Research with AI without outsourcing your judgment.

A polished report is the beginning of the review

AI-assisted research can help organise a question, find leads and assemble material into a readable draft. The danger is that a coherent report can feel more settled than the evidence underneath it. Headings, citations and a confident conclusion are useful presentation features, not proof.

OpenAI's introduction to deep research describes a system designed to search, analyse and synthesise online sources. It also discusses limitations, including hallucinated facts and incorrect inferences. That launch article is background on the capability, not a current statement of every plan, limit or feature.

The method here is tool-independent. Your goal is not to delegate judgment about what is true. It is to use AI to reduce some of the collection and organisation work while keeping the evidence chain inspectable.

Define the decision before the question

“Research this for me” is too open to produce a decision-ready result. You might be choosing a place for a sketching group, learning a subject, or checking a claim before sharing it. “Help me understand which capabilities matter for turning approved meeting notes into reviewed action lists” gives the research a job. It identifies the workflow rather than inviting a catalogue of popular products.

Specify the audience, geography if relevant, date sensitivity and what the output must support. Are you learning a subject, comparing alternatives, or preparing a recommendation? A broad educational overview and a purchase decision need different levels of detail and verification.

Write the exclusions too. If you are not selecting a vendor yet, say so. If you need primary sources, name that requirement. If a feature is not publicly documented, ask for it to remain unconfirmed rather than inferred from marketing language.

Request a source plan, not just an answer

Before asking for a full report, ask what kinds of sources would answer the question. For a product capability, official documentation and the actual interface may matter. For a research claim, the paper's methods and results matter. For a policy, the controlling document matters.

This step can reveal that your question combines different kinds of evidence. An official page can establish what a company offers without proving that the product will save you time. A customer case study can describe one outcome without establishing that it generalises to your organisation.

Ask for gaps to be listed alongside findings. A source plan is useful when it makes missing evidence obvious, not when it produces a long list of impressive domains.

A worked example: could a sketching group use this room?

Suppose you want to invite fifteen people to an evening sketching session. Before asking AI to recommend a venue, define the checks: room capacity, availability on your chosen date, and whether the proposed activity is permitted.

For this teaching example, imagine a fictional library policy containing two sentences: “Room bookings are limited to twelve people.” “Weekday evenings may be available by agreement.” These are invented excerpts, not information about a real venue. The film shows how to inspect them.

From a claim to a checked conclusion

A fictional room policy shows why a plausible recommendation still needs an evidence check. The film is an authored example, not a live AI result.

Read the video explanation

Question: could a room work for an evening sketching group of fifteen? Establish capacity, the actual date and whether the activity is permitted.

Inspect the fictional policy: bookings are limited to twelve people; weekday evenings may be available by agreement.

The claim that fifteen people fit is contradicted. The source supports a limit of twelve. The actual date and activity remain unresolved.

Choose another room or change the group size. Confirm the date and activity before inviting anyone, and keep the checked source.

A compact evidence note could look like this:

The useful conclusion is not “this venue is perfect.” It is “the proposed group exceeds the stated limit, and the date and activity still need confirmation.” A clear answer may tell you not to proceed.

A research prompt you can adapt

Help me investigate [question] for [purpose]. First suggest the kinds of primary sources needed and the questions they should answer. Use only the sources you can actually inspect or that I supply. For important claims, give the source URL or document location, a supporting passage, its date where relevant, and any conditions. Separate direct evidence, your interpretation, and unanswered questions. Do not invent sources or imply that an untested proposal is proven.

If the tool cannot browse, ask it to work from material you supply or help make a search plan. A list of remembered titles is a lead to investigate, not a verified evidence trail. Open links yourself; if a source is inaccessible, mark the claim unresolved rather than accepting a guessed summary.

Build a claim ledger

For each consequential claim, record four things: the claim, the source location, what the source directly establishes, and what still needs checking. This can be a simple document. It does not need to be a specialist research system.

Consider a fictional claim: “Tool A can process every file in our shared drive.” The source says it can connect to a drive service. That does not establish support for every file type, permission scope, size limit or workspace configuration. The ledger should expose that gap before the statement reaches a recommendation.

Use status labels such as supported, partly supported and unresolved. Avoid a single confidence percentage unless there is a defensible method behind it. A precise-looking number can hide a vague judgment just as easily as a polished paragraph.

Open the sources that carry the conclusion

You do not have to read every background link with equal intensity. Identify the sources that determine your conclusion and inspect those carefully. Find the exact passage. Check its date, context and conditions. If the link redirects or the relevant claim is absent, investigate.

Be especially careful with numerical claims. What was measured? Against what baseline? Who was included? Was it a controlled experiment, a survey, a vendor example or an estimate? A percentage separated from those details may communicate more certainty than the source supports.

If a result is important enough to repeat in public, make sure you can explain it without relying on the AI report's phrasing. When you cannot, either do more checking or present the uncertainty plainly.

Separate retrieval from inference

“The documentation lists feature X” is a source observation. “Feature X might help our workflow” is an inference. “We should adopt the tool” is a recommendation that also depends on cost, access, risk and alternatives.

Ask the report to keep those layers distinct. You can use a format such as finding, implication, caveat and next check. That makes it easier to disagree with a recommendation without discarding the underlying evidence.

For example: finding—an export format is documented; implication—it may fit our downstream tool; caveat—we have not tested the import; next check—use a non-sensitive sample file. This is less dramatic than announcing seamless integration, but much more useful to the person who must implement it.

Look for evidence that could change your mind

If the initial report supports what you already hoped, ask what would make the conclusion wrong. Search for limitations, unsupported file types, exceptions or contradictory results. This is not an instruction to manufacture balance where the evidence is clear; it is a way to avoid stopping at the first convenient answer.

Ask whether several sources ultimately repeat the same original claim. Five articles quoting one press release are not five independent confirmations. Trace the important statement back to its origin and describe the dependence honestly.

When sources disagree, preserve the disagreement long enough to understand it. They may be measuring different tasks, using different versions, or referring to different dates. Do not average incompatible claims into a smooth but meaningless conclusion.

A first experiment and a stopping rule

Choose one modest question you can check independently. Give the tool two or three short, non-sensitive sources rather than requesting a huge report. Ask for three consequential claims and the passage supporting each. Open the sources, compare the wording, and correct the note yourself.

Keep a small record: question, source location, when you checked it, supported conclusion and unresolved points. For a file, include the page or section and version if available; for a changing web page, the check date matters. This makes a later revision possible without relying on the original chat.

Stop when a missing fact determines your next action and you cannot establish it. Get the information, narrow the question or leave the conclusion open. More fluent wording does not resolve an evidence gap.

End with a decision and a boundary

A useful research note should say what you can reasonably conclude, what remains unresolved and what to do next. Sometimes the next step is a small test. Sometimes it is a question to a vendor. Sometimes there is not enough evidence to recommend anything.

Keep the underlying sources and your verification notes with the result so it can be revisited when circumstances change. Record when you checked time-sensitive information. A report without its evidence trail is harder to maintain than its polished appearance suggests.

AI can help you cover ground. Your responsibility is to know which ground you have actually checked. The best outcome is not the longest report; it is a clearer decision with a visible chain of support.

Go to the source

Primary sources checked on 13 October 2026. Publication dates and product details may differ; check the source for its scope.

• OpenAI: Introducing deep research (2025)
• OpenAI Academy: Deep Research video (2025)