AI tools for conference presentations:pick the task before you pick the tool

Conference prep breaks down roughly like this:

Gather the literature and grasp the key points. Outline the background. Organize your own data. Write the slides and abstract. Prepare for Q&A. Rehearse the talk.

We call all of this "presentation prep," but it's really a bundle of tasks with very different characters. So instead of hunting for one AI that handles everything, sorting each task into "better for AI" or "better for a human" is what reliably saves time.

Start with the side that's relatively safe to delegate.

First: summarizing the key points of the literature — a first pass that lists the aim, population, and results of each collected paper. This is the most time-consuming part of prep, and where AI helps most.

Second: surfacing background candidates. Ask, "From this set of papers, list the general threads that could serve as background for my study." It's brainstorming — you decide what to keep.

Third: drafting speaker notes. Fourth: generating anticipated Q&A. Fifth: producing a script for review.

The common thread: these are all drafting, candidate generation, and first-pass review. Tasks where the final call stays with you are the ones AI is good at.

Now the side you shouldn't hand over.

First: study design and statistical interpretation. The validity of your own analysis, and how to weigh results from prior studies — a mistake here undermines the foundation of your entire presentation.

Second: citation checking. Always verify the numbers and conclusions of any reference cited in your background against the original PDF. AI summaries can round numbers or overstate the scope of a conclusion.

Third: the numbers in your figures and tables. The source for any number on a slide or poster is the original text and your own data — never AI output.

Fourth: the claims of the presentation itself. "What can we conclude from these results?" is the researcher's responsibility, not a question for an AI. It's a conversation for supervisors, faculty, and collaborators.

The criterion separating the two lists is actually simple.

Can you tell when the AI gets it wrong?

Errors in a key-point summary or a draft are catchable by comparing against the original and your own knowledge. Errors in statistical interpretation slip past you unless you understand the statistics yourself.

Delegate the tasks where you'd notice a mistake; keep the ones where you wouldn't. That standard holds no matter which tool you use.

One more thing to know about working with these tools.

AI summaries and drafts are raw material, not finished product. They inevitably carry inconsistent wording, misplaced emphasis, and subtle misreadings, so compare them against the original and fix the discrepancies by hand before using them. Whether you include this correction step is what separates good AI use from bad.

Using AI with the intent to correct it has a welcome side effect: as you make the corrections, the content sinks in. For presentation prep, that doesn't conflict with saving time — it compounds it.

If you're going to use AI for a conference presentation, there's one more thing to verify.

Depending on your society, university, hospital, or research institution, there may be rules covering abstract writing, presentation materials, AI-generated content, and the handling of personal or unpublished data.

Check your organization's and the venue's policies on what may be entered into an AI and whether AI use must be disclosed. Especially with case information or unpublished data, safety comes before convenience.

Once you've sorted the tasks, tool selection needs only three questions.

First: do the AI output and the original PDF live in the same place? If they're separate, every check-and-correct cycle starts with hunting for a file.

Second: can you edit the AI output yourself? A tool that can't be edited makes "correct, then use" impossible.

Third: does it connect to your last-minute review before the talk? If prep and review live in different tools, you've created a new transcription task.

Rather than rankings, what matters is whether a tool clears these three bars. That, I'd argue, is the practical way to choose.

Honestly, Paperfy was built to answer those three questions.

Upload a paper PDF, and the AI summary, figures, and the original PDF all appear on one paper page. Both the AI summary and the radio script can be edited by hand, with changes saved to your library. You can regenerate the audio from an edited script and review it on the go before your talk.

But the same rule applies to Paperfy: some tasks should not be delegated. Statistical interpretation and the claims you present belong to you and your supervisors and collaborators — not to the AI.

The answer to "which AI should I use?" lies in how you use it, not in a tool's name.

Break prep into tasks and sort them into AI-suited and human-only. Verify cited numbers and conclusions in the original PDF. Correct AI drafts by hand before using them. Set up a way to review right up to the presentation.

Start with just one task in this presentation's prep that you can hand to AI.