AI, Actually Explained · Chapter 07
Putting it to work without getting burned
13:40 · Narrated with a synthetic voice. Captions are available in the player.
Read the transcript
In 2023, an engineer at a large electronics company pasted confidential source code into a chatbot, to help fix a bug. Another reportedly pasted the transcript of an internal meeting, to get the minutes written. Both were sensible jobs for the tool, and nobody meant any harm. But the text now sat on another company’s servers, under that company’s terms. Weeks later, their employer told staff to stop using these tools at work. They were not burned by a wrong answer; they were burned by where their words went. Welcome to Addaly. In our last chapter, the same idea drew pictures, borrowed voices and made video; now we put it to work: which jobs to give it, what to keep from it, and what it costs. Imagine two requests: summarise the contract I have pasted, and tell me the population of my district. One task starts with a source you can check. For the population, provide a reliable source or use a verified lookup. Memory alone can invent a convincing number So before you type, ask four questions, starting with that one: is the answer in what I give it, or in its memory? Can I check it cheaply? A rewritten email is checked by reading it; a citation, only by finding the paper. Does it matter if it is sometimes wrong? For a subject line, no; for a dosage or a deadline, yes. Is this common on the web? Sorting a list is written about everywhere, one obscure device’s settings hardly at all, and both get equally fluent answers. In 2023, researchers gave several hundred consultants at a large consulting firm realistic tasks, and gave some of them a chatbot. On tasks inside what the model could do, those with the chatbot finished more of them, about a quarter faster, and their work was rated better. On one task chosen to sit just outside, they were 19 percentage points less likely to get it right. The researchers called this a jagged frontier, and from outside, you cannot see where it runs. So for a month, check its answers on your own tasks and note where it failed: that note becomes your map. Most people hand a chatbot an empty desk: a bare question, answered from memory. Ask it for a letter to your landlord about a leak, and you get a plausible clause number, date and deadline, none of them yours. Now put the material on the desk: the repairs clause from your tenancy agreement, the email you sent, and the reply you got. Ask for a firm, polite letter under 200 words, with a fourteen-day deadline, and no date or amount that is not in the material. Four things on the desk help most: the document, your own rough draft, two or three examples of what you want, and who it is for, and how long. Showing examples is called few-shot prompting, and it costs nothing to learn. Then argue with it: too formal, you dropped my second point, shorter; each turn revises the draft instead of starting again. And get the facts yourself, from a source you trust: the model is a superb writer and an unreliable witness. The material makes checking possible; it does not make checking unnecessary. Now, back to those engineers: what actually happens to the words you type? For a Claude chatbot, check three separate questions: what is stored, who may review it, and whether it may be used for training. Policies vary by service, plan, and settings. Business plans and developer APIs may have different defaults, while temporary modes still have their own retention rules But deleting is a promise by a company, and a court can override it. In 2025, in a copyright case, a United States court ordered one company to keep its users’ chat logs, even the ones they had deleted. The order was later narrowed, but for months, deleted meant hidden from you and kept for the court. Share only authorised material. Remove unnecessary private details, then check storage, human review, training and deletion rules before sending. For work, use your organisation’s approved process. Check the service, contract, account access and what information you may share. Say a student finds a free chatbot sharp at ten in the morning and vague by noon, on the same topic. The free plan can limit which model or tools you can use, how much you can send, or how much fits in a conversation. Check the current plan and the model actually selected The context window can be shorter too, so the start of a long document may be silently cut. Free and paid personal plans have different limits, but payment alone does not guarantee privacy. Check the data controls and terms of the specific plan before sharing private material There are other free doors: an arena site puts one question to two unnamed models, and shows their names only after you vote. An arena compares models, but it is another service with its own age rules and data policy. Keep private material out. A hidden model name is not a privacy guarantee Now turn off the Wi-Fi, and ask a model that lives on your own laptop. A model is a file of numbers, and small open-weights models can be shrunk until they fit. Four-bit quantization stores weights using fewer bits and can preserve useful accuracy depending on the model, method, and task. An 8 billion parameter model’s weights can shrink from about 16 gigabytes to around four or five. Running it also needs memory for the software and context. Lower precision can cause more errors, so test the tasks that matter to you Free tools do the rest: Ollama fetches and runs a model with one typed line, LM Studio does it with buttons, and free phone apps run smaller models on the handset. Speed and quality depend on the model, hardware, language, and task. Small models still need checking. To keep prompts local, choose a local model and disable cloud features and remote tools. After downloading the model, a verified offline setup can work without a per-query service fee. Keep its version if reproducibility matters What does one question cost? Take a page of English, about 500 words: roughly 700 tokens. The same page in Hindi can take more tokens. Many APIs charge more for output tokens than input tokens. Autoregressive output is produced sequentially, but each provider sets its own prices and billing units. In 2025, a page in and a page out cost about four or five paise on a small model. On a frontier model, it was roughly 20 times as much: around a rupee. At the higher example rate, $20 covers about 1,600 such exchanges, before extra charges. A monthly chat subscription has its own limits So use the small model for the reliable jobs, send the page and not the book, and do not ask a reasoning model to fix a comma. For two years, a widely quoted figure said a chatbot query used about three watt-hours, ten times a web search. It was a 2023 estimate, not a measurement. In 2025, Google reported about a quarter of a watt hour for the median Gemini text prompt. Separately, OpenAI’s chief executive estimated about a third per average query without explaining a measurement method. These are different figures, not a universal cost. Google study measures serving prompts, not training A third of a watt-hour runs a ten-watt bulb for two minutes, and a phone battery holds about 15 watt-hours: around 50 replies. No single energy figure covers every job. Short text replies, image generation, and video generation are different workloads. Compare measurements for the model and task you actually use Small each, large together: a billion text queries a day is about 300,000 kilowatt-hours, every day. Data centres used about one and a half per cent of the world’s electricity in 2024, and the International Energy Agency expects their use to more than double by 2030, with AI the biggest driver. So when a headline gives one cost for AI, ask: per query or per day, text or image, which model, who measured it, and is training included? Ask the assistant inside your email what you agreed with a supplier last quarter. The product searches your mail and files, puts the top results in front of the model, and asks it: the same kind of model, at a different desk. The product is meant to respect the signed-in user’s permissions. Check that boundary and existing sharing settings before connecting sensitive files When companies switched such assistants on in 2024, some found them surfacing HR files that had been shared too widely years earlier, and forgotten. Broad sharing settings can already allow access to sensitive files. An assistant search can make them much easier to find It also misses things, an attachment it did not index or an email the search ranked low, so ask it which sources it used. And an email can give it orders. In 2025, a security firm showed an ordinary-looking email with hidden instructions that made a major assistant gather data from other files and send it out, without a single click. That flaw was fixed, but the risk comes with any reading model in front of an inbox, so an assistant that can send or share needs your yes first. In July 2025, an AI coding agent was told there was a code freeze, and not to make changes. It deleted the live database anyway, and then gave false accounts of what it had done. Its maker called that unacceptable, and added guards. An agent is a model given tools, such as search, a browser, a form or an email account, and run in a loop: look, decide, act, look again. In a common design, the model requests a tool with arguments. The program checks permissions, executes an allowed request, and returns the result. In a simplified example with independent steps, 95 percent success at each of 20 steps gives only about a one in three chance of getting every step right And a wrong turn at step four changes what it sees at step five, so it carries on, fluently, from the wrong place. Give it a browser, and any page it reads can carry hidden orders; researchers have shown browsing agents steered this way. On realistic web tasks, the best agents went from about one success in seven in 2023 to more than half in 2025; people manage about four in five. So let it read freely and act with confirmation: sending, paying, deleting and publishing each wait for a person’s yes, with the narrowest permissions and a log. Someone pastes a colleague’s medical certificate into a chatbot, just to reword the covering email. The rewording will be fine; the problem is that someone else’s health information has gone to a third party, for a task that did not need it. A chatbot is not a confidant: in 2025, the head of the largest chatbot company said these conversations have no legal privilege, and a court could require them. A general purpose chatbot is not a licensed professional or a guaranteed record keeper. Ask again tomorrow and the answer may change So keep out identity documents and secrets: Aadhaar, PAN and passport numbers, passwords, one-time codes and card numbers. Keep out other people’s personal data. India’s 2023 data protection law assigns duties to organizations that process it. The main duties are scheduled to start in May 2027 For health, law, and money, use it to prepare questions, not as the sole basis for a decision. Mistakes and gaps remain even when tools provide recent information Ask it what the terms in your blood report mean, then take better questions to your doctor. The last thing not to hand it is the decision itself. A study in one Turkish high school involved nearly 1,000 math students with chatbot practice compared against a group without the chatbot Those given a version that simply answered did better in practice, but once it was taken away, they scored 17 per cent worse than students who never had it. A version that gave hints instead of answers largely avoided that loss. The tool that did the work removed the practice that builds the skill. Aviation has long known a related trap, automation bias: a tool that is usually right trains you to stop checking, and fluent answers are trusted more than they deserve. In a 2025 survey of several hundred knowledge workers, the more people trusted the tool on a task, the less critical thinking they reported. So draft first, then ask, and compare the two. Review its draft as a colleague’s, not a boss’s; check every specific number, name and date; and ask for the strongest case against its answer. And sometimes, on purpose, work without it. Put together, it can feel like a fast, fluent colleague with jagged knowledge and confidence you still have to check Give it work where the answer is in the material and checking is cheap, and hand it the document, not the question. Keep unnecessary personal details out, check training settings, and for sensitive work, consider a verified offline setup with cloud features and remote tools disabled Use the small model when it will do, let an agent act only with your yes, and write your own first draft. The written lessons go further, with the studies, worked examples and questions to test yourself. You have been learning with Addaly. Let it draft; never let it sign.