The test that makes this list useful
Most articles on this topic are a list of things AI got wrong, written at a moment in time. Within a year, half the list is obsolete and the reader has no way to tell which half. So apply one test to every claimed limitation: did it survive the last two model generations?
Context length, arithmetic, long-form code, and multi-step reasoning were all described as fundamental limits of language models within the past three years. All four moved, some dramatically. They were capability gaps wearing the word "limitation". The four below have not moved, because they are properties of the approach rather than the size of the model.
1. No ground truth
A language model is trained on text, and text contains true and false statements in similar grammatical clothing. The model learns the distribution, not the truth value. This is why a model can produce a citation that is perfectly formatted, plausibly titled, attributed to a real author in the right field, and entirely invented. The output is doing exactly what it was trained to do.
The practical consequence: a model cannot check its own work. Asking it to verify its answer produces another generation from the same distribution, which is why "are you sure?" is not a control. The fix is architectural: an external source of truth the model reads from rather than recalls.
2. No persistent memory
Each session starts from nothing. Anything the model appears to remember was put back into its context by software outside it. This gets called a limitation of AI when it is closer to a property of the interface, but it has a real cost: every long-running process needs someone to decide what gets carried forward, and that decision is where most of the engineering actually lives.
3. No calibrated confidence
This is the expensive one, and it is the least discussed. Models produce fluent, assured output whether the answer is solid or fabricated. There is no reliable internal signal that says this one is shaky.
Consider two systems. One is wrong 5% of the time and tells you which 5%. The other is wrong 2% of the time with uniform confidence. The first is more deployable, because you can route its uncertain cases to a human and trust the rest. The second requires you to check everything, which erases most of the value. Accuracy is the number people quote; calibration is the number that determines whether you can ship.
The research here is more nuanced than the flat claim above, and the nuance is the useful part. Anthropic's Language Models (Mostly) Know What They Know found that larger models are in fact well calibrated on multiple-choice and true/false questions, a constrained format where the model is scoring a fixed set of options. The same work found calibration degrading on new tasks. That split matches what I see in production: confidence is usable where the output space is enumerable and unreliable where it is open-ended, which is exactly where most real work sits.
4. No accountability
A model cannot be answerable for an outcome. When an AI-produced answer causes a loss, the responsibility sits with whoever deployed it, in exactly the proportion it would have without the model. This is the only item on the list with no technical roadmap (no release notes will ever address it) and it is the one that decides where AI belongs in a process rather than how well it performs there.
This is why every serious governance framework treats accountability as an organisational property rather than a system property. NIST's AI Risk Management Framework (AI 100-1, 2023) assigns trustworthiness characteristics to the people and processes around the system, not to the model. That is not a drafting convention. It is the only place accountability can sit.
Limitation versus risk
Worth separating, because they have different owners. A limitation is something the technology cannot do. A risk is what happens when you deploy it regardless. Hallucination is a limitation; a customer acting on a hallucinated refund policy is a risk. Limitations are handled in architecture. Risks are handled in process. Conflating them is why so many AI governance documents read as technical complaints.
Where this gets harder
Everything above concerns a single model call. Agents chain them, and chaining changes the maths: a small misinterpretation early becomes a confident wrong action later, with no exception thrown on the way. I have written about how that compounds across handoffs. If you are deciding whether to build something rather than trying to understand the technology, the more useful page is when not to use AI agents.
In this cluster
- LLM limitations — the same question at the model layer, for a technical reader
- When not to use AI agents — the decision version of this page
- AI agents vs automation — the most common wrong call
Frequently asked questions
What are the main limitations of AI?
Four survive every model release. AI has no ground truth. It models the distribution of its training text, not reality, so it cannot tell a true statement from a plausible one without an external check. It has no persistent memory across sessions unless something outside the model supplies it. It has no calibrated confidence, so it sounds equally certain when right and wrong. And it carries no accountability, so a human remains answerable for every output. Capability limits move with each release; these four have not.
What can AI not do?
It cannot verify its own output, because verification requires a source of truth it does not have. It cannot reliably tell you when it does not know, because uncertainty is not represented in a form that survives the generation process. It cannot carry responsibility for a decision. And it cannot guarantee that the same input produces the same output, which rules it out anywhere determinism is required.
Are the limitations of AI permanent?
Some are, some are not, and most published lists do not distinguish. Context length, arithmetic and code generation were all cited as fundamental limitations within the past three years and have all moved substantially — those were capability gaps. No ground truth, no calibrated confidence and no accountability are properties of the approach rather than the current models, and none has moved. The useful test is whether a limitation survived the last two model generations.
What is the difference between an AI limitation and an AI risk?
A limitation is something the technology cannot do. A risk is something that happens when you deploy it anyway. Hallucination is a limitation; a customer acting on a hallucinated policy is a risk. The distinction matters because they have different owners — limitations are handled in architecture, risks are handled in governance and process.
Do these limitations apply to AI agents too?
They apply more, because agents compound them. A single wrong answer is a wrong answer; a wrong answer at step two of a nine-step chain becomes a confident wrong action at step nine, with no exception thrown along the way. Everything on this page is the input to a harder question, which is when an agent is the wrong tool entirely.