Capstone: Test AI Vision With a Chatbot

Time to see everything you learned in action, with your own hands and no code. Today's popular AI chatbots can look at images you upload and describe them, classify them, read their text, and answer questions about them. They are, in effect, a friendly front end to the computer vision ideas from this whole course. In this capstone you will run a series of small experiments to watch AI vision succeed and watch it fail, so the concepts stick for good.

What You'll Learn

• Which chatbots can see images and how to upload one
• Simple experiments that show each vision task in action
• How to deliberately push a system until it makes mistakes
• How to connect what you observe back to the course

Which tools can "see"

Several widely used AI chatbots are multimodal, meaning they accept images as well as text. As of mid-2026, ChatGPT (OpenAI), Gemini (Google), and Claude (Anthropic) all let you upload an image and ask questions about it, and each offers a free way to try this, with daily or hourly usage caps on the free tiers. Any one of them works for this capstone. Pick whichever you can access.

A useful thing to notice: these assistants read and understand images, but that is separate from creating images, the recognition-versus-generation split from lesson one. In fact, Claude can analyze images but does not generate them at all, a clean reminder that "seeing" and "drawing" are different skills. For this capstone you only need the seeing side.

How to upload is simple. In the chat box, look for a paperclip, plus, or image icon, attach a photo from your device, then type your question and send. If you are on a phone, you can often take a photo directly.

A quick, sensible caution before you start: do not upload images of other people without their consent, ID cards, or anything private or sensitive. Use your own everyday photos, screenshots, and objects around you.

Experiment 1: Classification and description

Upload a clear photo of a single object or pet and ask:

"What is in this image? Describe it in one sentence."

This is image classification and description in action. Try several photos: a pet, a plant, a meal, a household object. Notice how it does. Then upload something a bit unusual, an odd angle, a partly hidden object, a rare item, and watch whether it stays accurate or starts guessing. You are testing the "something it never saw clearly" failure mode from the limits lesson.

Experiment 2: Object detection and counting

Upload a photo with several things in it, a desk, a shelf, a street, a group of objects, and ask:

"List everything you can see in this image, and tell me how many of each there are."

Now you are exercising object detection and counting. Check the count carefully. Chatbots are often good at listing what is present but can miscount when objects overlap or crowd together. Try a photo with, say, a bowl of fruit and count along yourself. Where it miscounts, you have found the edge of its ability, and you understand exactly why from the detection lesson.

Experiment 3: OCR, reading text

Find something with text: a receipt, a book cover, a food label, a handwritten note, a sign. Upload it and ask:

"What does the text in this image say? Type it out exactly."

This is OCR. Start with clear printed text, which usually works well. Then push it: try messy handwriting, a blurry photo, glare on a label, or a stylized logo font. Watch where it slips, especially on handwriting and on look-alike characters like zero versus the letter O. You are reproducing, on purpose, the OCR limits you read about.

Experiment 4: Break it on purpose

This is the most instructive experiment, so do not skip it. Your goal is to make the AI confidently wrong, because seeing that with your own eyes is the lesson that lasts.

Try these:

• Upload a photo in poor conditions: very dim, blurry, or with strong glare. Ask what it is. Notice if the answer gets vaguer or wrong.
• Upload something ambiguous, an object that looks like two things, or an unusual item, and see if it commits confidently to one wrong answer.
• Ask it a trick question about the image, like asking it to count something that is not there, and see whether it pushes back or plays along.
• Upload the same photo twice with slightly different questions and see if the answers stay consistent.

For each miss, ask yourself which failure mode from the limits lesson you just triggered. That connection, from a real mistake you caused to a concept you named, is the whole point of this course.

Experiment 5: Ask it to show its uncertainty

Upload any image and add this to your question:

"How confident are you, and what would make you unsure?"

A well-behaved assistant will often admit when an image is unclear or when it is guessing. This mirrors the confidence-score idea from the classification lesson. Learning to ask for uncertainty, and to trust an "I am not sure" more than a slick, certain-sounding answer, is a habit that will serve you across all of AI, not just vision.

Wrapping up the course

Step back and look at what you can now do. You understand that an image is a grid of numbers, that AI learns to see by studying labeled examples, and that the field breaks into a handful of tasks: classification, detection, face recognition, and OCR. You can spot computer vision across daily life, and, most importantly, you know why these systems fail and how to use them without being fooled. In this capstone you tested all of it live.

That is a genuine foundation. Vision is one core piece of how modern AI works, and the same "learn from examples" engine drives language and other AI too. If you finish the final exam, you will have earned a free certificate you can add to your LinkedIn profile, your resume, or a school application. Well done for getting here.

Key Takeaways

• Popular chatbots like ChatGPT, Gemini, and Claude are multimodal: upload an image and ask questions, with free tiers available and usage caps.
• These tools read images, which is separate from generating them. Claude, for instance, analyzes images but does not create them.
• Run experiments for each task: describe and classify, list and count, read text, then deliberately break the system.
• Making the AI confidently wrong, in bad light, on ambiguous items, or on messy text, is the fastest way to truly understand its limits.
• You now have a real, no-code foundation in computer vision. Finish the final exam to earn your free certificate.