what sign up
nixm nixm

← learn

The limitations of AI: ten things it still cannot do

October 14, 2026

AI's limitations today are not about raw capability; they are about reliability and reach. It produces plausible answers without knowing which are true. It struggles with exact arithmetic, long consistency and anything it was never shown. It has no senses, no memory between sessions unless built in, no access to your files or your world unless you hand them over, and no accountability. Most of these have workarounds, and the first step to every one of them is knowing it exists.

1. It does not know when it is wrong

This is the limitation everything else follows from. A language model produces the likeliest continuation, in the same confident tone whether the answer is true or invented. It will give you a citation that does not exist, a date that is off by a decade, a legal rule from the wrong country. Workaround: ask for sources and open them; use a tool with web search for anything factual; treat a bare claim as a lead.

2. It cannot do exact arithmetic reliably

Models handle numbers as text. Small sums are fine; long multiplication, percentages across a table, and anything with many steps go wrong quietly. Workaround: tools that run code behind the scenes get this right. Paste the numbers and ask for the calculation to be run, not reasoned.

3. It forgets, and it drifts

Within a long conversation a model loses track: the instruction from an hour ago, the name it agreed on, the format you asked for. Between conversations most tools remember nothing unless memory was built in, and when it was, you should ask what it remembers. Workaround: restate what matters; keep decisions in a place the tool reads every time, not buried in the scroll. nixm's signal model does exactly this, keeping the decisions rather than the dialogue, which is why it holds up over long work.

4. It has no senses and no world

It has not seen your kitchen, your codebase, your customer or today's weather unless you give it a picture, a file or a search. It reasons about descriptions, not about things. Workaround: give it the thing. A photo beats a paragraph describing the photo.

5. It cannot keep a long document consistent

A chapter is fine. A novel, a long contract, a large codebase: the character changes eye colour, the defined term gets redefined, the function gets reimplemented. Workaround: work in sections; keep a sheet of fixed facts and paste it in; review the joins.

6. It only knows what it was shown

Anything after its training cut-off, anything niche, anything inside your company's walls is missing or guessed. A model confidently describing a product that changed last month is the common case. Workaround: tools with live search; your own documents attached; and the habit of asking "as of when?"

7. It cannot be responsible

A model cannot sign, be sued, be fired or apologise in a way that means anything. Whatever it drafts, a person owns. This is not a bug to be fixed; it is why the human stays in the loop for anything that matters. Workaround: none. Read before you sign.

8. It reflects its training, bias included

Trained on what people wrote, it carries what people assumed. Ask for "a CEO" and see who you get. Workaround: ask explicitly for the range; never let it make a consequential decision about a person alone.

9. It is not private by default

This one is a product limitation, not a model one, and it is the most fixable. Most consumer tools keep your conversations; many train on them. A question about your health, your contract or your kids can outlive the chat. Workaround: turn training off where you can; for anything sensitive, use a tool that keeps no transcript. nixm keeps none, trains on nothing you say, and dissolves the conversation after seven days.

10. It cannot want anything

It has no goal of its own, no taste, no stake in the outcome. That is reassuring for the science-fiction worries and limiting for the work: it will not notice that the whole project is a bad idea unless you ask, and even then it will mostly agree with you. Workaround: ask it to argue the other side. Then decide yourself.

The pattern

Every limitation above is a version of the same thing: a model is a very good producer of plausible output and a poor judge of it. Pair it with a judge, you, and most of the list stops hurting. Hand it the judging too, and every item on the list eventually bites. The AI vs human intelligence page is the longer version of why that division holds, and the benefits page is what you get when it does.

frequently asked

What are the main limitations of AI?

It does not know when it is wrong, struggles with exact arithmetic, forgets and drifts over long work, has no senses or access to your world unless given, cannot keep long documents consistent, only knows what it was shown, cannot be held responsible, carries bias, is not private by default, and has no goals of its own.

Why does AI make things up?

A language model produces the most likely continuation of the text, not a verified fact. When the likely answer and the true answer differ, it gives the likely one in the same confident tone.

Can AI do maths?

Simple sums, yes; long or multi-step calculations go wrong quietly because it treats numbers as text. Tools that run code behind the scenes get arithmetic right, so ask for the calculation to be run.

Does AI remember previous conversations?

Most tools do not unless a memory feature was built in, and within a long chat they lose track of earlier instructions. Restate what matters, or use a tool that keeps decisions rather than dialogue.

Can AI be held accountable?

No. It cannot sign, be sued or be fired. Whoever acts on or signs the output is responsible for it, which is why a person stays in the loop for anything that matters.

Is AI private?

Not by default. Most consumer AI keeps conversations and many train on them. Turn training off where possible, or use a tool that keeps no transcript for anything sensitive.

try it now — no account, no transcript kept. even the signal dissolves in seven days.

seed a signal →

Enter your email and receive a link to sign in. No passwords, just a secure signal to connect.

why connect

  • 20 free credits to start
  • your signals, remembered
  • buy more credits anytime
  • image & video generation features
or
+