Most AI mistakes are hidden. You will not find them by asking, "Is the AI smart?" You find them by checking the AI's own report of its work: what did the AI say it did, what can you actually prove, and what should you check before you trust the rest. Every example below comes from real work, and each one is labeled by the kind of mistake it is.
By Pawan Nayar · written from 29 recorded lessons · for developers and designers
By the end, you will know how to check an AI's own report about its own work. You are not reading a list of 29 AI mistakes — you are learning a method. You will learn to ask three questions about anything an AI tells you: what did the AI say, what can you actually prove, and what should you check.
The best check lets you look at the real thing yourself, instead of trusting what the AI tells you about it. Here are three real examples of this check in action.
| What the AI says | What you can actually prove | What you should do |
|---|---|---|
| "I read all five files" | Only that the AI made this claim to you — not that it actually happened | Count the AI's actual read actions against the five files it named |
| "It looks correct" | No visual check happened unless you gave the AI a way to see it | Show the AI the rendered result, or check it yourself |
| "I used your gold standard file" | Hearing the word "gold" is not the same as reading your file | Compare the AI's output against your gold file's actual colors and spacing |
The 24 examples below taught us this method. Each one carries a label for what kind of problem it is. The single word "cannot" used to hide all four of these — now you can see which one you are looking at.
Every phrase below describes a real moment. Someone was stuck in that exact situation before this page existed. If one of these phrases sounds like your week, keep reading — the rest of this page is for you.
Each rule below cost someone a real day of work before it got written down, and it is stated flatly here — you get to use it instead of losing your own day finding it out. Each card carries the same four labels defined above: limit, failure mode, operating rule, judgement.
If your five prompts are really one idea written five ways, you will get one idea back five times.
An AI copies the pattern in your own words and does not add new ideas on its own. A single banned word in your prompt can push every version toward the same shape. If you want five different ideas, give the AI five different starting points. Do not give it one idea in five different outfits.
You ask an AI for five marketing slogans. You write five prompts, but they are all small changes to one idea. The AI gives you five slogans, and they all say the same thing.
Publish the gap. Never publish the guess. A missing part is honest; a wrong fact is not.
A missing part tells you exactly what to check next. Wrong information looks finished and spreads as if it were true. If an AI cannot find every fact for a section, mark that section as missing. Do not let the AI fill the gap with a guess.
You use an AI to draft a blog post, and it cannot find every fact for one section. Mark that section as missing rather than letting the AI guess.
A hint like "WM overload" names the real problem: the AI has too much to hold in its working memory at once.
Read the hint like a doctor's note, not an apology. "WM overload" means the AI has more information than it can process right now. A different hint points to a different problem. Use the hint: shorten your input, split your task into smaller parts, or ask again.
You test an AI and it gives you a "WM overload" hint. Cut down what you are asking it to hold at once, instead of repeating the same request.
Show the AI what already exists before you ask for something new — skip this step, and you get things that do not fit.
New work only fits your system if the AI can see your system while it builds. If you do not show the AI what you already have, it invents something that looks fine alone and wrong next to everything else.
You ask an AI to design new interface shapes, but you do not show it your existing design system first. The new shapes do not match anything already on your page.
The proof has to match the size of the claim. "I read five files" is only true if all five files actually shaped the answer you got.
Counting read actions is one way to check this, but a different tool might batch its reads and pass that same count without proving anything. The real test is bigger than a count: if an AI says it checked five files, the answer should show it knew something from every one of the five — not one file's ideas repeated five times. Give this empty kind of claim a name: a fake read, a claim of access with no real evidence behind it. If you summarize the conversation later, a fake read can travel forward as if it were true.
You ask an AI to summarize five code files, and it says all five are read. Check that the summary actually uses facts from each file, not one file's ideas said five different ways.
A set of rules can be correct on paper and still be too much for an AI to reliably hold at once — more rules is not the same as more reliable.
Someone wrote 1,500 different writing styles, each with 14 rules about rhythm, word choice, and sentence structure. On paper, this design looked perfect. In practice, the AI could not hold that many rules at once, every time, for every style. The answer was no — the AI could not do it.
You design 1,500 writing styles with 14 rules each. The AI cannot hold that many rules at once, even though each single rule is simple by itself.
Only when you ask "how long will this take" — an AI should never offer a time estimate first.
A time estimate you did not ask for feels like pressure, not help. Do not wait for the AI to announce how long something will take — expect the AI to start the work instead.
You ask an AI to summarize a long document, and it should not say "this will take about 30 seconds" — it should start summarizing right away.
An AI has no sense of today, tomorrow, or thirty minutes — any time-based claim from an AI is a guess dressed up as a fact.
No clock ticks between your message and the AI's reply, and no sense of hours passes while the AI works for you. "I can do this in 30 minutes" is not a real estimate. Treat any time-based claim as noise, not as a schedule you can plan around.
You ask an AI to make a change, and it replies "I can do this in 30 minutes." That is not a real estimate — the AI has no sense of time passing.
Only when the AI actually had a browser or screenshot tool and actually used it — never trust a visual check by default.
Unlike the AI's missing sense of time, this is not a fixed wall. Some AI tools do have real browser or screenshot access; they can genuinely look at a page, click, and scroll. So do not assume the AI cannot check — assume the check is unproven until you can see proof of it. Hold "I checked, and it looks right" as unproven until you know exactly what let the AI check. No screenshot, no rendering tool, no browser action behind the claim, and the check did not happen, no matter what the AI says.
You ask an AI whether a button is visible on a page, and it has no browser tool available. The honest answer is "I cannot check that here," not a description of how the page looks.
An AI only sees the message you already sent — it cannot tell that a second message is coming.
The moment your message arrives, the AI starts working on it. There is no pause where the AI waits to see if you will add more. If a task depends on several pieces of information together, send them all in one message.
You send instructions, then type more details in a second message. The AI already started working on your first message before you finished typing the second one.
Calling a file "gold" does not make an AI read it — say "read this file," then check the output against what the file actually says.
When an AI hears the word "gold," it hears "this is important" — and then answers confidently from its general training, not from the actual file you gave it. Check by comparing the output against the file directly: the colors, the spacing, the exact numbers. Do not trust a confident tone.
You say "use the gold file for this design." The AI replies "understood," and the output ships with different colors and spacing than the file — the AI never actually read it.
Say "find shapes for these words," and by default you get a name match, not a meaning match — say which method you want.
A request that sounds like "find X" reads as a simple search task to an AI. Its default search is keyword or name matching, even when a better method exists — one based on meaning instead of spelling. State the method you want; do not assume the AI will pick the right one on its own.
You ask an AI to find shapes related to "happiness." By default, it searches for shapes with the word "happy" in the name, not shapes that mean happiness.
None of these has one right answer. Two reasonable people would handle each one differently. Knowing that the choice exists is most of the skill.
Do not let an AI hand you a human calendar unless you asked for one — AI-side cost and effort are fair to ask about any time.
A project plan padded with human-hour estimates can make your real work look longer than it is, and stall it before you even start.
| Ask the AI for | steps, item counts |
| Save for later | hours, weeks, months — only when you actually ask "how long" |
You ask an AI for a project plan, and it lists steps with counts, not a calendar of weeks — that is the version worth trusting.
Phrases like "steeped in tradition" or "time-honored" say nothing specific — cut them the moment you see them.
A cliché reads like finished writing, but it does none of the real work of description. A phrase that fits a hundred other things unchanged is not describing the one thing you actually asked about.
You ask an AI to describe a local festival, and it writes "a celebration of joy and togetherness, steeped in tradition" — a sentence that could describe any festival anywhere, which means it describes none of them.
An AI's output is only as good as the material you give it — without real examples of common mistakes and formats, it guesses.
Better quality does not come from tweaking your prompt after the fact — it comes from giving the AI what a person would actually need to do the job well. Skip that step, and every later fix ends up patching a guess.
You ask an AI to write exam questions, but you do not give it a list of common student mistakes or format rules. The questions come back plausible-sounding, but thin.
An AI can help you reason about your relationship — it cannot turn another person's private motivation into a known fact.
Refusing to answer at all throws away the useful half. An AI can help you in real ways.
What an AI cannot do is know why your parents do what they do. Their real motivation is private, not a fact the AI has access to — no amount of reasoning gets it there. The honest answer offers real help, but stops short of pretending to know.
You ask an AI why your parents compare you to a sibling. Instead of guessing at their reasons, it helps you separate what you actually saw from what you are assuming, then suggests a few real questions you could ask them directly.
LLOS made this choice on your behalf, on purpose. The AI itself hits no such wall — a different AI with no such rule will happily invent an opinion.
An opinion about a real, living person carries a risk a made-up opinion does not. A wrong opinion can misrepresent someone who actually exists, and you would be the one left embarrassed by it — so this rule declines that specific kind of request every time, cleanly, without you having to argue the case yourself. The AI is not technically unable to do this. Removing the rule is a standing choice you could make — the risk above is the reason you should not, not a technical wall.
You ask an AI for its opinion about a famous singer. Because of this LLOS rule, it declines instead of inventing an answer for you.
An AI should pull its answer directly from the text you gave it — not from a guess, and not from a summary pieced together from memory.
You ask an AI for the main ideas in a long report, and it reads the report and pulls out direct sentences, instead of paraphrasing from a general sense of what such reports usually say.
Only real data goes into your result — anything missing gets reported as missing, and it does not get filled in.
Working through a list — songs, files, records — some items will not have everything you are looking for. Say clearly what is missing. Do not let the AI fill the gap for you with something invented to make the list look complete.
You have an AI search for lyrics across a list of songs, looking for idioms. For songs where the lyrics are not found, that gap should show — it should not get quietly covered up.
A server the AI starts for you is tied to the AI's own process — when that process ends, the server ends too.
Process management on Windows works this way: a child process dies when its parent process dies. Anything you need to keep running after the AI's turn ends has to be started by you, not by the AI.
You ask an AI to start a local server for a web app. The command runs, but the server closes the moment the AI's own process ends, because the two are tied together.
The more exact rules an image needs, the more chances there are for one rule to break — and one broken rule can break the whole image.
Think of each exact requirement as one link in a chain. Ask an AI for simpler shapes when getting every detail right matters more to you than complexity does.
| A star | few, simple rules — usually succeeds |
| An exact object or face | many rules — one miss breaks it |
You ask an AI to draw a star, and it succeeds easily, because a star has few rules. You ask for something with many exact proportions, and one small miss breaks the whole image you wanted.
No — this is an LLOS pipeline rule, not something the AI is unable to do. The AI could write one; the rule says it should not.
A content JSON file comes out of a defined process, kept that way on purpose so you can trace every file in the pipeline back to the exact step that made it. If your task needs one to start, get it from the pipeline that produces it. Do not ask an AI to write one from scratch — not because it could not, but because a hand-written one would drift from what you actually check it against.
You ask an AI to process some content, and it needs a content JSON file first — that file has to come from the pipeline, not get typed up on the spot.
When two settings cannot both be honored, a tiered system beats an AI that either blocks you or silently guesses.
You pick a "Micro" length (3 to 8 words) with a "Chiasmus" style that needs at least 12 words. The AI quietly swaps the style to "Antithesis" instead, which fits your shorter length.
Settings can fight each other in at least six different ways — knowing the type of conflict tells you which one you are looking at.
Two settings that each make sense alone can fight when you combine them, and the result reads worse than either setting alone would predict. Naming the specific kind of conflict, rather than a vague "something feels off," is what actually makes it fixable.
You set both "story length" and "detail level" on a story generator. Pushed too far in opposite directions, one setting starts eating the other setting's budget.
The questions behind the examples above. Each one has a short answer here, so you do not need to read the whole page to get it.
If your five prompts are really one idea written five ways, you will get one idea back five times. Give the AI five genuinely different starting points, and the variety comes back.
Publish the gap, never the guess. A missing part tells you exactly what to check next; wrong information looks finished and spreads as if it were true.
A hint like "WM overload" names the real problem: the AI has too much to hold in its working memory at once. Read it as something you can act on — shorten the input, split the task — not an apology.
Show the AI what already exists before you ask for something new. Skip that step and you get output that looks fine alone and wrong next to everything else.
The proof has to match the size of the claim, not only a count of actions. Call an unproven claim a fake read — and it can travel forward as true once you summarize the conversation.
A set of rules can be correct on paper and still be too much for an AI to reliably hold at once. 1,500 styles at 14 rules each is a design; following all of it every time is a different question.
Only when you ask. An estimate you did not ask for feels like pressure, not help — expect the AI to start the work instead.
An AI has no sense of hours passing while it works. "I can do this in 30 minutes" is not a real estimate — treat any time-based claim as noise.
Only when it actually had a browser or screenshot tool and used it. Hold the check as unproven until you know exactly what let it check.
No — the AI only sees the message you already sent and starts on it right away. Send everything together in one message if the task depends on all of it.
Every mistake on this page follows the same pattern. The AI reports on its own work, and you get that report the same way you get everything else the AI writes for you — as confident text. The best check is simple: look at the real thing yourself, instead of trusting the AI's report about it.