Internet Is Fun 
Most writing about the limits of AI is either dismissal or advertising. What is more useful is a working list of where these systems fail in practice β the failures you will actually hit, rather than the philosophical ones.
One caveat up front, and it is a real one: this is a fast-moving target. Specific weaknesses get fixed, sometimes abruptly. What follows is written as of September 2026, and the honest instruction is to test rather than to trust any list, including this one.
Knowing what it does not know
This is the big one and it underlies most of the others. These systems produce fluent output regardless of whether they have the information, and the confidence of the answer is a poor guide to its reliability. A fabricated citation reads exactly like a real one.
The practical consequence: anything checkable should be checked, and the things most worth checking are names, numbers, dates, quotations and references β precisely the details that make an answer feel authoritative.
Counting and exact arithmetic
Text models are poor at operations over exact symbols: counting characters, long multiplication, keeping a running tally. This is a consequence of how text is represented internally rather than a bug to be reasoned around.
It is largely solved by tools β a model that can run code will do arithmetic properly, because the code does it. If a system is producing numbers without computing them, treat them as estimates.
Long-horizon consistency
Over a long task β a big document, an extended piece of work, a multi-step plan β small inconsistencies accumulate. A decision made early gets quietly contradicted later. Nothing flags the contradiction, because each individual step looked locally sensible.
The workaround is structural: keep the requirements written down somewhere outside the conversation, and check the output against them rather than against your memory of what was agreed.
Anything that needs to be current
A model’s own knowledge has a cutoff, and it will answer questions about the period after it with whatever seems most plausible. Systems with search attached handle this much better, but then you have inherited the quality of whatever they found β which, given the state of search results, is not nothing to worry about. See AI slop.
Taste, and knowing when to stop
Ask for feedback and you will usually get improvements. Ask repeatedly and you will get changes, because producing changes is the task. Knowing that a thing is finished, or that the third version was better than the fifth, is not well represented in the training signal.
This is the one that catches people out most, because it looks like progress the entire time it is happening.
The image side
Image models have improved fastest, and the classic complaints β hands, text β are much less reliable as criticisms than they were. What has persisted: keeping a specific character consistent across many images, rendering a precise diagram where the relationships carry the meaning, and following a long list of spatial constraints without dropping one.
The general pattern is the same as in text. They are excellent at producing something of the right shape and unreliable at producing something exactly right.
What this implies about how to use them
The systems are strong where being approximately right is useful and being wrong is cheap to notice: drafting, summarising, explaining, translating, generating options, getting unstuck. They are weak where an error is expensive and invisible.
That is the whole heuristic, and it is more durable than any specific list of weaknesses. Before relying on an output, ask what happens if it is subtly wrong and nobody notices. If the answer is bad, verify it.
For sorting the useful tools from the demos, see is this AI tool useful or just a demo, and what open source AI actually means for the terminology.
