Expert systems are computer programs designed to mimic the decision-making ability of a human specialist in a narrow domain, using a structured collection of domain-specific rules and facts rather than general intelligence. They were among the earliest practical successes of artificial intelligence, and their core architecture still influences how we think about embedding human expertise in software. While the term may sound like a relic of the 1980s AI boom, expert systems remain operational in industries from oil refining to medical diagnostics, and their DNA is showing up in unexpected places inside today’s large language models and hybrid AI research.
How an Expert System Works
At its simplest, an expert system has three main parts. The knowledge base stores the domain expertise, usually as a collection of “if-then” rules or structured facts. The inference engine is the reasoning mechanism that applies those rules to a specific situation, deciding which rules fire and in what order. And the user interface handles communication between the system and whoever is asking it questions, sometimes including an explanation facility that can walk the user through why a particular conclusion was reached.
The inference engine can reason in two directions. Forward chaining starts from known facts or symptoms and works toward a conclusion. If you type in that your laptop screen is flickering and the display driver recently updated, forward chaining would march through rules until it lands on a diagnosis. Backward chaining works the other way: it starts with a hypothesis and asks whether the evidence supports it. Research comparing the two approaches in a laptop-repair expert system found that forward chaining suits symptom-based input better, while backward chaining is more effective when you already have a suspected cause and want to verify it.1Al-DYAS. Perbandingan Metode Backward Chaining dan Forward Chaining pada Rancangan Sistem Pakar Deteksi Kerusakan Laptop Warrior Computer Mendalo Jambi Many production systems use a blend of both, switching direction depending on the query.
The rules themselves can be expressed in several formal ways. A comparative analysis of four common representation schemes evaluated logic-based representations, production rules, semantic nets, and frames, each with different strengths for different kinds of knowledge.2Information Processing & Management. A framework for the comparative analysis and evaluation of knowledge representation schemes Production rules are the most widely used because they map naturally onto the way human experts articulate their reasoning: “If the patient has a fever and a rash, then consider measles.” Frames and semantic nets work better when the knowledge is more about relationships between concepts than about step-by-step reasoning.
The Knowledge Acquisition Bottleneck
Building an expert system sounds straightforward in principle: find a human expert, extract what they know, and encode it as rules. In practice, this has always been the hardest part. Knowledge acquisition was recognized early as the major constraint on developing expert systems, with the core difficulty lying in identifying the varying structures of domain knowledge and matching them to suitable encoding techniques.3Expert Systems. The Knowledge Acquisition Bottleneck: Time for Reassessment?
The problem goes deeper than logistics. Human experts often cannot articulate why they make the decisions they do. A seasoned geologist might glance at a rock formation and sense that it is promising for copper, but if you press her on the exact rules she is applying, she may struggle to decompose her intuition into discrete steps. Traditional programming skills turn out to be of little use when faced with the challenge of extracting potential knowledge-base information from a human expert.4Computers and Electronics in Agriculture. Knowledge acquisition for expert system development This spawned an entire sub-discipline called knowledge engineering, whose practitioners acted as translators between domain experts and the formal rule language the system could ingest.
Some projects tried to sidestep the bottleneck by automating knowledge acquisition. The DENDRAL project, one of the earliest expert systems, was specifically designed to automate the acquisition of task-specific knowledge in chemistry, using detailed heuristics about molecular structure rather than relying entirely on manual rule writing.5Artificial Intelligence. DENDRAL: A case study of the first expert system for scientific hypothesis formation That idea of learning rules from data rather than from interviews would later become the foundation of machine learning, though the two traditions developed largely in parallel for decades.
Where Expert Systems Have Been Deployed
Expert systems found homes in industries where human expertise was scarce, expensive, or needed around the clock. A few domains stand out for how long these systems stayed in active use.
In mineral exploration, the Prospector project began in the mid-1970s as a research effort to help geologists explore for hidden mineral deposits. By the late 1980s it had grown into a full-sized knowledge-based system used for regional mineral resource assessments.6Mathematical Geology. Prospector II: Towards a knowledge base for mineral deposits Prospector famously predicted the location of a molybdenum deposit in Washington state that human geologists had missed, which became one of the most cited success stories in AI history.
Industrial process monitoring is another natural fit. The LUBRES system was implemented for real-time fault diagnosis of a lubricating oil refining process, monitoring the chemical process continuously and providing operators with suggestions when abnormal situations arose.7Expert Systems with Applications. LUBRES: An expert system development and implementation for real-time fault diagnosis of a lubricating oil refining process A similar system was developed for the cryogenic plant of the EAST fusion reactor, where the cryogenic control system could detect alarms but struggled to diagnose root causes without expert-level reasoning layered on top.8Fusion Engineering and Design. Design of a real-time fault diagnosis expert system for the EAST cryoplant In both cases the selling point was the same: the expert system never takes a break, never retires, and can be deployed at sites where the handful of human experts cannot always be physically present.
Finance adopted expert systems for credit risk evaluation. One approach combined fuzzy logic and neural network techniques to model the thought process a credit-rating expert goes through when evaluating an entrepreneur’s creditworthiness.9Expert Systems. Expert system design for credit risk evaluation using neuro‐fuzzy logic This hybrid approach allowed the system to handle the vague, graded judgments that pure rule-based systems struggle with, like the difference between “mostly reliable” and “highly reliable” repayment histories.
More recently, a rule-based expert system for intelligent power line detection and diagnosis achieved a diagnostic accuracy of about 92% across complex operational conditions and environmental factors.10AIP Conference Proceedings. Design of a rule base for intelligent power line detection and diagnosis based on a mivar expert system That is not perfect, but it is good enough to serve as a first-pass filter that escalates only the ambiguous cases to a human inspector.
How They Handle Uncertainty
Real-world expertise is rarely black-and-white. A doctor does not say “the patient definitely has pneumonia”; she says “it is very likely pneumonia given the X-ray and the fever.” Expert systems need a way to reason with this kind of fuzziness, and the field has tried several approaches. A review of uncertainty models identified five popular candidates: symbolic representation, MYCIN-style certainty factors, Bayesian probability, Dempster-Shafer theory, and fuzzy set logic, each with its own strengths and weaknesses for medical applications.11PubMed. The representation of uncertainty in medical expert systems
MYCIN, one of the most famous medical expert systems, introduced certainty factors as a practical workaround. Each rule carried a number between negative one and positive one indicating how strongly the evidence supported or denied a conclusion. This was easier for doctors to provide than formal probability estimates, but it lacked the mathematical rigor of Bayesian approaches. Bayesian methods, in turn, required data on how common each disease was in the population, data that often did not exist for rare conditions. Fuzzy logic offered a third path, allowing terms like “high fever” and “moderate fever” to be represented as overlapping categories rather than hard thresholds.
No single uncertainty model won the debate. In practice, many modern systems mix approaches, using crisp rules where the domain knowledge is clear and probabilistic or fuzzy reasoning where it is not. The credit-risk system mentioned earlier is a good example of this hybrid strategy, layering neural-network learning on top of fuzzy categories to capture the subtleties of human judgment.
Brittleness and the Boundaries of Knowledge
The most persistent criticism of expert systems is that they are brittle. They work well within the boundaries of their knowledge base, but the moment a situation falls outside those boundaries, performance can collapse without warning. A human expert encountering something unfamiliar will usually recognize that she is out of her depth and say so. A rule-based system may simply apply the closest matching rule and deliver a confident but wrong answer.
A pilot study in immunohematology illustrated this vividly. Researchers compared two design approaches: one in which the expert system automatically completed subtasks for the technologist, and another in which it served as a monitoring tool that critiqued technologists as they worked. The auto-completion design induced a 29 percent increase in errors compared to the critiquing design.12Immunohematology. Dealing with brittleness in the design of expert systems for immunohematology The takeaway was not that expert systems are dangerous but that how you position them relative to the human operator matters enormously. When the system takes over and the human disengages, the human loses the ability to catch the system’s mistakes.
Brittleness also shows up in maintenance. Real-world domains change: new diseases are discovered, regulations shift, equipment gets upgraded. The rules in the knowledge base need to be updated accordingly, and since those rules were originally extracted through painstaking knowledge engineering, updating them can be almost as expensive as building the system in the first place. This is one reason many 1980s-era expert systems fell into disuse. They worked fine when first deployed but gradually drifted out of alignment with reality.
The 1980s Boom and the AI Winter
Expert systems were the commercial face of artificial intelligence throughout the 1980s. Companies invested heavily in specialized hardware and software, and the market for expert-system shells and consulting exploded. Governments poured funding into AI research programs partly on the promise that expert systems would automate high-value professional work.
The bubble burst around 1990. Hardware costs dropped, making the specialized machines less attractive. The knowledge acquisition bottleneck meant that many projects took far longer and cost far more than projected. And the brittleness problem proved harder to solve than optimists had hoped. The resulting disillusionment contributed to what is commonly called the AI winter, a period of reduced funding and public skepticism toward artificial intelligence that lasted through much of the 1990s.
Recent scholarship has drawn parallels between that boom-and-bust cycle and today’s enthusiasm around large language models, noting the similarities, analogs, and differences between the two eras and projecting potential outcomes for AI research if the current enthusiasm wanes.13AAAI Publications. Implications for AI Research: Applying Lessons from the Expert Systems Boom and Bust to the Current Large-Language Model Boom The comparison is instructive. Expert systems failed to generalize beyond narrow domains; large language models generalize impressively but struggle with reliability and factual grounding. The two failure modes are nearly mirror images of each other, which is part of why researchers are now trying to combine them.
Legal Liability When Software Gives Advice
When an expert system gives bad advice and someone follows it, who is responsible? This question was raised almost as soon as expert systems entered commercial use. As expert systems receive increasing application, legal issues inevitably arise concerning contractual matters, potential liability, and possible governmental regulation, with liability centering on losses alleged to have resulted from someone following erroneous advice.14Computers & Industrial Engineering. Legal liability and system safety applied to expert systems
The liability question has only grown more urgent. If a medical expert system recommends a treatment and the patient is harmed, potential defendants include the software developer, the knowledge engineer who encoded the rules, the domain expert who provided the knowledge, and the institution that deployed the system. Traditional product-liability law, designed for physical goods, fits awkwardly. Is the expert system a product or a service? Is the knowledge base a form of professional advice? Courts in different jurisdictions have answered these questions differently, and the rise of AI-assisted decision-making has renewed the debate without settling it.
One practical consequence is that many deployed expert systems are explicitly positioned as decision-support tools rather than autonomous decision-makers. The LUBRES and EAST cryoplant systems both provide “operation suggestions” rather than issuing commands. That framing is partly engineering and partly legal strategy: if the human operator retains final authority, the liability picture is clearer.
The User Interface Problem
An expert system is only useful if people actually interact with it well, and interface design has historically been underrated as a factor in success or failure. One early research effort explored building a dialogue system that explicitly modeled both the person interacting with the system and the person being reasoned about, in this case a clinician and a patient, in order to generate responses that were empathetic and meaningful to both.15PubMed. Modeling all dialogue system participants to generate empathetic responses That idea, building a user model into the system so it can tailor its communication, was ahead of its time and anticipated the personalization features we now expect from modern software.
The immunohematology study’s finding about auto-completion versus critiquing designs is relevant here too. A system that silently does work for the user encourages passivity; one that actively engages the user’s reasoning encourages vigilance. The interface is not a cosmetic layer on top of the intelligence. It shapes how much of the intelligence the user actually benefits from. Many well-built expert systems failed in practice not because their rules were wrong but because the interface made it too easy for users to accept outputs uncritically or too hard to understand why the system reached a given conclusion.
Transparency as a Lasting Advantage
One feature expert systems have that most modern AI does not is built-in explainability. Because the reasoning is encoded as explicit rules, the system can trace back through its chain of inference and show you exactly why it reached its conclusion. When a medical expert system says “this patient likely has condition X,” it can list the rules that fired and the evidence that triggered each one.
This matters more now than it did in the 1980s. A review of ethical guidelines from multiple organizations found that building and maintaining user trust was the most commonly cited reason for incorporating transparency in AI systems. Other organizations highlighted that transparency supports security and helps users understand when AI is making an actual decision versus when it is offering a recommendation to support human judgment.16ScienceDirect. Transparency and explainability of AI systems: From ethical guidelines to requirements Neural networks and large language models are notoriously opaque; they can give you an answer but not a transparent chain of reasoning. Expert systems can. In regulated industries like medicine, aviation, and finance, where you may need to explain and defend every decision, that transparency is not a quaint feature. It is a regulatory requirement.
Expert Systems Meet Modern AI
The most interesting development in recent years is the convergence of expert-system-style symbolic reasoning with neural network-based machine learning, an approach called neuro-symbolic AI. The idea is to get the best of both worlds: the learning ability and pattern recognition of neural networks combined with the logical consistency and explainability of rule-based systems.
One technique involves embedding symbolic rules directly within a neural network as constraints. In Logic Tensor Networks, for example, logical rules are represented as tensors, which allows neural networks to learn while maintaining logical consistency. The primary challenge is integrating symbolic knowledge without compromising the learning flexibility of the neural networks.17Intelligent Systems with Applications. A review of neuro-symbolic AI integrating reasoning and learning for advanced cognitive systems Get it right, and you have a system that can both learn new patterns from raw data and respect the hard logical constraints that the domain requires.
Another approach injects structured knowledge graphs into large language models. A recent study found that augmenting a smaller model with a knowledge graph achieved reasoning scores close to those of much larger models like GPT-4o while using roughly 100,000 times fewer active parameters.18arXiv. Injecting Knowledge Graphs into Large Language Models The knowledge graph is doing what the expert system’s knowledge base always did: providing structured, verified facts that the reasoning engine can draw on. The engine just happens to be a neural network now instead of a rule-matching algorithm.
This convergence suggests that expert systems did not disappear so much as dissolve into the broader AI landscape. The principles that made them work, structured domain knowledge, explicit reasoning chains, uncertainty management, keep resurfacing in new forms. Whether the label “expert system” survives is mostly a matter of branding. The underlying architecture, a curated knowledge base paired with an inference mechanism, keeps proving too useful to abandon entirely.
When an Expert System Is Still the Right Tool
For all the excitement around deep learning and large language models, there are situations where a classical expert system remains the better choice. If you need full auditability of every decision, if the domain is narrow and well-understood, if the knowledge changes slowly, and if the cost of a wrong answer is high, a rule-based system has advantages that no neural network can match today. Air traffic control decision support, pharmaceutical drug interaction checkers, and industrial safety interlocks are all areas where expert systems or their direct descendants remain in production.
The economics also favor expert systems in some settings. Training a large neural network requires enormous amounts of data and compute. An expert system can be built from a handful of interviews with the right specialist and run on modest hardware. For a small company with a well-defined problem and access to a domain expert willing to spend a few weeks articulating their knowledge, a rule-based system can be running in months rather than years. The knowledge acquisition bottleneck is real, but so is the data acquisition bottleneck that afflicts machine learning in domains where labeled training examples are scarce or expensive to produce.
The question is not whether expert systems are better or worse than neural networks. They solve different problems with different trade-offs. The interesting frontier is the space in between, where structured knowledge and statistical learning meet, and where the next generation of AI systems is being built on foundations that expert systems laid decades ago.