Which AI should I use? A plain guide to picking one

About half of American adults now use an AI chatbot, according to Pew Research Center’s February 2026 survey of 5,119 people. Ask the other half what’s stopping them and you rarely hear fear of the technology. You hear a shopping problem: there are ten of these things, they all claim to be the smartest, and nobody wants to pick wrong.

Here is the secret the regular users already know: for most everyday tasks, picking wrong is nearly impossible. The leading chatbots overlap so much that any of them will draft your email, plan your trip, or explain your lease competently. The question that actually decides your pick isn’t “which is smartest.” It’s four smaller questions: what’s the job, does the content need to stay private, what will you pay, and do you need it to work offline. This guide walks them in order, then names names in a clearly dated section, because the names change faster than the advice does.

“Which AI is best?” is the wrong question

Start with what people actually use. In Pew’s survey, 44% of US adults had used ChatGPT, 24% Gemini, 17% Copilot, 14% Meta AI, 8% Grok, and 6% Claude. It’s tempting to read that as a quality ranking. It mostly isn’t. It measures defaults and distribution: ChatGPT got famous first, Gemini rides along with Google’s products, Copilot ships inside Windows and Office, Meta AI sits inside apps billions already open daily.

Meanwhile, at the capability top, the products converged for everyday work. Hand a friend three answers to “draft a polite email rescheduling my dentist appointment,” one from each major chatbot, and they will not reliably tell you which came from which. Real differences survive at the edges: very long documents, serious coding, live information, tone. Reviewers argue about those edges for a living. But the tasks most people actually bring, the emails, the plans, the explanations, the lists, sit comfortably inside the overlap. (For a tour of what those everyday abilities are, see our plain-English tour of what AI can actually do.)

So the honest answer to “which AI should I use?” is usually: several would be fine. Which turns the real decision into finding the tiebreaker. And the tiebreaker is almost never raw capability.

Four questions decide it, and smartness isn’t one of them

When someone asks us which AI to use, we ask four things back, in this order.

First: what’s the actual job? “Using AI” is as vague as “using a computer.” Writing help, image making, voice, video, and question-answering over your own files are different product categories with different leaders.

Second: does the content need to stay private? Medical questions, finances, your employer’s documents, other people’s information. This is the one question that can completely change the answer, so it gets its own section below.

Third: free or paid? The free tiers are genuinely good now, and the standard paid rung costs about the same everywhere. The budget question is mostly about how often you hit limits.

Fourth: always online, or sometimes offline? Cloud chatbots need a connection and an account. Local models run on your own computer: no connection, no account, no usage meter. Most people never need this; the ones who do, need it badly.

Notice what’s missing: benchmarks, model version numbers, leaderboards. For a first pick, none of them matter. Answer the four questions and you land on a shortlist of two or three tools, any of which will serve.

Start from the task and pick a category, not a brand

Brands churn; categories are stable. The table below maps the five jobs people actually show up with onto the category that serves each, plus the thing worth comparing within that category.

Two notes on reading it. For most people, the first row is 80% of the story: a general chatbot handles writing, planning, explaining, summarizing, and light research, and every major one now also makes decent images and reads uploaded files. You don’t need five subscriptions; you need one general tool plus, occasionally, one specialist.

The specialist cases are real, though. If you’re making images in volume (a shop’s product photos, a comic, a brand), dedicated image tools give you controls chatbots hide. Voice and video are their own worlds with their own pricing, and video in particular is the one modality where costs jump from cents to dollars per attempt. When a task graduates from occasional to serious, graduate the tool with it.

The last row deserves a word, because it’s the fastest-growing job on the list. “Ask questions about my own files” covers the contract you don’t understand, the 60-page report you were supposed to read, the spreadsheet that won’t balance. Every major chatbot now accepts uploads and answers questions about them, and for public or harmless documents that’s the right tool. The moment the documents are sensitive, though, this row jumps the fork into the next section. Where your files can go matters more than which model reads them.

Privacy is the one hard fork

Everything above says “several would be fine.” Here is where that stops being true. The major consumer chatbots learn from your conversations unless you tell them not to, and the defaults are not in your favor.

The specifics, from the vendors themselves. Google’s Gemini privacy hub says plainly that your activity is used for “training generative AI models” and that “a subset of chats are reviewed by human reviewers.” OpenAI’s equivalent toggle, “Improve the model for everyone,” ships enabled, per TrustScan’s opt-out guide. Anthropic asks at signup: allow training and your data is kept up to five years; decline and retention is 30 days. None of this is hidden, but none of it is loud either.

There is also a middle path worth knowing: every major chatbot now offers a disposable mode. ChatGPT’s Temporary Chats and Claude’s incognito-style chats aren’t used for training regardless of your settings, and Gemini’s temporary chats aren’t reviewed for improvement. For the occasional sensitive question inside an otherwise ordinary account, a disposable chat is the two-second habit that covers you.

So the privacy branch has three exits. For ordinary content, flip the toggles once and move on; paid business tiers and APIs generally don’t train on your inputs at all. For sensitive-but-not-secret content, a good trick is to strip identity from the question: AI can analyze a lease or a budget without knowing whose it is, a pattern our personal-finance guide is built on. And for genuinely private material (client files, medical records, anything you’d shred), the clean answer is a local model on your own machine, where nothing leaves the building. That is the one case where the “which AI” question has a hard answer, and it isn’t a brand: it’s a location.

Free is real now, and ads are the new price

The free tiers stopped being demos. Google’s free Gemini runs a current model with real allowances. ChatGPT’s free tier runs a recent model with daily limits. Claude’s free plan covers web, phone, and desktop. For a person with occasional needs, $0 is a complete answer, and trying before subscribing is the entire point of the free rung.

But free found a new price in 2026: attention. On February 9, OpenAI switched on advertising in ChatGPT’s Free and $8 Go tiers in the US: labeled “Sponsored” results under answers, with OpenAI promising advertisers “won’t have access to chat history, memories, or personal user information.” Expect others to follow; an ad-free chatbot is becoming a paid feature, the way ad-free email did.

Above free, the pricing converged to the point of parody: ChatGPT Plus is $20, Claude Pro is $20 ($17 billed annually), and Google’s AI Pro is $19.99. Twenty dollars is the standard rung; the differences live in what rides along (storage, office tools, coding tools), not in the number. Budget rungs exist below it ($8 ChatGPT Go, $4.99 Google AI Plus), and $100–$200 professional tiers above it that almost nobody reading this needs. Our advice is boring: start free, and upgrade only after you’ve hit the free limits twice in one week. A subscription bought to “find out if it’s useful” usually just finds out you forgot to cancel.

The named examples, as of August 2026

Everything above should survive a year untouched. This section won’t, so it’s dated on purpose: here is where the categories land on names this month.

How to use the table. If you want the most popular default with the biggest ecosystem, ChatGPT. If your life already runs on Gmail, Docs, and Android, Gemini’s free tier is the natural start and its paid tier bundles storage you might be buying anyway. If your work is heavy on writing and long documents, Claude has the strongest reputation there. Microsoft’s Copilot, the 17% player, is the same idea pre-installed for people living in Windows and Office. And for the privacy exit, local apps running open models cover the chatbot job with zero dollars and zero data leaving your machine. The realistic bar is a computer from the last few years with 16 GB of memory or an Apple Silicon Mac; what you give up is some polish and the frontier edge, which for private paperwork is usually a fair trade.

If you read this in 2027: the prices will have drifted, the model names will have rolled over, and the four questions will still work.

Pick one, use it for a week, switch without guilt

The last thing to know is that this decision is cheap to unwind, which makes it cheap to make. Your chats export. Your files aren’t locked in. And the one real skill you build, asking clearly for what you want, transfers completely between tools; the habits in our guide to writing prompts that work work identically in all of them. Switching chatbots is closer to switching search engines than to switching banks.

The one-week trial matters more than the choice itself. A tool evaluated on a demo question tells you nothing; a tool that survives a week of your real email, your real planning, and your real documents tells you everything. Give it actual work, notice where it saves you twenty minutes and where it wastes ten, and let that ledger make the subscription decision for you.

So the whole guide compresses to one breath:

• Name the task, and pick the category that serves it; a general chatbot covers most of them.
• Check the privacy question; flip the training toggles, use disposable chats, or go local when it’s truly sensitive.
• Start free; pay $20 only after you’ve hit the free ceiling twice in one week.
• Need offline or absolute privacy? A local model on your own machine is the answer, whatever the brand weather.

Pick whichever shortlisted tool you’ll actually open. The differences at the top are smaller than the marketing implies, and the best AI, like the best camera, is the one you have with you.