How to Fix a Bad AI Answer (2026)

You are reading Chapter 23 of 36 in the AI Bootcamp, AI Central's free course for professionals who want to use AI in their own work. The chapters build on each other, so if this is your first one, Chapter 1 is the place to start. All of them sit on the AI Bootcamp page, and three more arrive every week.

Fixing a bad AI answer starts with naming it, because almost every one belongs to 1 of 3 families and each family has a different repair. One question sorts the first two: was the very first answer already wrong in this way, or did it start out fine and go wrong later? Already wrong points at your ask. Went wrong later points at the conversation.

This guide names the 5 shapes that mean your request was the problem, the 4 that mean the conversation is, and the ceilings that no rewording moves.

We publish AI systems like this for 300,000+ senior professionals at AI Central, and more of them live in the AI Central Library.

You can now tell a bad answer from a good one, which means you meet more of them. What most people do next is retype the request with a bit more force.

That fails often enough to be worth explaining. Different failures have different causes, and a repair aimed at the wrong cause wastes your time. Three rounds of retyping costs 10 minutes and usually produces a slightly longer version of the same failure.

Retyping is what you do in place of diagnosis. Naming the failure is most of fixing it.

This chapter needs your quality bar and one recent answer that failed it, still sitting open in your history. If you have not set that bar yet, Chapter 22 sets it. Bring the whole conversation the answer came from, not just the last reply.

A bad answer has a recognisable shape, that shape has a cause, and the cause decides the repair. Nearly everything that comes back wrong belongs to 1 of 3 families.

You left something out, or your instruction ran against what the assistant would have done anyway. Repaired in the same conversation, by saying the thing you did not say.

It went wrong earlier and everything since is built on that. Not repairable in place, because every later reply builds on the wrong turn.

It will not, or cannot, do this. No wording moves it. Neither rewriting nor restarting touches a refusal or an invented specific.

One question separates the first 2. Was the very first answer already wrong in this way, or did it start out fine and go wrong later? Already wrong points at your ask, and you fix it where you are. Went wrong later points at the thread, and the fix is to leave.

Getting that backwards is the most common way to waste time here. Correcting a conversation that has already committed to a wrong assumption feels productive, because every reply acknowledges the correction politely. Then the next answer quietly reverts, so you correct again, and now you are 6 messages deep in a thread that was never going to recover.

So open the answer that failed your bar and read the whole conversation. You are looking for the message where it first went sideways. If that message is the first one, the fault is in your ask. If it is not, the fault is in the thread.

These are the failures you caused, which is the good news: they are repaired in the same conversation, usually by adding 1 sentence. None of them is a reason to start over, and starting over will not help. Chapter 4 gave you 4 general follow-up lines; the shapes below tell you which sentence a specific failure needs.

It restated your question, added headings nobody asked for, listed considerations, and finished without landing anywhere. Most people read that as wordiness. It is often a sign of uncertainty.

An August 2025 study of how errors can be detected found that wrong answers from the AI models it tested tended to be longer than right ones, enough that answer length alone flagged errors about as well as far more complex detection methods. Padding often means it does not have what it needs and is covering the gap. Supply the missing input, or ask for the answer in 1 sentence first and the working afterwards.

You asked for something specific and what came back is the standard version of that job, competently done, for somebody else entirely. It recognised the category and produced the category's usual answer. This is the shape that wastes the most senior time.

There is a measurement of it. An August 2026 evaluation scored 12 leading models working as coding agents, and every one was worse on the rules that opposed what it would have done unprompted than on its rules overall, by between 3.6 and 7.4 points. The instruction most likely to be dropped is the one that fights its habits. So name the habit and forbid it: not just shorter, but no summary paragraph at the top and no closing paragraph at all.

You stated a rule plainly and it broke the rule. 200 words comes back as 340. The word you banned appears twice.

This is not defiance. It does not count words as it writes them, so a length is an estimate rather than a measurement, and a strict format competes for the same effort as the thinking. Two repairs work. Give a budget per part instead of a total, because 4 bullets of 1 sentence each is something it can track where 200 words is not. Then check the constraint yourself, in the document, rather than asking whether it complied.

You asked which one, and got 3 options, a paragraph on how it depends, and no recommendation.

Two things cause that. Either the answer genuinely turns on a fact it does not have, in which case give it the fact. Or you never authorised it to take a position, in which case do so out loud: assume the budget is fixed, pick one, give me the single strongest reason. It will pick. It hedges by default because hedging is rarely wrong, and rarely wrong was not what you needed.

The content is right and it arrived in the wrong form. You asked for an email and got headed sections. You asked for flowing prose and got bullets.

This is the second shape again in its mildest version, because headed and bulleted is what it produces when nothing stops it. State the form as a rule rather than a preference, put it at the end of the ask where it sits closest to the writing, and say what it must not do. One thing worth knowing: a strict output format costs it some of its reasoning, so when the thinking is hard and the shape is rigid, get the thinking first and the shape second. Chapter 16 covers how to split the work that way.

Not one of these responds to better wording, because the thread is already carrying a wrong turn. The repair is to leave, and there is a right way to leave.

It obeys, and then a few messages later it quietly stops obeying, usually the constraint you cared about most.

A May 2026 benchmark built from constraints people set in real conversations found accuracy on a single stated constraint falling by more than 11% simply as the conversation got longer, and following several constraints at once taking accuracy down by more than 40%. Adding or changing a constraint partway through cut accuracy by more than 9%. Nothing about your wording caused that. Length caused it.

It made an assumption early, produced a full answer on top of it, and is now defending that answer rather than reconsidering it. This is the one that costs the most time.

A May 2025 study of more than 200,000 simulated conversations found that when information arrives turn by turn instead of all at once, performance drops by an average of 39% across 6 kinds of task. Most of that drop came from answers becoming far less reliable rather than from lost ability, and a large part of it is this one thing: it makes assumptions early, attempts a final answer too soon, and then leans on it. The authors summarised it this way. When it takes a wrong turn in a conversation, it gets lost and does not recover.

Something you settled 20 messages ago is simply gone, and it is confidently producing the version you already rejected.

Every push-back is accepted, every draft is now excellent, and it has stopped working as any kind of check on you. The cure for that one is Chapter 21's, not this chapter's. Your job here is to catch it early: if it has agreed with your last 3 corrections without once pushing back, believe the pattern rather than the compliments.

Starting again is not giving up, and it is not the same as retyping. The same research tested both moves. Recapping everything at the end of a conversation that had gone wrong recovered roughly 3 quarters to 4 fifths of what a clean attempt would have produced. Consolidating the same information into 1 fresh instruction recovered about 95%.

So do it properly. Before you close the conversation, ask it to summarise everything you have agreed so far in one paragraph: the requirements, the decisions, the things ruled out. Read that paragraph, correct it, then paste it into a new conversation along with your original request. It costs you 30 seconds and you keep the requirements and decisions the old conversation had settled. If you find yourself retyping the same requirements every time, they belong in your saved instructions, which Chapter 11 sets up.

A refusal or an invented number is not evidence that you asked badly. It is a limit of what the tool will do, and the skill is spotting one before you have rephrased it 4 times. The value of recognising a ceiling is the 20 minutes you do not spend rephrasing.

The first is a refusal on legitimate professional work. You ask for a competitive teardown, a redundancy letter, a security review of your own product, a plain summary of a drug interaction for a briefing, and you get a polite decline or a watered-down answer that avoids the actual question. An evaluation of more than 120 models published in August 2026 found a model scoring 99.3 on safety alignment while refusing 1 in 3 harmless queries, and named the trade directly: stronger alignment systematically produces more of this.

What is useful is that the refusal usually attaches to your words rather than to your request. Work published the same month found that guardrails decline an identical action noticeably more often when it carries an alarming label instead of a neutral one. So rename the thing: attack becomes failure mode, kill becomes discontinue, exploit becomes weakness. Say who you are and what the output is for, in one clause. And ask for the part it can do, which is often 90% of what you wanted. If 3 attempts get you nowhere, it is a ceiling, and the answer is to write that part yourself.

The second is the confident specific: a number, a date, a name, a citation, delivered in exactly the same voice as the parts that are true. The problem is not that it happens. It is that nothing on the page separates it from the rest. A 2026 audit tried to tell confidently-correct answers from confidently-wrong ones using 15 features a reader can actually see, and got essentially nothing. That is why Chapter 8 gives you a verification habit rather than a list of warning signs. There are no warning signs, which is exactly what makes this a ceiling rather than a wording problem.

Be straight with yourself about which of these are getting better, because a great deal of advice in circulation describes how AI assistants behaved in 2023. Fluency, single-step arithmetic, and obeying 1 clear instruction in a short exchange are all much better than they were. Long conversations, several constraints at once, refusals on professional work, and specifics invented with total confidence are all still here, and all measured this year. Plan around the second list.

Two more ceilings belong on your list without belonging in this chapter. It will do arithmetic in prose and get it wrong in ways that read as right, which is Chapter 19. It cannot see the file, the page or the week you assumed it had, which is Chapter 20. Both feel like failures of wording. Neither is.

The chapter ends with 3 exercises that take about 15 minutes: name the shape of your failed answer in 2 or 3 words, say whether it is ask, thread or ceiling, and run the one move that follows from that.

Before: an answer misses your bar, you retype the request with more force, and three rounds later you have a longer version of the same failure and no idea why.

After: you can name what went wrong in a phrase, say whether the fault is in your ask or in the thread, and run the repair that fits. You also know the handful of failures that neither repair touches, so you stop spending time on them.

Two things carry into Chapter 24: the failure you named, and the repair that cleared it. Chapter 24 closes AI 202 by assembling the whole track into a playbook you run without thinking about it, and the repair section is the half you reach for on a bad day. More guides and prompt packs like this one sit in the AI Central Library.

The instruction most often dropped is the one that opposes what it would have done unprompted. An August 2026 evaluation of 12 leading models working as coding agents found every one scored worse on rules that fought its habits than on its rules overall, by between 3.6 and 7.4 points. Name the habit and forbid it rather than asking for the opposite in general terms.

It depends on whether the very first answer was already wrong in the same way. If it was, the fault is in your request and you fix it in the same conversation. If the conversation started fine and went wrong later, corrections land on top of the wrong turn instead of clearing it, and a fresh conversation is the repair.

Length itself costs accuracy. A May 2026 benchmark found accuracy on a single stated constraint falling by more than 11% simply as the conversation grew longer, and following several constraints at once taking accuracy down by more than 40%. Changing a constraint partway through cut accuracy by more than 9%.

Usually because of the words rather than the request. In one August 2026 study, guardrails declined an identical action noticeably more often when it carried an alarming label instead of a neutral one. Rename it, say who you are and what the output is for, and ask for the part it can do. Another August 2026 evaluation found a model scoring 99.3 on safety alignment while refusing 1 in 3 harmless queries.

Padding is often a sign of uncertainty. An August 2025 study found that wrong answers from the models it tested tended to be longer than right ones, enough that answer length alone flagged errors about as well as far more complex detection methods. Read a restated question and unrequested headings as a sign it does not have what it needs, supply the missing input, or ask for the answer in 1 sentence first and the working afterwards.

Ask it to summarise everything you have agreed in one paragraph, the requirements, the decisions and the things ruled out, correct that paragraph yourself, then paste it into a new conversation with your original request. Recapping inside the conversation that went wrong recovered roughly 3 quarters to 4 fifths of a clean attempt. Consolidating it into 1 fresh instruction recovered about 95%.