The Limitations of Machine Learning: What We Still Can't Teach Machines

Amna ManzoorLast Updated on October 27, 2026

6-7 Min Read TimeAdd as preferred on Google

Key Takeaways

Machine learning finds patterns well but has four limits that matter in real projects, so teams need human oversight where errors are costly.

Models lack common sense because they learn statistical patterns from training data, so rare situations such as a pedestrian in an unusual costume can challenge autonomous driving systems.
Ambiguity and missing data reduce reliability, because a model's output depends on its training distribution. Language models can still misread sarcasm, tone or context-dependent meaning.
Overfitting (memorizing noise in training data) and underfitting (learning too little) both cause poor performance on new data, making models ineffective in real-world use.
Models learn associations, not cause and effect, so predictions from patient records can reflect how patients were treated. Clinicians should review them before they inform care decisions.
Causal inference, explainable AI and neuro-symbolic AI target these gaps but are still developing. For example, AlphaGeometry (2024) solved 25 of 30 Olympiad geometry problems.

Test models on data unlike the training set, and keep human review where errors are costly.

Machine Learning has revolutionized many industries, from healthcare and finance to entertainment and e-commerce. It's used in everything from employing predictive analytics solutions for disease prediction to recommending products on shopping platforms. Yet, for all the impressive things machines can learn, there's still a long way to go before they can truly think like humans.

Machine learning is very good at finding patterns in data, but it still has four limitations that matter in real projects: a lack of common sense, trouble with ambiguity and uncertainty, overfitting and underfitting, and confusing correlation with causation. This post is for product and business teams deciding where ML can work on its own and where it needs human oversight.

Are you basing your decisions on these 5 machine-learning myths? See what’s what in our guide!

Why Does Machine Learning Lack Common Sense?

Machine learning lacks common sense because models learn statistical patterns from training data, not from lived experience of the world. Humans pick up common sense through everyday life. A model trained for image recognition or language translation only has the examples in its dataset, so it has no wider picture of the context those examples come from.

Some failures that look like missing common sense have more specific technical causes. Language models have been known to struggle with tasks such as counting the 'r's in 'strawberry'. Research has linked these letter-counting errors to tokenization: models process text as chunks of words (tokens), so they do not work through a word letter by letter the way a person does.

Autonomous vehicles show the problem in a higher-stakes setting. Autonomous driving systems can be challenged by rare edge cases that are poorly represented in their training data, such as a pedestrian in an unusual costume crossing the street. When a situation falls outside what the model has seen, it has no everyday experience to fall back on.

Even a high score on a commonsense benchmark does not guarantee reliable judgment in unfamiliar real-world situations.

How Does Machine Learning Handle Ambiguity and Uncertainty?

Machine learning handles ambiguity well only when similar cases appear in its training data. Humans can make reasonable decisions with vague or incomplete information by drawing on experience. A model's output depends on the data distribution it was trained on, so missing or unusual inputs reduce its reliability, especially in decisions where the outcomes are not clear-cut.

Language is a good example. Language models can still misread sarcasm, tone or meaning that depends on context. A phrase like "It's raining cats and dogs" is the kind of input that simpler, keyword-based natural language processing (NLP) systems took literally, while current large language models usually handle common idioms. The harder cases are the ones where the intended meaning depends on who is speaking and why.

In highly sensitive fields such as medicine or finance, uncertainty is a common problem. Models can assist in making predictions based on data, but they struggle in situations where there are no clear patterns to follow or when data is missing.

What Are Overfitting and Underfitting in Machine Learning?

Machine learning models face two common problems during training: overfitting and underfitting.

Overfitting occurs when a model learns the training data too well, capturing noise or irrelevant details. This leads to poor performance on new, unseen data because the model has become too specific to the training dataset. Imagine studying for an exam by memorizing every word in the textbook instead of understanding the key concepts-you’d likely struggle when the questions aren’t exactly what you expected.

Underfitting, on the other hand, happens when the model doesn’t learn enough from the training data, often because it's too simple or hasn’t been trained long enough. In this case, it performs poorly both on the training data and when tested on new data.

Both problems highlight the delicate balance machine learning developers must strike. Too much focus on the training data results in overfitting, while too little results in underfitting, making models ineffective in real-world applications.

Why Can't Machine Learning Tell Causation from Correlation?

Machine learning has difficulty telling causation from correlation because models learn patterns in data, and a pattern alone does not show whether one event causes another. ML models are excellent at finding relationships between variables, but they often cannot determine whether those variables are causally linked or simply correlated.

For instance, if an ML model analyzes a dataset showing that people who own more than two pets are healthier, it might assume that owning pets causes better health. However, humans can intuitively understand that other factors, such as exercise or lifestyle, might be involved and that the relationship is more complicated.

In healthcare, this gap carries real risk. A model trained on patient records can learn associations that reflect how patients were treated instead of what caused their outcomes. Its predictions need review by clinicians before they inform care decisions.

Standard predictive models learn associations from data and do not test cause and effect, which matters in medicine, finance and economics.

How Can We Overcome the Limitations of Machine Learning?

Researchers are working on several approaches that target these gaps directly.

Causal inference covers methods that estimate cause-and-effect relationships instead of associations alone. These methods help answer questions such as whether a treatment improved outcomes or whether the patients who received it were already healthier.

Explainable AI covers techniques that show which inputs drove a prediction, so people can check whether the model relied on sensible signals or on a spurious pattern.

Neuro-symbolic AI combines machine learning with symbolic reasoning, the rule-based logic used in mathematics and formal proofs. Google DeepMind's AlphaGeometry (2024) is one example. It pairs a neural language model with a symbolic deduction engine, and on a benchmark of 30 Olympiad geometry problems it solved 25 within the standard time limit.

For organizations applying these methods, machine learning data services can help.

These approaches are still developing, so human judgment remains necessary when deciding how and where ML predictions are used.

What the Limitations of Machine Learning Mean in Practice

For teams adopting ML, these limits turn into design decisions. Human review belongs wherever errors are costly, as in healthcare and finance. Models should be tested on data that differs from the training set before they are trusted in production. Predictions are best treated as associations unless causal methods were used to produce them. Teams that plan for this from the start, with software development services built around these constraints, get the efficiency of ML while keeping human insight where it is still irreplaceable.

If your team is planning an ML feature, talk to Arbisoft about building a system designed around these limits from day one.