Try it free

Resource

AI research assistant: how to choose one in 2026

By Jules Spector, a journalist (French press card no. 145851) and the founder of Verbatim.

Published on 17 October 2026

In short

An AI research assistant searches the web or your own documents, reads them and writes a summary with its sources. A good one shows the passage behind each statement so that you can check it. In October 2025, a study by 22 public service broadcasters found at least one significant issue in 45% of the answers AI assistants gave about the news. Choose a tool for its sources, its handling of private documents and the place where it stores your data.

This guide sets out what AI research assistants do in 2026, how to choose one and which errors to check for before you use their answers. It ends with the way Verbatim’s agents carry out research.

What an AI research assistant does

An AI research assistant takes a question and returns an answer built from sources. In 2026, these tools cover four tasks.

  • Searching. The tool queries the web and lists the pages it used. According to its help pages on 17 October 2026, Perplexity searches the internet in real time, and each answer includes numbered citations that link to the original sources.
  • Reading a corpus. The user uploads documents and asks questions about them. Google renamed NotebookLM to Gemini Notebook in July 2026. According to its help pages, the tool accepts PDF files, websites, YouTube videos, audio files, Google Docs and Google Slides, and its answers carry inline citations to these sources.
  • Summarising. The tool condenses dozens of pages into a note, a timeline or a comparison table.
  • Citing. The tool attaches a reference to each statement, in the form of a link, a page number or an excerpt.

AI agents add a fifth task, which is work over time. An agent follows a subject for weeks, reads new documents as they arrive and reports what has changed. It can also read thousands of pages in one go and cross-check them.

How to choose an AI research assistant

Six criteria separate one research assistant from another.

  • Visible sources. Each statement should link to the exact passage, with the page or the paragraph. A list of websites at the end of an answer does not let you check a given fact.
  • Exact quotes. A quote must reproduce the original word for word. Compare a few quotes with their source before you rely on the tool.
  • Private documents. Check which formats the tool reads, how many pages it accepts and how long it keeps your files.
  • Hosting and training. The terms of use state where data is stored and whether it is used to train the provider’s models. These terms sometimes differ between free and paid plans. Gemini Notebook’s help page, for example, states that user data is not used to train the tool unless the user sends feedback, in which case Google may review the full interaction.
  • Language. A tool that searches mostly in English can miss sources published in other languages. Test it in the language of your subject. Gemini Notebook says it supports more than 80 languages.
  • Work over time. A one-off question suits a search tool. A subject that develops over months calls for a tool that monitors sources and keeps all the project’s documents together.

Test a research assistant before you rely on it

Test the tool on a subject you know well. About ten questions are enough to judge it.

  • Ask a question whose answer you know, and open every source cited.
  • Check that each passage says what the answer claims.
  • Check that the cited source is the original publication.
  • Ask a question whose answer appears in none of your documents, and see whether the tool says so.
  • Upload a long document and ask for a figure from a specific page.
  • Ask for a quote and compare it word for word with the original.

Note the answers in a table with the date of the test. Run the same test on each tool you are considering. The results show which tool you can trust with which task.

The limits: invented citations and misread sources

Several studies measure how AI assistants handle sources.

In October 2025, the European Broadcasting Union and the BBC published the study News Integrity in AI Assistants. Journalists from 22 public service media organisations in 18 countries assessed more than 3,000 answers from ChatGPT, Copilot, Gemini and Perplexity, in 14 languages. Of these answers, 45% had at least one significant issue. Serious sourcing problems affected 31% of them, with missing, misleading or incorrect attributions. Major accuracy issues, including hallucinated details and outdated information, affected 20%.

In March 2025, the Columbia Journalism Review published a test of eight AI search tools carried out by the Tow Center for Digital Journalism. The researchers gave each tool excerpts from news articles and asked it to identify the headline, the publisher, the date and the link. Across 1,600 queries, the tools gave incorrect answers more than 60% of the time. Most of them presented their wrong answers with confidence. Some produced links that led to error pages, and the tools often credited a copy republished by another site in place of the original publisher.

Invented citations also appear in court cases. The database maintained by Damien Charlotin lists court decisions in which a party relied on content invented by an AI, most often fake citations. It counted 2,199 cases on 16 October 2026.

A link placed next to a statement does not show that the source says what the answer claims. Every quote and every figure has to be checked against the original before it is used. Our glossary defines AI hallucination.

Good practice with any research assistant

Six habits limit the risk, whatever the tool.

  • Keep the question, the answer, the sources and the date together in your notes.
  • Open the original of every source you plan to use.
  • Copy quotes from the original document, never from the assistant’s summary.
  • Note the page or the timestamp of each passage.
  • Treat any statement without a source as a lead to check.
  • Keep confidential documents out of tools whose terms you have not read.

These rules apply to every tool, Verbatim included.

Research with Verbatim

Verbatim gives you a team of AI agents, from the first lead to the final draft. You describe your subject in plain words. The agents search, read and follow your sources for hours or for weeks. Every fact stays tied to its source, and you make every call.

  • The reading agent searches the web and public registers, and reads up to 40,000 pages of your documents. It answers with the exact page.
  • The monitoring agent follows hundreds of sources day and night, including the press, public registers, official decisions, social networks and the sites you add. It flags what changes your subject, with the passage that proves it.
  • Verification matches every claim, whether a figure, a fact, a quote or an accusation, with its source and the exact excerpt. Each claim is marked verified, contradicted, outdated or inconclusive. A quote found only in an AI summary never counts as verified.
  • Data analysis ties every figure to the cells it comes from.
  • Transcription handles audio and video files, recordings made in the app and live interviews. It separates the voices and timestamps every passage, and quotes are copied word for word.
  • The People agent finds who knows the subject and how to reach them.
  • Writing turns the research into a note, a report or a summary in your style, learned from your published texts.

Several agents work on the same project in parallel and check one another’s work. A finding from the reading agent is checked against its source before it reaches your notes. Verbatim checks every sentence against the project’s sources and rejects any claim it cannot trace.

Verbatim tests models from several labs on each task and keeps the one that does it best. Verbatim has built its own technology to tie every fact to its source. Your documents are hosted and processed in Europe, and they never train a third-party AI. An iPhone app is on the way, along with Verbatim inside Claude and ChatGPT.

Journalism is the first field Verbatim serves, because it demands a source for every sentence. The same agents work for researchers, analysts and authors on their own corpora. VerbatimAgent costs from €49.95 a month including VAT, with a 14-day free trial.

Questions

What is the best AI for research?

The best tool depends on the task. For a quick question, a search tool that cites its sources is enough. For a large set of documents or a subject that develops over weeks, choose a tool that reads the whole corpus, shows the exact passage behind each fact and stores your data under terms you have read.

Can an AI research assistant make up sources?

AI tools can attribute a statement to the wrong source, produce broken links or invent references. In a 2025 test by the Tow Center for Digital Journalism, eight AI search tools gave incorrect answers to more than 60% of requests to identify the source of a news excerpt. Open every source before you use it.

Can I use an AI research assistant on confidential documents?

Read the terms of use first. They state where documents are stored, how long they are kept and whether they train the provider’s models. These terms can differ between free and paid plans. Verbatim hosts and processes documents in Europe and never uses them to train a third-party AI.

How do I check an AI assistant’s citations?

Open each link and find the passage. Check that it says what the answer claims, that it comes from the original publication and that it is still current. Copy quotes from the original document, never from the summary.

Sources

Other resources

See Verbatim on your own work.

Try it free for 14 days, or ask us for a demo built on one of your subjects.

Try it freeRequest a demo