{"course":{"slug":"ai-at-work","title":"AI at Work","url":"https://addaly.com/learn/ai-at-work","level":"beginner","track":"work"},"lesson":{"slug":"the-shape-of-the-work","title":"What it is actually good at","url":"https://addaly.com/learn/ai-at-work/the-shape-of-the-work","position":1,"of":73,"minutes":8,"answer":"If you did not supply the facts, treat every fact it returns as unverified: fluency is not evidence.","keyIdea":"If you did not supply the facts, treat every fact it returns as unverified: fluency is not evidence."},"body":{"format":"markdown","text":"Most advice about AI at work is either \"it will replace you\" or \"it makes things up, ignore it.\" Both are useless when you have a job to do on Tuesday. You need a sharper question: which parts of my work does this help with, and which parts does it quietly damage?\n\nHere is the most useful distinction I know. It is not about topic. It is about where the truth lives.\n\n## Truth in the text vs truth in the world\n\nWhen the answer is **inside the text you provided**, the model is strong. Summarise this eight-page circular. Rewrite this complaint letter so it is firm but not rude. Pull every deadline out of this contract. Turn these rough notes into a paragraph. The material is right there. The model is reorganising, compressing, rephrasing. It has everything it needs.\n\nWhen the answer lives **out in the world**, the model is guessing from patterns. What is the current stamp duty rate in my state? Did this supplier actually deliver in March? What was the exact wording of the 2019 amendment? What is my patient's potassium level? None of that is in the conversation. The model will still answer, fluently and in full sentences, because producing plausible text is what it does. That fluency is the danger. A wrong answer does not arrive looking wrong.\n\nSo the first sorting rule: **if you did not give it the facts, do not trust the facts it gives back.**\n\n> Figure: Where the truth lives. In the text you supplied: Summarise this eight-page circular; Rewrite this complaint so it is firm, not rude; Pull every deadline out of this contract; Turn these rough notes into a paragraph; Sort sixty survey answers into eight themes. Out in the world: The current stamp duty rate in my state; Whether the supplier delivered in March; The wording of the 2019 amendment; This patient's potassium level; What my manager said in the corridor. Same model, same fluent sentences, opposite reliability. On the left it is reorganising material it can see; on the right it is guessing, and the guess arrives in the identical register.\n\n## Where it earns its keep\n\nAcross very different jobs, the same shapes of task keep paying off:\n\n- **The blank page.** A first draft of anything. A parent letter, a tender response, a policy note, a product description. Drafts are cheap to fix and expensive to start.\n- **Compression.** Forty pages down to one. A two-hour meeting down to seven decisions.\n- **Translation between registers.** Legalese into plain language for a client. Clinical notes into something a family understands. English into Spanish, Hindi, Yoruba, Tagalog: good, not perfect, and better if you can read the output.\n- **Structure from mess.** Sixty open-ended survey answers sorted into eight themes. A pile of receipts turned into a table.\n- **The thing you can nearly do.** A spreadsheet formula, a regex, a SQL query, a bit of Excel VBA. You cannot write it, but you can test whether it worked.\n\n## Where it will hurt you\n\n- **Anything where being 95% right is a failure.** Dosages. Tax filings. Court deadlines. Structural loads. Not \"never use it\": never use it *unchecked*.\n- **Recent and local specifics.** Last month's circular from your ministry. Your state's fee schedule. Your company's leave policy. Unless you paste it in.\n- **Counting and exact arithmetic across long documents.** \"How many invoices are over 50,000 naira\" is a spreadsheet question, not a chat question.\n- **Judgment that is yours to make.** Should we fire this person. Should this student repeat the year. Is this client lying. It will give you an answer. That answer carries no accountability, and yours does.\n\n## Do this today\n\nWrite down the five things you did last week that took over thirty minutes and were mostly typing or mostly reading. That list is your real curriculum. Most people find two or three of them are pure text-shuffling, and those are the ones to hand over first."},"check":{"question":"A district health officer needs two things done: condense a 60-page WHO guidance PDF she has open into a one-page briefing, and confirm the current national reimbursement rate for a particular vaccine. Which task is the riskier one to hand to a chatbot?","options":[{"text":"The reimbursement rate, because the number is not in anything she gave the model, so a confident answer may be invented","correct":true},{"text":"The 60-page summary, because long documents are where models lose accuracy and start hallucinating","correct":false,"tempting":"Believes length itself causes hallucination, when the real risk factor is whether the source material was supplied at all."},{"text":"Both equally, since the model has no way to verify anything it says in either case","correct":false,"tempting":"Treats all model output as uniformly unreliable, which throws away the genuine difference between reorganising given text and recalling world facts."},{"text":"The reimbursement rate, because health policy is a regulated domain and models are restricted from answering it","correct":false,"tempting":"Attributes the risk to subject-matter sensitivity rather than to the absence of source material."}]},"prev":null,"next":{"slug":"what-it-is-doing-when-it-answers","title":"What it is actually doing when it answers you","url":"https://addaly.com/learn/ai-at-work/what-it-is-doing-when-it-answers"},"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) 'What it is actually good at', in AI at Work. Edited by M. Mun. Addaly. Available at: https://addaly.com/learn/ai-at-work/the-shape-of-the-work (Accessed: 31 October 2026).","licence":{"name":"CC BY-NC-SA 4.0","url":"https://creativecommons.org/licenses/by-nc-sa/4.0/"},"reader":{"schema":1,"title":"What it is actually good at","courseTitle":"AI at Work","module":{"index":1,"slug":"deciding-what-to-hand-over","title":"Deciding what to hand over"},"answer":null,"keyIdea":"If you did not supply the facts, treat every fact it returns as unverified: fluency is not evidence.","hero":{"kind":"figure","figure":{"t":"figure","label":"Figure 1.1","figure":{"left":{"title":"In the text you supplied","items":["Summarise this eight-page circular","Rewrite this complaint so it is firm, not rude","Pull every deadline out of this contract","Turn these rough notes into a paragraph","Sort sixty survey answers into eight themes"]},"type":"compare","right":{"title":"Out in the world","items":["The current stamp duty rate in my state","Whether the supplier delivered in March","The wording of the 2019 amendment","This patient’s potassium level","What my manager said in the corridor"]},"title":"Where the truth lives","caption":"Same model, same fluent sentences, opposite reliability. On the left it is reorganising material it can see; on the right it is guessing, and the guess arrives in the identical register."},"alt":"Figure: Where the truth lives. In the text you supplied: Summarise this eight-page circular; Rewrite this complaint so it is firm, not rude; Pull every deadline out of this contract; Turn these rough notes into a paragraph; Sort sixty survey answers into eight themes. Out in the world: The current stamp duty rate in my state; Whether the supplier delivered in March; The wording of the 2019 amendment; This patient’s potassium level; What my manager said in the corridor. Same model, same fluent sentences, opposite reliability. On the left it is reorganising material it can see; on the right it is guessing, and the guess arrives in the identical register."}},"blocks":[{"t":"p","runs":[{"text":"Most advice about AI at work is either “it will replace you” or “it makes things up, ignore it.” Both are useless when you have a job to do on Tuesday. You need a sharper question: which parts of my work does this help with, and which parts does it quietly damage?"}],"lede":true},{"t":"p","runs":[{"text":"Here is the most useful distinction I know. It is not about topic. It is about where the truth lives."}]},{"t":"h2","id":"truth-in-the-text-vs-truth-in-the-world","runs":[{"text":"Truth in the text vs truth in the world"}]},{"t":"p","runs":[{"text":"When the answer is "},{"text":"inside the text you provided","b":true},{"text":", the model is strong. Summarise this eight-page circular. Rewrite this complaint letter so it is firm but not rude. Pull every deadline out of this contract. Turn these rough notes into a paragraph. The material is right there. The model is reorganising, compressing, rephrasing. It has everything it needs."}]},{"t":"p","runs":[{"text":"When the answer lives "},{"text":"out in the world","b":true},{"text":", the model is guessing from patterns. What is the current stamp duty rate in my state? Did this supplier actually deliver in March? What was the exact wording of the 2019 amendment? What is my patient’s potassium level? None of that is in the conversation. The model will still answer, fluently and in full sentences, because producing plausible text is what it does. That fluency is the danger. A wrong answer does not arrive looking wrong."}]},{"t":"p","runs":[{"text":"So the first sorting rule: "},{"text":"if you did not give it the facts, do not trust the facts it gives back.","b":true}]},{"t":"h2","id":"where-it-earns-its-keep","runs":[{"text":"Where it earns its keep"}]},{"t":"p","runs":[{"text":"Across very different jobs, the same shapes of task keep paying off:"}]},{"t":"ul","items":[[{"text":"The blank page.","b":true},{"text":" A first draft of anything. A parent letter, a tender response, a policy note, a product description. Drafts are cheap to fix and expensive to start."}],[{"text":"Compression.","b":true},{"text":" Forty pages down to one. A two-hour meeting down to seven decisions."}],[{"text":"Translation between registers.","b":true},{"text":" Legalese into plain language for a client. Clinical notes into something a family understands. English into Spanish, Hindi, Yoruba, Tagalog: good, not perfect, and better if you can read the output."}],[{"text":"Structure from mess.","b":true},{"text":" Sixty open-ended survey answers sorted into eight themes. A pile of receipts turned into a table."}],[{"text":"The thing you can nearly do.","b":true},{"text":" A spreadsheet formula, a regex, a SQL query, a bit of Excel VBA. You cannot write it, but you can test whether it worked."}]]},{"t":"h2","id":"where-it-will-hurt-you","runs":[{"text":"Where it will hurt you"}]},{"t":"ul","items":[[{"text":"Anything where being 95% right is a failure.","b":true},{"text":" Dosages. Tax filings. Court deadlines. Structural loads. Not “never use it”: never use it "},{"text":"unchecked","i":true},{"text":"."}],[{"text":"Recent and local specifics.","b":true},{"text":" Last month’s circular from your ministry. Your state’s fee schedule. Your company’s leave policy. Unless you paste it in."}],[{"text":"Counting and exact arithmetic across long documents.","b":true},{"text":" “How many invoices are over 50,000 naira” is a spreadsheet question, not a chat question."}],[{"text":"Judgment that is yours to make.","b":true},{"text":" Should we fire this person. Should this student repeat the year. Is this client lying. It will give you an answer. That answer carries no accountability, and yours does."}]]},{"t":"h2","id":"do-this-today","runs":[{"text":"Do this today"}],"role":"try"},{"t":"p","runs":[{"text":"Write down the five things you did last week that took over thirty minutes and were mostly typing or mostly reading. That list is your real curriculum. Most people find two or three of them are pure text-shuffling, and those are the ones to hand over first."}]}],"check":{"question":[{"text":"A district health officer needs two things done: condense a 60-page WHO guidance PDF she has open into a one-page briefing, and confirm the current national reimbursement rate for a particular vaccine. Which task is the riskier one to hand to a chatbot?"}],"options":[{"text":[{"text":"The reimbursement rate, because the number is not in anything she gave the model, so a confident answer may be invented"}],"correct":true,"tempting":null},{"text":[{"text":"The 60-page summary, because long documents are where models lose accuracy and start hallucinating"}],"correct":false,"tempting":[{"text":"Believes length itself causes hallucination, when the real risk factor is whether the source material was supplied at all."}]},{"text":[{"text":"Both equally, since the model has no way to verify anything it says in either case"}],"correct":false,"tempting":[{"text":"Treats all model output as uniformly unreliable, which throws away the genuine difference between reorganising given text and recalling world facts."}]},{"text":[{"text":"The reimbursement rate, because health policy is a regulated domain and models are restricted from answering it"}],"correct":false,"tempting":[{"text":"Attributes the risk to subject-matter sensitivity rather than to the absence of source material."}]}],"order":[2,0,3,1],"why":null},"prev":null,"next":{"slug":"what-it-is-doing-when-it-answers","title":"What it is actually doing when it answers you","minutes":10}}}