{"course":{"slug":"ai-for-students","title":"AI for Students","url":"https://addaly.com/learn/ai-for-students","level":"beginner","track":"work"},"lesson":{"slug":"follow-ups-that-find-the-weak-joint","title":"Follow-ups that find the weak joint","url":"https://addaly.com/learn/ai-for-students/follow-ups-that-find-the-weak-joint","position":13,"of":73,"minutes":8,"answer":"Replace \"are you sure\" with follow-ups that produce checkable material (what would make this wrong, argue the opposite, what would a marker criticise) and test stability by re-asking cold in a fresh chat.","keyIdea":"Replace \"are you sure\" with follow-ups that produce checkable material (what would make this wrong, argue the opposite, what would a marker criticise) and test stability by re-asking cold in a fresh chat."},"body":{"format":"markdown","text":"## Stop asking whether it is sure\n\n*Are you sure?* is the most common follow-up and one of the least useful. It is another prompt, and the model has no lookup to consult. Worse, the tuning that makes assistants agreeable means a challenge often produces a retraction whether or not the original was wrong. You can talk a correct model out of a correct answer, which should tell you exactly how much information the retraction carries.\n\nSo replace it. The follow-ups below are the ones that actually move you forward, because each asks for something that can be checked outside the conversation or that changes the shape of the answer rather than its confidence.\n\n## The five that earn their place\n\n**1. What would make this wrong?**\n\n> What would have to be true for this answer to be incorrect? What is the strongest case against it?\n\nThis converts an assertion into a claim with stated conditions. It surfaces the assumptions and it works in every subject. In an essay it gives you your counterargument paragraph; in a lab report it gives you your limitations section.\n\n**2. Which part of this are you least confident about?**\n\nNot a real confidence report, for reasons already covered. But it reliably points at the parts that are unusual, contested or specific, because those are what surrounding text hedges about. Treat the answer as a list of things to verify, not as a probability.\n\n**3. Argue the opposite, as strongly as you can.**\n\nThe best single move for essay subjects. Ask for the strongest version of the position you disagree with, written by somebody who believes it. Then look at whether your own argument survives it. If the opposing case is easy to write and hard to answer, you have picked the wrong thesis, and it is much better to find that out now.\n\n**4. What would a marker in this subject criticise?**\n\nBetter with the mark scheme pasted in, but useful even without. Framing the model as an examiner rather than a helper changes the output substantially, because it shifts from completion to critique.\n\n**5. What am I not asking that I should be?**\n\nWorks surprisingly well when you are new to a topic and do not know its shape. It surfaces the standard next questions in a field. Verify anything specific that comes back, as always.\n\n## The rephrase-and-repeat check\n\nWhen a specific answer matters, do this:\n\n1. Open a new chat.\n2. Ask the same question in different words, without pasting the earlier answer.\n3. Compare.\n\nStable answers across cold rephrasings are more likely to reflect well-represented knowledge. Divergent answers mean you need a real source. It costs one message and it is the closest thing to a free reliability test you have.\n\nThe reason it must be a new chat: everything in the current context conditions the next token strongly. A model that has just written an answer will tend to be consistent with it. Consistency inside a conversation measures almost nothing.\n\n## Asking it to disagree with itself\n\nA structured variation worth knowing:\n\n> Here is an answer to a problem. Act as a sceptical examiner. Find the weakest step. Do not be polite about it.\n\nPaste the previous answer into a *fresh* conversation rather than continuing the old one. Stripped of the context in which it produced the answer, the model critiques much more freely. This is not magic and it will sometimes invent objections, but the objections are cheap to evaluate and it catches real errors often enough to be worth the message.\n\n## The follow-up that is actually about you\n\nThe last one is not addressed to the model:\n\n> Can I now explain this without the window open?\n\nIf yes, the session worked. If no, you have a fluent answer and no new capability, and the right move is to close it and try to reconstruct from memory before reading again. Everything in Module 3 is built on that question.\n\n## What to do when it will not budge\n\nSometimes a model repeats the same wrong thing however you ask. This usually means the error is well-represented in its training data: a common misconception, an outdated standard, a widely-copied wrong answer. That is a signal, not a nuisance. Go and find out why the wrong version is so common; it is often the exact thing your marker is testing whether you know."},"check":{"question":"Why must the rephrase-and-repeat reliability check be done in a new conversation rather than further down the same one?","options":[{"text":"Because a long conversation eventually exceeds the context window and the original question is truncated away.","correct":false,"tempting":"A real effect at length, but the check fails immediately in a short chat too, so truncation is not the reason."},{"text":"Because the earlier answer sits in the context and heavily conditions what follows, so consistency proves little.","correct":true},{"text":"Because models apply a lower temperature to later messages in a conversation, reducing variation artificially.","correct":false,"tempting":"Invents a mechanism; sampling settings do not change as a conversation proceeds."},{"text":"Because a repeated question in the same chat is treated as a challenge, and the tuning makes the model retract rather than repeat.","correct":false,"tempting":"Describes the sycophancy problem, which is real, but it predicts change rather than the consistency that actually defeats the check."}]},"prev":{"slug":"ask-for-the-working","title":"Ask for the working, then check it","url":"https://addaly.com/learn/ai-for-students/ask-for-the-working"},"next":{"slug":"explain-it-at-three-levels","title":"Explain it at three levels","url":"https://addaly.com/learn/ai-for-students/explain-it-at-three-levels"},"authorship":"Founder and series editor: Dr Mustafa Mun. Written by AI models to a written standard. Exam keys checked blind by a second model: 3,665 of 3,666 agreed. Academic review invited.","review":{"status":"Draft, open for academic review"},"cite":"Addaly (2026) 'Follow-ups that find the weak joint', in AI for Students. Edited by M. Mun. Addaly. Available at: https://addaly.com/learn/ai-for-students/follow-ups-that-find-the-weak-joint (Accessed: 13 October 2026).","licence":{"name":"CC BY-NC-SA 4.0","url":"https://creativecommons.org/licenses/by-nc-sa/4.0/"},"reader":{"schema":1,"title":"Follow-ups that find the weak joint","courseTitle":"AI for Students","module":{"index":2,"slug":"asking-well","title":"Asking well"},"answer":null,"keyIdea":"Replace “are you sure” with follow-ups that produce checkable material (what would make this wrong, argue the opposite, what would a marker criticise) and test stability by re-asking cold in a fresh chat.","hero":{"kind":"plate","image":null,"alt":null,"caption":null},"blocks":[{"t":"h2","id":"stop-asking-whether-it-is-sure","runs":[{"text":"Stop asking whether it is sure"}]},{"t":"p","runs":[{"text":"Are you sure?","i":true},{"text":" is the most common follow-up and one of the least useful. It is another prompt, and the model has no lookup to consult. Worse, the tuning that makes assistants agreeable means a challenge often produces a retraction whether or not the original was wrong. You can talk a correct model out of a correct answer, which should tell you exactly how much information the retraction carries."}]},{"t":"p","runs":[{"text":"So replace it. The follow-ups below are the ones that actually move you forward, because each asks for something that can be checked outside the conversation or that changes the shape of the answer rather than its confidence."}]},{"t":"h2","id":"the-five-that-earn-their-place","runs":[{"text":"The five that earn their place"}]},{"t":"p","runs":[{"text":"1. What would make this wrong?","b":true}]},{"t":"quote","runs":[{"text":"What would have to be true for this answer to be incorrect? What is the strongest case against it?"}]},{"t":"p","runs":[{"text":"This converts an assertion into a claim with stated conditions. It surfaces the assumptions and it works in every subject. In an essay it gives you your counterargument paragraph; in a lab report it gives you your limitations section."}]},{"t":"p","runs":[{"text":"2. Which part of this are you least confident about?","b":true}]},{"t":"p","runs":[{"text":"Not a real confidence report, for reasons already covered. But it reliably points at the parts that are unusual, contested or specific, because those are what surrounding text hedges about. Treat the answer as a list of things to verify, not as a probability."}]},{"t":"p","runs":[{"text":"3. Argue the opposite, as strongly as you can.","b":true}]},{"t":"p","runs":[{"text":"The best single move for essay subjects. Ask for the strongest version of the position you disagree with, written by somebody who believes it. Then look at whether your own argument survives it. If the opposing case is easy to write and hard to answer, you have picked the wrong thesis, and it is much better to find that out now."}]},{"t":"p","runs":[{"text":"4. What would a marker in this subject criticise?","b":true}]},{"t":"p","runs":[{"text":"Better with the mark scheme pasted in, but useful even without. Framing the model as an examiner rather than a helper changes the output substantially, because it shifts from completion to critique."}]},{"t":"p","runs":[{"text":"5. What am I not asking that I should be?","b":true}]},{"t":"p","runs":[{"text":"Works surprisingly well when you are new to a topic and do not know its shape. It surfaces the standard next questions in a field. Verify anything specific that comes back, as always."}]},{"t":"h2","id":"the-rephrase-and-repeat-check","runs":[{"text":"The rephrase-and-repeat check"}]},{"t":"p","runs":[{"text":"When a specific answer matters, do this:"}]},{"t":"ol","start":1,"items":[[{"text":"Open a new chat."}],[{"text":"Ask the same question in different words, without pasting the earlier answer."}],[{"text":"Compare."}]]},{"t":"p","runs":[{"text":"Stable answers across cold rephrasings are more likely to reflect well-represented knowledge. Divergent answers mean you need a real source. It costs one message and it is the closest thing to a free reliability test you have."}]},{"t":"p","runs":[{"text":"The reason it must be a new chat: everything in the current context conditions the next token strongly. A model that has just written an answer will tend to be consistent with it. Consistency inside a conversation measures almost nothing."}]},{"t":"h2","id":"asking-it-to-disagree-with-itself","runs":[{"text":"Asking it to disagree with itself"}]},{"t":"p","runs":[{"text":"A structured variation worth knowing:"}]},{"t":"quote","runs":[{"text":"Here is an answer to a problem. Act as a sceptical examiner. Find the weakest step. Do not be polite about it."}]},{"t":"p","runs":[{"text":"Paste the previous answer into a "},{"text":"fresh","i":true},{"text":" conversation rather than continuing the old one. Stripped of the context in which it produced the answer, the model critiques much more freely. This is not magic and it will sometimes invent objections, but the objections are cheap to evaluate and it catches real errors often enough to be worth the message."}]},{"t":"h2","id":"the-follow-up-that-is-actually-about-you","runs":[{"text":"The follow-up that is actually about you"}]},{"t":"p","runs":[{"text":"The last one is not addressed to the model:"}]},{"t":"quote","runs":[{"text":"Can I now explain this without the window open?"}]},{"t":"p","runs":[{"text":"If yes, the session worked. If no, you have a fluent answer and no new capability, and the right move is to close it and try to reconstruct from memory before reading again. Everything in Module 3 is built on that question."}]},{"t":"h2","id":"what-to-do-when-it-will-not-budge","runs":[{"text":"What to do when it will not budge"}]},{"t":"p","runs":[{"text":"Sometimes a model repeats the same wrong thing however you ask. This usually means the error is well-represented in its training data: a common misconception, an outdated standard, a widely-copied wrong answer. That is a signal, not a nuisance. Go and find out why the wrong version is so common; it is often the exact thing your marker is testing whether you know."}]}],"check":{"question":[{"text":"Why must the rephrase-and-repeat reliability check be done in a new conversation rather than further down the same one?"}],"options":[{"text":[{"text":"Because a long conversation eventually exceeds the context window and the original question is truncated away."}],"correct":false,"tempting":[{"text":"A real effect at length, but the check fails immediately in a short chat too, so truncation is not the reason."}]},{"text":[{"text":"Because the earlier answer sits in the context and heavily conditions what follows, so consistency proves little."}],"correct":true,"tempting":null},{"text":[{"text":"Because models apply a lower temperature to later messages in a conversation, reducing variation artificially."}],"correct":false,"tempting":[{"text":"Invents a mechanism; sampling settings do not change as a conversation proceeds."}]},{"text":[{"text":"Because a repeated question in the same chat is treated as a challenge, and the tuning makes the model retract rather than repeat."}],"correct":false,"tempting":[{"text":"Describes the sycophancy problem, which is real, but it predicts change rather than the consistency that actually defeats the check."}]}],"order":[0,1,2,3],"why":null},"prev":{"slug":"ask-for-the-working","title":"Ask for the working, then check it","minutes":8},"next":{"slug":"explain-it-at-three-levels","title":"Explain it at three levels","minutes":8}}}