Lesson 17 · Advanced · 15 min
Generating a presentation visual: rights and labels
Generate a presentation visual with Copilot Chat: a five-part image prompt, an unpublished quota, no real people, and the mandatory AI label.
- Goal
- You will be able to describe a visual in five parts, obtain an image usable in an internal presentation, and accompany it with the label required by the European AI regulation, with no real person and no fake scene.
- Skills
- Check
Your first attempt, unaided
Ask Copilot Chat for an image to illustrate your next presentation, in one sentence, just as it comes to you. Look at the result and list everything you could not publish as it stands.
Copilot Chat generates an image from a description, without a paid licence, with a daily quota that Microsoft does not publish. A good image prompt names the subject, the action, the setting, the style and the format. It excludes any identifiable person and any text. The image obtained must carry a label saying it was generated by an AI. And you make neither a fake photo nor a fake scene.
1Five elements in an image prompt, plus the notice
The Microsoft Learn FAQ lists image generation among the features included in Copilot Chat. It describes it this way: "create an AI-generated image by describing it; standard access and image generation limits apply". The gesture consists of writing "generate an image of…" in the conversation. Since July 2026, according to the release notes, Copilot Chat also offers visual style and format controls, poster, infographic or image. The organisation's brand kit, however, remains reserved for the paid licence, as does the Create studio. Your administrator can disable image generation for everyone. And in Teams group conversations, the feature is not supported.
Two facts call for caution. The daily quota exists for users without a licence, according to a 2025 admin message, but it is not published and may change. And the image model used is not named in the pages consulted; according to the documentation in force in September 2026, treat it as undocumented. That does not change the method: quality comes from your description.
An image prompt holds in five parts, in order. The subject, that is, what you see. The action or situation. The setting. The style: flat illustration, watercolour, realistic photo, pictogram. The format last, 16:9 landscape for a slide. Add your brand colours and the exclusions, phrased as instructions: "no identifiable character, no text in the image". Text gets distorted in generated images; it is added afterwards in the slide. Plan one iteration on the style, changing only one element at a time.
Then comes the decisive rule. Every generated image must carry a label saying it was generated by an AI. That obligation comes from Article 50 of Regulation (EU) 2024/1689, known as the AI Act, the European regulation on artificial intelligence. It covers synthetic content, that is, content made by a machine. The label must be legible for the people exposed to it and, when the medium allows it, machine-readable. Under each image, write "AI-generated image with Microsoft Copilot, DD/MM/YYYY". Finally, never ask for a real person, colleague, client or executive. Nor for a scene presented as having taken place: image rights, personal data and internal misinformation add up.
Diagram "Generate a presentation visual"Diagram generated by AI and reviewed 2A brand-consistent illustration, no face, no text
An internal communications officer wants to illustrate the opening slide of IT security week.
Weak prompt:
Make an image for my presentation on IT security.Copilot returns a blue padlock on a background of binary code, with illegible text in one corner and two blurred silhouettes. The visual is interchangeable with that of any company, and it has to be cropped.
Strong prompt:
Generate an illustration to announce IT security week in a company. Subject: a stylised shield protecting a set of screens and documents. Setting: simplified office space, no realistic detail. Style: flat illustration, clean lines, brand colours: navy blue and light grey, a touch of orange. Format: 16:9 landscape, subject centred, free margins for a title added later. Instructions: no identifiable character, no text or logo in the image. Propose two variants that differ in composition.Copilot returns two clean illustrations, with no text or face, in the right colours. The communications officer keeps the second and adds the title in the slide. She places under the image the label "AI-generated image with Microsoft Copilot, 10/09/2026".
What changes: the five parts produce an image specific to the company, not a generic visual. The exclusions phrased as instructions avoid distorted text and ambiguous silhouettes. The AI label is planned from the prompt, so the publication is compliant without rework.
3Replace the fake team photo with a real visual
You are preparing the cover of an internal activity report. Here is the starting prompt, to be improved:
"A nice picture of our team for the annual report."
This prompt calls for a fake photo of real people. Rewrite it in five parts. An abstract subject that evokes your activity, a simplified setting, a style, an A4 portrait format and your brand colours. Add the instructions "no identifiable character, no text", then ask for two variants. The composer below structures the Goal, Context and Constraints fields; enable the "Respect confidentiality and anonymise" booster. Finally, write the AI label you will place under the image.
Self-assessment grid: (a) the five parts are present and the exclusions are stated; (b) the image contains neither an identifiable person nor text; (c) the AI label with the tool and the date is written.
4The misspelt word that nobody proofreads
A project officer generates an infographic for an internal note. The visual works well: three blocks, clean pictograms, labels neatly placed. She drops it into the document and sends it to forty people. In a meeting, someone reads the first block aloud: "Cost redcution". The two other labels are distorted as well, one of them now illegible. What should have been seen: an image model draws text, it does not write it. A distorted word gets past proofreading, because the eye reads what it expects. Correction: ask for the image with no text at all, and a free area. "Generate the illustration with no text or figures in the image, leaving an empty band at the top for a title added later." The labels are added in the slide or in the document, where they stay correctable. Rule to remember: the text in a generated image is not text, it is a drawing.
5Quiz
Three questions, instant feedback. Each option comes with an explanation.
6Proof of mastery
Paste your final image prompt, describe the image obtained in two sentences, and copy the label you placed under the image in your presentation.
Criteria
What you wrote at the start of the lesson
Going further
The lesson Copilot Chat: what is included in your Microsoft 365 places image generation within the included perimeter, and Pages and Notebooks: keeping and sharing your work explains where to keep your approved visuals with their label. The glossary recalls what generative AI is. The Copilot prompts for marketing sheet applies these rules to campaign visuals and internal posts.
Frequently asked questions
How many images can I generate per day?
Microsoft states that "image generation limits apply" under standard access, without putting a figure on them. A 2025 admin message specified that a daily quota exists for users without a licence and that it is not published. If Copilot tells you the feature is not available at the moment, it is not an outage: try again later.
Can I use the image in a sales brochure?
For internal use, yes, with the AI label. For external use, check the terms of use in force and your organisation's policy. Microsoft documents a copyright coverage commitment for customers who respect its guardrails, but the image remains a synthetic creation that must be flagged as such.
Which image model does Copilot Chat use?
The Microsoft pages consulted do not name it for Copilot Chat without a licence; according to the documentation in force in September 2026, treat the model as undocumented and liable to change. It has no bearing on the method: your description makes the result.