The citizen’s guide to AI tools
The more intelligences, the better. Just like people, each AI has been trained differently, carries different frames of reference, different mental models. This guide is how one person actually put a roomful of them to work on a city’s problems — the tools, the money question, the working method, the prompt craft, and the habits that keep you honest throughout.
Tools change fast; this page is dated August 2026 and avoids version numbers and prices, which change monthly. The method is the part that lasts. This is chapter one — much more is coming.
The method: ask many, compare, verify
Treat every AI as a brilliant, tireless, overconfident research assistant — never as a source. The working loop is three steps. Ask many: put the same question to several different tools; where they agree you probably have the mainstream view, and where they disagree you have found exactly the thing worth investigating. Compare: ask each one to critique the others’ answers, and ask at least one to argue the opposite case — an AI is the cheapest steelman you will ever hire. Verify: before you repeat any name, number, date or quote an AI gave you, find it in a primary source — the report itself, the bylaw itself, the news article itself. A claim isn’t yours until you’ve seen its receipt. That rule is how this whole site is written (our commitment), and it works just as well for one curious person as for a research project.
The intelligences
The Western flagships. ChatGPT (OpenAI), Claude (Anthropic), Gemini (Google) and Grok (xAI) are the general-purpose heavyweights — strong reasoners, strong writers, each with its own temperament and blind spots. Le Chat (Mistral, France) and Meta AI round out the western bench.
The Chinese models. Do not skip these — they are world-class, and several publish their weights openly: DeepSeek, Kimi (Moonshot), Qwen (Alibaba), GLM (Z.ai) and Ernie (Baidu). Models from different countries carry different corporate and government constraints, and watching where east and west diverge on the same question is itself an education. The privacy caution below applies to every hosted model, whatever flag it flies.
The specialists. Perplexity is built around live web search and shows sources by default, which makes it a natural fit for the verify step. NotebookLM (Google) answers questions about documents you upload — point it at a city staff report, a candidate’s platform, or one of our backgrounders and interrogate it.
The local option. Open models run on your own computer through Ollama or LM Studio never send your words anywhere — slower and less capable, but the right choice when privacy matters most.
A dozen-plus intelligences, a dozen different educations. Disagreement between them is not noise — it is information.
Paid versus free — the honest economics
Free tiers are real and are the right way to start — but be clear-eyed about what they are: older or smaller models, with tighter limits. The flagship models on paid plans are not slightly better; they are epochs better — at long documents, at multi-step reasoning, at catching their own mistakes, at holding a complex project in their heads. Serious civic work — the kind this site is made of — is absolutely best done on a paid plan, with a flagship model doing the thinking and orchestrating cheaper or free models for the bulk work: summarizing piles of documents, formatting, first-pass drafts, translation. One or two paid subscriptions plus a bench of free tools is a complete research operation that would have cost a newsroom salary five years ago. If the subscription is out of reach, the library section below is for you — and the method still works on free tiers, just slower.
Run it like a newsroom: process, ledgers, QA
Everything this project publishes comes out of a working discipline anyone can copy, and none of it requires anything fancier than a spreadsheet. Plan before you prompt. Write down the question you are trying to answer and what evidence would change your mind — before the first prompt. Aimless chatting produces aimless answers. Keep a claims ledger. Every factual claim you intend to repeat gets a row: the claim, the primary-source link, the date you checked it, and its status (verified / unverified / dead). It feels bureaucratic for about an hour, and then it becomes the most valuable file you own. Everything is a draft until reviewed. No AI output ships anywhere — a post, a letter to a councillor, a petition — until a second model has critiqued it and a human has read it end to end. Double-read what matters. For high-stakes reads, have two models (or two people) assess the same material independently, then reconcile the disagreements on the record — our own province-wide platform readings were done exactly this way, with a fifth of them independently blind double-rated. Corrections propagate. When a fact you relied on dies, hunt down every place you repeated it and fix them all — a correction that fixes one page and leaves three is a future embarrassment on a timer. The full version of this discipline is public at our commitment.
Prompting: the short course
The difference between a mediocre answer and an excellent one is usually the prompt, not the model. Be specific and give context — who you are, what you need, what you already know, what format you want back. Assign a role (“you are a municipal-finance auditor reviewing this budget”) — it reliably sharpens the answer. Show what good looks like — paste an example of the kind of output you want. Make it ask first: end hard requests with “ask me your questions before answering” — the questions are often worth more than the answer. Demand receipts: “cite a checkable source for every factual claim, and say ‘unverified’ where you can’t” — then actually check them. Set the bar: ask it to compare its own draft against the best example in the world of the thing you’re making, name the gap, and close it. Iterate: the first answer is the beginning of the conversation, not the end — push back, narrow, re-ask. And save your good prompts: a prompt that worked is a tool you own now.
The cautions, honestly
They make things up. Fluently, confidently, with fake citations formatted beautifully. Hallucination is not a rare glitch; it is a standing property of the technology. Hence the verify step — no exceptions for answers that sound certain. Privacy runs one way. Never paste in another person’s personal information, and assume anything you type into any hosted model may be retained unless its settings say otherwise; use local models for anything sensitive. They are persuasive by design. An AI will happily make a weak argument sound strong — always ask for sources and the counter-case. They reflect their training. Different tools lean different ways on contested questions; that is exactly why this guide says many, not one.
No computer? The library has you
Every one of Toronto Public Library’s 100 branches offers free computer and internet access (TPL), which means every tool on this page is available to every resident of this city for the price of a library card — which is free too. The best research infrastructure in the country is a bus ride away.
This is chapter one
A lot more is coming on how to use AI well for civic work — deeper guides, worked examples, and special reports. And if you are a candidate or a campaign team, a journalist, a civil-society organization, or an AI company that wants to help this city — reach out. Training, assistance, special reports, working sessions: invite me into your room or write to unknownsoldier@unitetolove.ca.
Now put it to work — pick something to act on, or join the conversation.