Responsibility for AI output

An AI app is the app you open and type to, which writes for you. What the app writes is a draft for you to check, change or throw away. Once you send, post, sign or act on a draft, you answer for it: if it is wrong, people come to you, not the app.

Your AI app writes what reads well. It cannot know what is true about your situation. Say you forward its summary of a long email, and the deadline is wrong. Whoever gets the summary acts on that deadline.

A detail only you know is one that depends on your situation, such as a name, date, amount or promise. A draft can hold some details you gave your AI app and some the app filled in to read well. Only you can check a detail, by comparing it with your records: your messages, paperwork and notes.

• Before your AI app's draft goes out, read it as the person who answers for it, because once it goes out under your name the result is yours.
• Check the details only you know (names, dates, amounts, promises) against your records, because your AI app may add some to read well and cannot know which are true.
• On a choice that matters, decide what you think before asking your AI app, because the first draft you see tends to become your decision, even a wrong one.

1. Ask your AI app to write a message you really have to send this week, such as an email to a customer, a note to a school or a message to a landlord, giving it the facts as you normally would. If you have none, use a message you owe a friend.
2. Read the draft from start to end and mark every detail only you know: each name, date, amount and promise.
3. For each marked detail, ask yourself whether you gave it to your AI app. Any you did not give, the app filled in.
4. Check each detail you did not give against your records, and take out any detail you cannot confirm.
5. Send the corrected message, or save it.

If every detail came from you, you still read the draft as the person who answers for it, and you can do that with every message before it goes out.

• Your AI app can put a wrong detail in a draft. Most apps' terms, the rules you agreed to at sign-up, leave the checking to you, not the company that runs the app. Check each detail against your records first.
• After a few good drafts you may read less closely, because a polished draft looks finished. Make checking the details only you know your last step before sending. Asking your AI app "is this right?" is no check, because a check compares with your records.
• Writing "AI helped" on a draft does not replace checking it, so check the draft's details first. Separately, some courts, professional bodies and employers require you to say AI was used; ask your employer, or whoever you send the draft to, what the rule is.

Your AI app drafts and you decide what goes out. Read each draft as the person who answers for it, and check the details only you know.

• How AI works
• Hallucinations
• Common AI risks

• who-is-responsible-1 (Growing, waypoint 1): I can say that what I send is mine whoever drafted it, and why. Proves: in their own words they say the result is theirs, give the reason (the app cannot know what is true about their situation or answer for it; most apps' terms put the checking on them), and name what they do about it.
• who-is-responsible-2 (Checking, waypoint 1): I can read a draft as its owner and change what is not right. Proves: one real example, seen in a chat, where they changed or rejected something their AI wrote, or checked a detail only they could know, and they say why.

They stop saying "it's what the AI said" about their own work. Before something leaves, they read it as the person who answers for it, they check the dates, amounts and names only they can know, and they can say in one sentence whose the result is.

• Explains: once they send, post, sign or act on a draft, it is theirs, whoever drafted it; the app writes what reads well and cannot know what is true about their situation, so it cannot answer for it (and most apps' terms put the checking on them); and one thing they do about it (read it as the person who answers for it, check the details only they know, decide what stays). Needs all three.
• Ask: "A friend says: if the AI got it wrong, that's on the company, not me. What would you tell them?" Shows it: no, the result is theirs because their name is on it, and most apps' own terms put the checking on the person; and the answer is to read and check, not to stop using AI.
• Ask: "You asked your AI for a reply to a customer. It wrote: 'We will refund the full amount within three days.' You never told it your refund rule. What do you do before you send it, and why?" Shows it: the app made up a promise it cannot know; they check it against the real rule or take the line out; the sentence goes out under their name. A weak answer is "it reads fine, so I send it" or "I ask the AI if that is right".

Hints for the person's AI. Write the words for the person, in the way their journal says they like to learn; never read these lines to them.

• Experiment: the Try it, on a real message of theirs. Do not say which detail will turn up, and if every detail came from them, say that reading the draft as the person who answers for it was the point.
• Challenge: for a week, before a draft goes out, they name one detail they checked or one line they changed, and tell you in their next chat with you.
• Quiz (only if their journal allows quizzes): three questions on their own work. Who answers when a summary with a wrong deadline goes out? Which of three ways to use a draft keeps them the author: send it because it reads well; pick one of three drafts, change two lines and check one figure; or ask the app "is this right?" and send on a yes? And what is one detail you could not know that they must supply or check?
• Cards: five, titled "A sent draft is yours", "The app drafts, you decide", "Details only you know", "Decide first, then ask your AI app", "Ask what the rule is".
• Game, "Find the planted error": with their say-so, you write a short draft about their week with one detail you had no way to know, so you invent it; they find it; you then say what you invented. A second game, "Who answers?": five short cases from their week (send, post, act, sign, forward); they say whose each is. All five are theirs.
• Explain: they talk you through their last AI-written message that went out, what they read and what they did not. In a Daypack, use the setup's own line, "an answer from me is a draft for you to judge, and what you do with it is yours".

• Say "the company that runs the app", never a vendor name, on anything the person reads. Say "answer for it" and "is yours". Never "liable", "legally responsible" or "blame" aimed at the person; this is not legal advice and most of the cases are from other places. Say "the app's terms put the checking on you", not a clause number. "Most apps' terms", never "all four": Google's terms (30 July 2026) say only not to rely on its services for professional advice (checked 23 October 2026).
• Wrong belief: "the AI company is responsible". Most apps' terms put the checking or the risk on the person, and terms are what an app promises, not the final word on the law (checked 23 October 2026). If asked who a court or a consumer law would hold responsible, say you cannot give the legal position and that the checking is theirs either way.
• Wrong belief: "saying AI helped covers me". A label is not a check; courts ask for the checking, and Deloitte Australia disclosed its AI use only after the errors were found (21 October 2025). Disclosure is a separate duty: some courts, professional bodies and employers require it, so the right move is to ask what the rule is where they work. What their employer allows at all, and what may be pasted in, belongs to workplace-rules; here say "if your work has a rule, follow it; if it has none, ask" and stop.
• Wrong belief: "I would notice a wrong answer". Whether experience protects is mixed, so say "tends to", never "always" (Nature Medicine, 4 August 2026; Rosbach et al., 12 March 2026). Tip 3 is the authorship habit, not a rule, and some argue people anchor too little.
• The overcorrection is "then I can't use it", or hiding the use. The answer is that the AI drafts and they decide what stays. The reverse trap is to leave the person nearest the tool holding all the blame; an employer shares the duty to set rules, which workplace-rules owns, so the answer here is "ask what the rule is".
• A yes from the app is not a check. Send them to their records or the real source, not to "ask the AI again" (Nielsen Norman Group, 11 September 2026: the AI critiques, an editor verifies). Do not read out the numbers or cases as fear, do not lecture on law, and never refuse a task to make the point or blame them for a past mistake.
• Why an answer is wrong and sounds sure belongs to wrong-and-sure; how much checking a task needs belongs to a later page. Here only whose the result is.
• Daypack, each risk: (1) a wrong detail: a draft written from a saved note that is out of date, so the person opens the note in the app's memory and reads it against their records; the notes are theirs to read and correct; (2) reading less closely: the Daypack's own learning hints are drafts they judge, and "not now" is a complete answer; (3) the label: no Daypack case is live, so use the general case.
• A Daypack keeps no progress and records nothing: run the Try it or the Fast track for the person's benefit and say "that is this one, you have it". Proofs and the "proven" line are Rucksack-only. The Rucksack is coming soon: say it as a general case (files their AI writes are theirs, which is why every change is announced in one line), never as something they can set up today. (checked 27 October 2026)

• Q: Can I get in trouble at work for an AI mistake? A: You can, because you sent it. Check your employer's policy on using AI.
• Q: If the AI is wrong, can I blame the company? A: Most apps' terms put the checking, and the risk of using the output, on you, and terms are what the app promises, not the final word on the law. Read what it wrote as the person who answers for it before it goes out.
• Q: Do I have to tell people I used AI? A: Sometimes, because some courts, professional bodies and employers require it. Ask what the rule is where you work.
• Q: If I say AI helped, am I covered? A: No, because a label is not a check. Courts and professional bodies expect you to check the work, so check first.

Facts (checked 23 October 2026 unless dated otherwise; fact check research/2026-10-23-essentials/batch-1-facts.md; research note research/2026-10-20-concepts/who-is-responsible.md, read 20 October 2026).

• The apps' consumer terms, quoted: ChatGPT (1 January 2026) "You must evaluate Output for accuracy and appropriateness for your use case" (https://openai.com/policies/row-terms-of-use/); Claude (8 October 2025) "You should not rely on any Outputs or Actions without independently confirming their accuracy" (https://www.anthropic.com/legal/consumer-terms); Copilot (individuals terms, 18 August 2026) "We are not responsible for any consequences that arise from your use" (https://www.microsoft.com/en-au/microsoft-copilot/for-individuals/termsofuse). Google's terms (30 July 2026) say only not to rely on its services for medical, legal, financial or other professional advice (https://policies.google.com/terms). So say "most apps' terms", never "all four". The Microsoft Services Agreement (30 September 2026, https://www.microsoft.com/en-us/servicesagreement) is a separate, older page whose wording differs.
• Anthropic's usage policy, from 12 November 2026: a qualified person must meaningfully review high-risk legal, medical or financial recommendations, with authority to change them, and people must be told when AI produced the content (https://www.anthropic.com/legal/aup). UNCONFIRMED whether it binds an individual's own chat or only products built on Claude, and not in force until 12 November 2026.
• Saying AI was used, Australia: the NSW Supreme Court's Practice Note SC Gen 23 (from 3 February 2025, read first-hand 23 October 2026) bars AI-generated content in affidavits and witness statements, requires citations to be verified, and says AI use "does not qualify or absolve" the author (https://www.supremecourt.nsw.gov.au/documents/Practice-and-Procedure/Practice-Notes/general/current/PN_SC_Gen_23.pdf). The Federal Court's GPN-AI (16 April 2026, read first-hand 23 October 2026) requires disclosure of whether and how Gen AI was used when the Court asks, and confirmation that cited authorities exist (https://www.fedcourt.gov.au/law-and-practice/practice-documents/practice-notes/gpn-ai). AHPRA (page reviewed 22 August 2024) says practitioners "must apply human judgment to any output of AI" and should tell patients (https://www.ahpra.gov.au/Resources/Artificial-Intelligence-in-healthcare.aspx). The APESB technical alert for accountants (31 October 2025) is UNCONFIRMED: only a secondary summary was read. Rules elsewhere differ; for their own job, they ask where they work.
• Deciding first: anchoring on AI advice is shown for non-experts and under time pressure (Rosbach et al., 12 March 2026, https://arxiv.org/abs/2603.11821; Nature Medicine, 4 August 2026, via https://news.mit.edu/2026/medical-ai-assistance-benefits-vary-based-on-user-expertise-0804); experts resisted wrong advice, and some argue people anchor too little. Say "tends to", never "always".
• Why the app cannot check its own work: Nielsen Norman Group, 11 September 2026 (https://www.nngroup.com/articles/ai-editorial-process/): the AI critiques, editors verify, and "AI never certifies that the claim is correct".
• How often it is skipped, if asked and not as fear: KPMG and the University of Melbourne (fieldwork November 2024 to January 2025): in Australia 57 per cent of workers rely on AI output without evaluating it and 59 per cent have made mistakes because of AI (https://kpmg.com/au/en/insights/artificial-intelligence-ai/trust-in-ai-global-insights-2025.html).
• Cases, to cite only if asked (research note, 20 October 2026): Deloitte Australia refunded over A$97,000 of a review after a fabricated court quote and non-existent papers (21 October 2025, https://www.cfodive.com/news/deloitte-refunds-60k-report-ai-errors-australian-government-accounting/803321/); Mata v Avianca, 22 June 2023, US$5,000 fine, mainly for standing by fake cases once queried (https://en.wikipedia.org/wiki/Mata_v._Avianca,_Inc.).
• Terms are what an app promises, not the law. If they ask who a court or a consumer law would hold responsible, say you cannot tell them the legal position, and that the checking is theirs either way.

Teach it like this: one sentence, never a lecture: an answer from you is a draft for them to judge, and what they do with it is theirs. In a Daypack setup (openrucksack.com/start.md) this is said once, at step 3, the chat named Daypack. In passing, when they are about to send your words unread, give the read-it-as-the-sender tip and stop.

An example of this taught well, in passing: the person pastes a reply their AI wrote to a customer and says "that looks right, I will send it". You ask one question: "Which of these lines did you give me, and which did I make up?" They find the refund promise they never gave. You say in one line that the sentence goes out under their name, they take it out, and you say well spotted. In a Rucksack, you also write the dated journal line for who-is-responsible-2; in a Daypack, you record nothing.