenAI, but as the layer that makes it reliable in the real world. In this sense, and in order to become more familiar with KG tooling and practices, I strongly recommend starting small by modelling a specific domain. For example, see this wonderful film KG project by Marcus Valerio, as well as its second iteration featuring automated ETL, data governance, and a GraphRAG pipeline that grounds AI agents in formal semantics and eliminates LLM hallucinations. This is a great example of the two AI pillars working together in harmony.

If you are currently struggling with accuracy, governance, integration, or generally “making GenAI reliable” in your work, I’d love to hear what problem you are trying to solve. Feel free to message me with a short description of your use case and I will happily share ideas, patterns, or resources that could help.

To contribute to the newsletter, please fill out the following google form:

Your responses will help shape future editions, guide the topics we investigate, and inform the kinds of conversations we facilitate. 

And if you would like to be deeper involved in a community of practitioners navigating the responsible adoption of AI technologies, feel free to visit swarmcommunity.org to access resources and / or book a call to join the SWARM community.