# How should I use AI as a speaker? ChatGPT, Claude and DeepL for abstracts, slides and rehearsal

Author: Marin Niehues, Stage Starter · Published: 2026-07-28 · Updated: 2026-11-16
Topic: Call for Papers & Submissions (https://stage-starter.com/blog/topic/call-for-papers/)
Source: https://stage-starter.com/blog/using-ai-as-a-speaker/

As a speaker you use AI well when it works with your own material, and badly when it is meant to supply the abstract or the evidence itself. Language models such as ChatGPT and Claude review sources and give a second opinion on clarity, while DeepL translates first drafts for international stages. None of these tools writes an abstract that holds its own against hundreds of others, because it lacks what a program committee looks for: your concrete experience and your angle on it. Using AI well as a speaker means giving the machine your own raw material and then letting it work.

**AI saves a speaker time as soon as it works with their own raw material, and it produces an interchangeable text as soon as it is asked to invent the content itself.**

## Where does AI really take work off a speaker's hands?

AI takes over the work whose result is already contained in your own material, above all reviewing sources and checking drafts. Reviewing material saves the most time. When you prepare a topic for a session, you collect articles and your own project documents, and sorting this pile is work a machine does faster than you. You still check every figure and every source that makes it into your session against the original.

The second area is the second opinion. A draft abstract, the short talk proposal you use to apply to a conference, can be checked for jumps in the argument and for passages that need prior knowledge. The answer comes immediately and without scheduling, and a human test reader still remains the stricter reviewer.

The third area is variants and translations. Twenty title suggestions for finished content take seconds, and you pick out the one usable idea yourself. For international stages, DeepL delivers a solid first draft that you then check for terminology and tone, and when a rewrite pays off instead of a translation is described in the article [Abstract in English](/en/blog/conference-abstract-in-english/). Then there is managing deadlines and pending acceptances, because anyone who submits regularly tracks dozens of statuses over months.

**AI yes, AI no**
AI is well suited to tasks whose result is already in your material, such as sorting sources or shortening a draft. It is unsuited to anything that requires new substance, meaning the abstract itself and the evidence in it. The question before every use is therefore: do I already have the content, or am I expecting it from the machine?

## Which AI tools are suitable for speakers?

Three groups of tools suit speakers, each taking on a different job. General language models such as ChatGPT or Claude review material and check drafts, and both handle long documents such as project reports or transcripts. Translation tools such as DeepL deliver first drafts for English-language submissions. Rehearsal tools such as Speaker Coach in Microsoft PowerPoint give feedback during practice on pace and filler words.

So can AI give you feedback while you rehearse your talk? For the measurable things, yes: a rehearsal tool reliably picks up pace and filler words, and a language model can check a transcript of your rehearsal for gaps in the structure. Whether a story lands and whether the room goes along at the right moment can only be judged by a human listener.

To check an abstract, a good prompt describes the conference and its audience along with the criteria of the call. It asks the model to read as a member of the program committee and name weaknesses. The more precisely conference and audience are described, the more useful the feedback, because otherwise the model checks against a general standard. You do not let it rewrite the text, or the interchangeable tone comes back.

## Should AI write my abstract?

AI should not write your abstract, because a generated abstract reads fluently and is worded in a language that fits everywhere. An abstract is always judged against everything submitted to the same track by the same deadline, and what fits everywhere stands out nowhere in that pile. Some organizers now actively ask about the use of generative tools or prohibit it in their submission terms.

The second kind of damage comes from invented facts and sources. Language models produce plausible figures, plausible study titles and case examples that do not exist, and in a specialist audience there are people who know the field. An invented case study is quickly exposed in front of an experienced audience, and the damage to your credibility exceeds any benefit of the smoother story.

The third kind of damage hits the preparation itself. A talk comes about largely through working your way through your own material and noticing which idea is yours and which story holds up. If you outsource this thinking, you stand in front of the room with material you have not fully understood yourself, and that shows at the first critical question at the latest. What makes a [good talk](/en/blog/how-to-build-a-good-conference-talk/) emerges in exactly this working through.

**What a submission is up against**
- Three figures | is the number of submissions per edition at mid-sized conferences
- A few minutes | is what a submission gets in the first reading round
- 5 percent at most | acceptances per conference are the market norm for large open calls

## Why a program committee recognizes generated texts

A program committee, the panel that selects a conference's talks, reads three-figure numbers of submissions at mid-sized conferences, often over a few evenings. At that volume, uniformity stands out long before anyone checks deliberately: the same sentence rhythm, the same transitions and phrasing that claims significance without naming anything.

Program committees recognize generated submissions above all by what is missing, namely the connection to their own audience and personal experience. The first filter checks whether the talk belongs in this program, and that requires the submission to pick up the organizer's language. The second filter asks why you in particular should give this talk, and the answer depends on a role, a project or a turning point from which the insight comes. A model knows neither and fills both gaps with generalities. How strongly this selection works is described in the article on the [call for papers](/en/blog/what-is-a-call-for-papers/), the public invitation through which a conference collects talk proposals.

On paper, automation still looks attractive, because an abstract costs you several hours and the machine two minutes. The calculation still does not work out, because the acceptance rate falls so far with generic submissions that more submissions bring fewer acceptances, and because an accepted submission with invented evidence costs you your credibility in front of an audience. As generated submissions pile up, committees weight evidence and references more heavily, and that hits newcomers without a list of references hardest.

## The raw material has to come from you

The substance of a session consists of things a machine structurally cannot supply. The evidence in the middle requires a case you were involved in, with real figures, hard decisions and the one detail that makes it credible. A talk with a clear position is memorable because the audience can take a stance on it, and a position grows out of experiences that cost you something. A generated text balances positions and lands in the middle.

Using AI makes sense at the point where your raw material already exists. Turning the recording of a talk you have given into text modules for your own channels, or drafting a version of your finished abstract for a differently focused conference, are tasks with a clear benefit. How a clear profile develops from this over the years is described in the article on the [personal brand as a speaker](/en/blog/personal-brand-as-a-speaker/).

For most consultants, the real task lies in the content: the material exists, it is scattered across twenty projects, and nobody has ever condensed it into a stage version. This inventory works best in conversation with someone who knows what a program committee is looking for. [Stage Starter](/en/) takes this work off your hands: we draw the material out of your practice, develop talk and abstracts from it, find the right calls, write the submission and submit it. With our own submissions for clients we achieve up to 19 percent acceptances per conference, while the market norm at this volume of submissions is 5 percent at most, and the collaboration starts from $999 net per month. What a speaker agency does overall and how the process works is shown in the [guide to speaker agencies](/en/blog/speaker-agencies-in-germany/).

**AI shifts a speaker's work to other places, because research, checking, variants and admin become faster. What an organizer books you for and what makes a room take you seriously stays just as demanding, namely your experience and your position.**

**Sources**
Weber-Wulff, Debora, et al. (2023): Testing of detection tools for AI-generated text. International Journal for Educational Integrity, volume 19. The study tested fourteen detection tools and found none that recognizes AI texts accurately and reliably. The acceptance rates come from Stage Starter's analysis of our own submissions for clients.

## Frequently asked questions

### Am I allowed to use AI for my abstract at all?

You can use AI to check, shorten and vary your abstract, as long as problem, angle, evidence and benefit come from you. Some organizers now ask about the use of generative tools or prohibit it. A look at the submission terms therefore belongs before every submission.

### Do detection tools recognize AI texts reliably?

Detection tools do not recognize AI texts reliably, they produce false hits in both directions and are useless as proof. An experienced reader recognizes arbitrariness even without a tool. Arbitrariness is reason enough for a rejection.

### Which AI tools are suitable for speakers?

Language models such as ChatGPT or Claude suit speakers for reviewing and checking, DeepL for translations and rehearsal tools such as Speaker Coach in PowerPoint for feedback on delivery. All three work best with material that already comes from you. None of them is suitable for evidence and case stories.

### Which prompt is suitable for checking an abstract?

A good prompt for checking an abstract describes the conference and its audience, names the criteria of the call and asks the model to read as a program committee and name weaknesses. You do the rewriting yourself. That way the text stays yours and the feedback becomes concrete.

### Can AI give me feedback while I rehearse my talk?

AI reliably gives you feedback during rehearsal on measurable things such as pace, filler words and pauses. A transcript of your rehearsal can be checked for gaps in the structure. Whether a story works in the room is better judged by a human listener.

### Can I have AI build my slides?

AI can be used for structural and text suggestions on slides, but hardly for the structure of the session itself. The deck follows your narrative and your pace, which are only fixed once you have spoken the session aloud. After that, AI speeds up the design of individual slides.

### Where does AI pay off most for speakers?

AI pays off most for speakers after the appearance. Turning a talk you have given into material for your own channels and for follow-up submissions is laborious, repetitive work. It relies entirely on raw material that already comes from you.

Which material lies in your projects and at which specialist stages your clients sit is something we work out together in an [initial consultation](/en/erstgespraech/). Afterwards you will know which talk we would submit for you.
