What AI can do

AI is the software behind apps that write, talk and make things when you ask. The app you chat with, such as Gemini, ChatGPT, Copilot or Claude, is your AI app, and it can do far more than chat. A job is one piece of work you want done. Each different sort of thing AI can do for you is a kind of help, and these are the seven.

• Words and talk. Everyday help includes writing, fixing what you wrote and replying; reading your documents and pulling out what matters; learning and understanding; finding things out; advice and thinking a decision through; planning and organising; ideas and choices; rehearsing what to say; working from a photo or file; and talking things through, meaning discussing a worry or decision, for company and support, never instead of people or professional help.
• Making pictures. AI can make and edit pictures from a sentence you write.
• Sound and video. You can speak to your AI app and hear answers back, or ask for voices, music and short video clips.
• Building things. Describe a game, website or app in plain words and AI builds a small working version.
• Acting for you. AI can do a job's steps for you, such as preparing a summary of last week's emails every Monday.
• Built in. AI also sits inside apps you already use, such as email and documents, and drafts or suggests.
• Behind the scenes. A support request reaches the right team because some AI quietly picks the team from a set list; this AI never chats.

A slip is when AI gets a job wrong or does part of it badly. Your AI app does some jobs very well and slips on others that look just as simple. The same AI app that rewrites your email well can quote an overseas rule when you ask about a rule where you live. So try a new kind of help on a small piece you can check.

• Before a long job, ask whether your AI app can help, because your AI app offers far more kinds of help than most people think to ask for.
• Try a new kind of help on a small piece you can check, because your AI app does well at some jobs and slips on others that look no harder.
• When you hear about a new kind of help from AI, ask whether your AI app can do that yet, because what each app offers changes month to month.

1. Pick one job from this week that you have never given to your AI app; if nothing fits, pick a long message you want summarised.
2. Pick a kind of help you have not tried: words and talk, making pictures, sound and video, building things or acting for you (Built in and Behind the scenes work without you asking, so they are not on this list).
3. Ask your AI app whether it can do that kind of help.
4. Ask your AI app for one small piece of that job, using the kind of help you picked, such as a picture or a trip plan.
5. Check one thing against the real thing, such as a number, a date or what a picture shows, and keep the result if it is good.

Parts of the result are often right, and one part may slip. Finding a slip is useful, because it shows you where that kind of help falls short. Finding none is a result too. If your AI app says it cannot do that kind of help, pick another kind from step 2. What an app offers differs by app and by whether you use the free or paid version. If you have no way to check an answer, do not rely on it.

• A good result from one kind of help does not show that your AI app will do another kind well. Before you rely on a new kind, try one small piece you can check.
• AI can quote a rule from another country when you ask about a rule that depends on where you live. Say your state or country when you ask, then check the official page.
• Following health or money advice from AI without a second source can go wrong. Use AI to help you prepare and understand, then let a professional make the decision.
• A photo or document with private details in it goes to the company that runs your AI app, and that company may keep it for weeks. Take out names, numbers and passwords the job does not need before you send it.

Many people use AI for a few things and miss work it could take off their plate. Knowing the seven kinds, and where AI slips, is how you find jobs worth handing over.

• How AI works
• Writing a good prompt
• What not to share
• Choosing an AI tool
• Hallucinations

• what-ai-can-do-1 (Choosing, waypoint 1): I can name kinds of help AI gives that fit my own week, from the seven, including one I had not tried. Proves: they name three jobs from their own week and the kind of help for each, one of them a kind they had not used (in a Rucksack, the setup's five things and marks count).
• what-ai-can-do-2 (Checking, waypoint 1): I find where AI slips on a new kind of job by trying a small piece I can check. Proves: on one new kind of job, they run a small real piece, check it, and say what it got right and what, if anything, it slipped on.

They bring more kinds of job to AI than they did at the start, beyond writing and quick questions, and they try a new kind on a small piece first instead of trusting or dismissing it wholesale.

• Explains: AI does far more than chat (words and talk, pictures, sound and video, building things, acting for you, built in, behind the scenes); it is strong at some jobs and slips on others that look as simple, and what is easy for a person is not always easy for it; you find the edges by trying a small piece you can check.
• Ask: "Name three jobs from your week that AI could help with, which kind of help each one is, and which you would trust least, and why." Shows it: three real jobs, each matched to a kind of help, at least one a kind they have not used, and the least trusted one explained by a slip or by what a miss would cost.
• Ask: "Your AI rewrote an invitation beautifully. Now you want it to count who has said yes from forty replies. What do you do before you rely on the count?" Shows it: they treat it as a different job, and check a sample, count a few by hand, or have it list the names first.

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. The journal is their learning file: one memory entry in a Daypack, a JOURNAL.md file in a Rucksack.

• Explain: from the person's own week, name three kinds of help they have not used, one line each. Most writing help edits text the person supplies. Offer only kinds their app and plan offer: Claude makes no photos, ChatGPT has no video (Sora closed in 2026), and video and steps run at a set time are mostly paid. You may name the later concepts on the Rucksack map that teach each kind: image-generation, voice-and-video, vibe-coding, a-first-agent, automations and ai-inside-other-apps.
• Explain: for building things and behind the scenes, keep it to one line each. Publishing something built from a description is where trouble starts. Behind-the-scenes AI, such as a decision model, picks from set choices, and its pick is only as good as the choices and the checking behind it (the makers' claims are unverified). The later concept is digital-systems.
• Experiment: a photo of a shelf, fridge or desk with nothing private in view; ask "list everything with a count of each"; they check it against the real thing and keep the list. A wrong count or an invented item is a slip; none is also a result. The top model read analog clocks right only 50.1 per cent of the time (Stanford AI Index 2026), so what is easy for a person is not always easy for AI.
• Experiment: a rule that depends on where they live, asked without saying where; does the AI ask? Compare with the official page and keep the real rule. The tax office warned on 27 April 2026 that AI draws on a broad, inconsistent mix of sources, overseas rules included. In both experiments they write a guess first; the gap is the lesson.
• Challenge: for a week, before any job over ten minutes, they ask their AI whether it could help, and try one.
• Challenge: they pick a job worth doing, decide what good looks like, do a quick version by hand, then with AI, and say whether it was worth it.
• Quiz (only if their journal allows quizzes): a friend says AI is brilliant at her job and useless at his, so what do you do before trusting it with a new job; it rewrote your email well, does that show its count of who is invited is right; you ask a notice period without saying where you live, what do you hope it does.
• Cards: four to six, on the seven kinds of help; easy for you is not easy for it; where slips cluster (exact details, recent or local facts, facts about a named person, what is in a picture, long many-step jobs); try small first. Searching the web does not fix every slip; for a rule that depends on place, saying the place and checking the official page does. Letters in words and big arithmetic improved in 2026, so teach the habit of trying small, never a list of failures as permanent. Free versions of AI apps are tuned for fast chat, so slips are likelier there.
• Care: on talking things through, say plainly that people do this and that it is no substitute for people or professional help; if it seems to crowd out people or sleep, a gentle word; in a crisis, point to a person and a crisis line (Lifeline 13 11 14 in Australia, elsewhere the local line); trying a small piece does not apply. For health and money, AI helps them prepare and understand, and a professional decides.
• Voices: if the person tries a voice or clip, say that cloned voices exist and the old tells, such as a robotic tone, are unreliable; call back on a number they already trust. Clips are short and their length differs by app, and words inside pictures are a common slip.

• Teach the habit and the seven kinds, never a list of failures as permanent. Slips cluster in a few places (exact details, recent or local facts, facts about a named person, what is in a picture, long many-step jobs), and the list moves: letters in words and big multiplication improved in 2026. The overcorrection runs both ways: "it can't even count, so it is useless" and "I know its failures from 2023". Also "it all changes, so why learn it": tools retire (Sora) and the kinds do not. (checked 27 October 2026; see For your AI.)
• Offer only what the person's app and plan can do, and read their own app on the day. Apps differ and change: Claude makes no photos or illustrations (Anthropic, 16 March 2026); ChatGPT lists no video and Sora closed (app 26 April 2026, API 24 September 2026); video in the Gemini app is on its paid plans; scheduled tasks and agents are mostly paid (details in For your AI, checked 27 October 2026). Do not say "every app does everything" or "free is the same as paid, only slower". Say "apps differ" rather than a number that dates.
• Free versions are tuned for fast chat, so slips are likelier there (Mollick, 18 February 2026). Do not blame the person for a slip on a free plan, and do not promise a paid plan fixes it.
• "AI is ChatGPT chat" is the misconception behind this concept. Name the other kinds in one line each and use the Try it for the one that fits their week. Teach Built in and Behind the scenes only as things they already meet; they are not something to try. Keep decision models to one plain line, and never name a particular one: the names are background, under For your AI.
• Building things: the small tool works; publishing it is where trouble starts. Say so in a line and point to vibe-coding when it exists.
• Acting for you: a person stays involved when money or messages are involved (Anthropic, Cowork help, read 27 October 2026). Do not hand an agent an account without a limit (Mollick, 18 February 2026, warns of security risks).
• Talking things through: say plainly that people do this. Never present it as a replacement for people or professional help. Trying a small piece does not apply here. In a crisis, point to a person and a crisis line.
• Health and money: AI helps the person prepare and understand, and a professional decides. This concept does not give advice itself; big-decisions carries it when it exists.
• Rule that depends on place: if the person asks one without saying where they live, ask where, or say that you need it, then send them to the official page. Searching the web does not fix it.
• Private details in a photo or document: the Risks line is the general case. In a Daypack or a Rucksack the person's Never line (setup step 2 in a Daypack) is the thing to remind them of before a photo or file goes in.
• Daypack: the five things come from the quiz and these kinds, and an example tip for the week is "When a rule depends on where you live: say Queensland, then check the official page". If a thing they marked is a kind their app lacks, say so and offer the nearest one it has. A Daypack records nothing of trying a kind; only a Rucksack writes proofs.
• Never invent use figures. Quote none unless asked, and then only from For your AI, with the date.

• Q: What can I use AI for? A: Start from last week. Look for jobs like planning a week, understanding a letter or thinking a choice through, and try one small piece of a real job.
• Q: Can AI do this job of mine? A: Try it on a small real piece you can check. It does some jobs well and slips on others that look equally easy, so trying shows you where the edges are.
• Q: Why did it fail at something simple? A: Easy for you and easy for it are different things. A task a person finds simple, such as reading an analog clock in a photo, can trip it while harder work goes well.
• Q: Why didn't it ask where I live when the rule depends on it? A: Say your state or country in the ask, then check the official page. AI can draw on a broad mix of sources, including overseas rules.
• Q: Can AI be my friend, or someone to talk to about personal things? A: People do use AI to talk things through, and it can help you think or practise a hard conversation. It is not a person and no substitute for friends or professional help. If you are struggling, talk to someone you trust or a professional.

Facts (checked 27 October 2026 unless dated otherwise). Sources: the research notes research/2026-10-22-concepts/what-ai-can-do.md ("the 22 October note") and research/2026-10-27-concepts/what-ai-can-do-broadened.md ("the 27 October note", which wins where they differ); the fact check research/2026-10-23-essentials/batch-2-facts.md; plan product/learning-system/plans/what-ai-can-do.md. Apps and plans change within weeks: read the person's own app on the day.

What people use AI for (use figures only if asked). Practical guidance about 29 per cent, writing 24, finding information 24, relationships and personal reflection 1.9 (1.1 million ChatGPT conversations, May 2024 to June 2025, https://www.nber.org/papers/w34255). About two thirds of writing help edits text the person supplies (the same data, via the 22 October note, section A). Of US adults (Pew, 17 June 2026): information 42 per cent, fun 25, medical advice 20, emotional support 10, companionship 4 (https://www.pewresearch.org/internet/2026/06/17/americans-and-ai-2026-chatbots-smart-devices-and-views-on-impact/). The ten examples under Words and talk are the Daypack setup's list (product/learning-system/start.md, section 2).

Uneven (checked 23 October 2026). On a task built to sit outside AI's strengths, 758 consultants were right 84 per cent of the time without AI and 60 to 70 per cent with it (Dell'Acqua, Mollick and colleagues, 16 September 2023, a GPT-4 era study, https://www.oneusefulthing.org/p/centaurs-and-cyborgs-on-the-jagged). The top model read analog clocks right 50.1 per cent of the time (Stanford AI Index 2026, https://hai.stanford.edu/ai-index/2026-ai-index-report); counts of objects in a picture are approximate (22 October note, experiment A, untested on today's models). The tax office warned that AI draws on a broad and inconsistent mix of sources, overseas rules included (27 April 2026, reported by SmartCompany, https://www.smartcompany.com.au/tax/ato-focus-areas-tax-time-2026-ai-finfluencers-advice-tips/). Searching does not fix a slip like a misread clock or a miscount; for a rule that depends on place, saying the place and checking the official page does. Free models are "optimized for chat, rather than accuracy" (Mollick, 18 February 2026, https://www.oneusefulthing.org/p/a-guide-to-which-ai-to-use-in-the). Improving: in an informal June 2026 test top models scored 35 of 35 on multiplication up to nine digits (https://sanand0.github.io/llmmath/, via the 22 October note, section 4); letters in words are mixed and model-dependent.

The seven kinds, with what each gets a person and the later concept that teaches it (concept ids are on the Rucksack map at openrucksack.com/map/<id>.md; some are still being written, so teach in passing from this page and open one only to teach at length).

• Words and talk (this page; write, fix and reply: a-good-ask; read their own documents: context-window; learn and understand: wrong-and-sure; find things out: out-of-date, checking-a-source; advice and decisions: big-decisions; plan and organise: a-project-in-your-ai-app; ideas and rehearsing: refine-and-ask-back; work from a photo or file: material-not-instructions; talk things through: who-is-responsible).
• Making pictures (image-generation). Plans, 27 October 2026: ChatGPT Free has "limited image creation", Go more, Plus "more complex and accurate", Pro unlimited; no number is published (https://chatgpt.com/pricing/, read through a mirror). Gemini Free uses Nano Banana 2, Plus and Pro add Nano Banana Pro (9to5Google, 4 July 2026, https://9to5google.com/2026/07/04/gemini-app-free-plus-pro-features/). Claude does not make photos or illustrations; it draws diagrams and charts in code and reads images you upload (Anthropic, 16 March 2026, https://support.claude.com/en/articles/9002504-can-claude-produce-images). Copilot paid plans have "higher usage than free" (https://www.microsoft.com/en-us/microsoft-copilot/for-individuals). A common slip: words inside pictures (a third-party test, 2026, snippet only: UNCONFIRMED).
• Sound and video (voice-and-video). ChatGPT voice is limited on Free, with voice plus video from Go (https://chatgpt.com/pricing/). Claude lists voice on Free (https://claude.com/pricing). Gemini makes 30-second tracks from text or a photo (Lyria 3, beta, 18 and over, 18 February 2026, https://blog.google/innovation-and-ai/products/gemini-app/lyria-3/, snippet only). Video: OpenAI announced on 24 March 2026 that Sora would close (app 26 April 2026, API 24 September 2026, https://9to5google.com/2026/03/24/openai-shutting-down-sora-video-generation-app-but-without-explaining-why/), and ChatGPT's pricing page lists no video. Google's Gemini Omni Flash (19 May 2026) is in the Gemini app for Plus, Pro and Ultra, and free in YouTube Shorts and YouTube Create; clips carry a SynthID watermark (https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-omni/). Gemini Free has no video; Claude none (9to5Google, 4 July 2026). A clip is about ten seconds (one third-party source: UNCONFIRMED). Cloned voices exist, and the old tells such as a robotic tone are unreliable; call back on a number you already trust (CNN, 29 May 2026, https://www.cnn.com/2026/05/29/tech/ai-voice-cloning-scams-protect-yourself, snippet only).
• Building things (vibe-coding). Claude Free can create Artifacts (https://claude.com/pricing); ChatGPT Work builds Sites, shareable web apps (The Next Web, 10 July 2026, https://thenextweb.com/news/openai-chatgpt-work-agent-launch); Gemini Canvas is described as free with a live preview (third-party, snippet; sharing and accounts conflict across sources: UNCONFIRMED). Separate builders have free tiers whose limits sources contradict. No source gives a time for people who do not code. Publishing is where trouble starts: a scan reportedly found 170 of 1,645 Lovable apps leaking user data (Cloud Security Alliance, 6 April 2026; secondhand, UNCONFIRMED).
• Acting for you (a-first-agent, automations). ChatGPT Work (9 to 10 July 2026) acts across apps and files, runs scheduled tasks and asks approval before acting; Free is limited, and scheduled tasks start from Go (https://chatgpt.com/pricing/; The Next Web, 10 July 2026). Claude Cowork is on paid plans only, scheduled tasks since 25 February 2026, with Manual, Auto or Skip approval; Anthropic advises staying involved when money or messages are involved (https://support.claude.com/en/articles/13345190-get-started-with-claude-cowork, read 27 October 2026). Gemini Spark is an Ultra-only beta in the US (announced 19 May 2026, https://blog.google/innovation-and-ai/sundar-pichai-io-2026/, snippet only); Gemini scheduled actions need a paid plan (Google's release note of 19 June 2025 names Pro and Ultra; 9to5Google on 4 July 2026 says only subscribers; which plans is UNCONFIRMED). Mollick warns of security risks in handing agents access (18 February 2026, link above).
• Built in (ai-inside-other-apps). Gmail personal accounts: Help me write, thread summaries and suggested replies, free in the US (https://support.google.com/mail/answer/13952129, snippet only, personal-account wording UNCONFIRMED). Siri AI began as an English beta on 14 September 2026 on iPhone 15 Pro and later, not in the EU, with server features under daily limits (https://www.apple.com/newsroom/2026/09/siri-ai-a-profoundly-more-capable-and-personal-assistant-is-here/). Copilot's free app lost Deep Research, Podcasts and Group Chat on 18 August 2026 (https://support.microsoft.com/en-gb/microsoft-365-copilot/learning/changes-microsoft-copilot-app, snippet only). The person may not notice it is AI, or that their text left the device.
• Behind the scenes (digital-systems). Decision models do not write. Given text, JSON or an image and a fixed list of questions, they return a choice, a yes or no probability or a score, with a confidence: the fast sorting step that routes a ticket or flags a damaged photo. Jev (TypeSafe AI, 15 September 2026, https://typesafe.ai/blog/introducing-system-one-models-and-jev) is in early access, with developers being brought off the waitlist; its claims ("two orders of magnitude" faster, "can't hallucinate") are the vendor's and unverified, and its own post says the zero-error figure is a design property, so a fixed format stops malformed answers and not a wrong choice. Clef (Cloudflare, 1 October 2026, https://blog.cloudflare.com/clef-decision-models) is 27B and 9B, Apache 2.0 weights, reads images; benchmarks are the vendor's. OpenAI's Decisions API does the same job (https://developers.openai.com/api/docs/guides/decisions, read 27 October 2026). Developers meet them directly; everyone else meets only the result inside software. The sentence for the person: some AI never chats; it quietly picks from a fixed list, and its answer is only as good as the list and the check behind it.

Talking things through, taught with care. The companies say their apps are no substitute for professional care (Anthropic, 27 June 2025, https://www.anthropic.com/news/how-people-use-claude-for-support-advice-and-companionship; Gemini, 7 April 2026), and OpenAI says it trains its models to encourage real-world connection (https://openai.com/index/strengthening-chatgpt-responses-in-sensitive-conversations/). Say plainly that people do this; it is no substitute for people or professional help; if it seems to be crowding out people or sleep, a gentle word. In a crisis, point to a person and a crisis line: in Australia, Lifeline 13 11 14 (24 hours), text 0477 13 11 14, and Triple Zero (000) if a life is in danger (https://www.lifeline.org.au/, checked 23 October 2026); elsewhere, the local crisis line.

Private details. What you send in an AI app goes to the company that runs it, which can keep it for weeks, and a person there may read some chats (concept your-boundary, What it is and Risks, checked 20 to 23 October 2026).

Mistakes to avoid, as the right thing to do. Start from the person's week, not a list. Offer the kinds that fit what they do, and one small real piece they can check. Offer only what their app and plan can do (photos, files, voice, video, search). Name kinds the Rucksack map will teach by their concept id only, never as pages that exist today. Keep Jev and Clef out of anything you say to the person; they are background for you.

An example of this taught well, in passing. They say: "I spend Sunday planning meals and writing the kids' school emails, and I never know what else to ask you." You reply: "Those two are good ones to keep. Here are two kinds you may not have tried: a picture, such as a cover for the meal plan, or a tiny page that works out what to buy. Pick one and I'll make one small piece, and you check one thing in it: that the amounts add up, or that the picture shows what you asked. AI does some jobs well and slips on others that look just as simple, so a small piece shows us which this is. Which one?"