Choosing the Right AI Tool for the Job

Most bad AI research starts with the wrong tool. Someone asks a chat assistant, from memory, what their county spent on road gravel last year, gets a confident number, and repeats it at a council meeting. The assistant never saw the county's budget. It produced a number shaped like an answer.

So the first question is what kind of job you have. AI research tools fall into a handful of families as of October 2026, and each fails in its own way. Products change names and features every few months, so learn the families. The brand names below are examples, not endorsements, and some will be renamed by the time you read this. (One already was: Google renamed NotebookLM to Gemini Notebook on July 16, 2026.)

The families

Academic search tools such as Elicit search scholarly papers and pull findings into tables, which helps when an issue turns on research studies. Local models, run with free software such as Ollama on your own Mac, Windows or Linux machine, keep your text off anyone else's servers. Local models are usually weaker than the big cloud ones. For sensitive material, weaker and private beats strong and leaky.

What the numbers say about AI search

The Tow Center for Digital Journalism at Columbia tested eight AI search tools in a study published March 6, 2025. Researchers gave each tool excerpts from real news articles and asked it to name the headline, publisher, date and link. Across 1,600 queries, the tools gave incorrect answers more than 60 percent of the time. The spread was wide: Perplexity got 37 percent wrong, Grok 3 got 94 percent wrong. For Grok 3, 154 of 200 citations led to error pages.

The tools rarely hedged; ChatGPT signaled low confidence in only 15 of its 200 answers. And the paid versions of two tools answered more questions correctly than the free ones but also got more wrong, because they guessed instead of declining. Paying more does not buy you humility.

A larger study by 22 public service broadcasters in 18 countries, coordinated by the European Broadcasting Union and the BBC and released in October 2025, found significant problems in 45 percent of more than 3,000 AI answers about news. Lesson 4 covers that one in detail.

Matching the job to the tool

Ask one question first: where does the answer live?

If the answer lives in a document you hold (a budget, an audit, minutes), upload the document or put it in a notebook. If it lives somewhere on the web, use an AI search engine and treat its links as a reading list. If you do not even know what exists yet, a deep research run can sketch the field in ten minutes, and then you go get the real documents. If the text is your own, a chat assistant works as a decent editor, and it adds nothing you did not give it.

Free versus paid, and what plans change

Plans shift constantly, so check the vendor's own help page before you rely on a limit. A few patterns hold as of October 2026. Deep research modes are rationed; OpenAI says usage varies by plan and shows a counter, and Claude's Research mode requires a paid plan with web search turned on. File uploads have size caps; Claude's help center lists 500 MB per file and 20 files per chat (30 MB per file in projects), and says it reads both text and visuals in PDFs of 100 pages or fewer. Business and enterprise plans from the major vendors generally do not train on your content by default, while consumer plans have settings you need to check. Lesson 8 covers privacy in full.