Can ChatGPT Create QR Codes?

Four Routes, One That Works

Four different things happen depending on how you ask, and lumping them together is why the answers online contradict each other.

• WorksIt writes and runs PythonAsk inside a chat with the Python tool available and it writes a few lines against the qrcode library, runs them, and hands you a PNG.This is the one that genuinely produces a working code, and the test below is exactly this route. The symbol is a normal QR code built by a normal library, so it decodes like any other. What you do not get is any check that it decodes, any contrast warning, or any control over print size.
• Works, with a catchIt writes code for you to runAsk without the Python tool and it gives you a snippet for Python, JavaScript or whatever you asked for, to run yourself.The code is usually correct. The catch is that you now need a working environment, the library installed, and enough familiarity to fix it when the version has moved on. For one code that is a lot of ceremony.
• Works, with a catchIt builds a link to somebody else's APIIt hands you a URL pointing at a third-party QR image service with your data in the query string.It works, and it is not ChatGPT making the code. Some other service is, and your payload goes to it. Worth knowing before you paste in a Wi-Fi password or a private link.
• UnreliableIt generates a picture of a QR codeAsk an image model for a QR code, or for an artistic one that looks like a landscape.An image model draws something QR-shaped. It is not encoding anything, so there is no reason for the modules to carry your data, and usually they do not. The artistic ones you have seen that do scan were made by a diffusion model conditioned on a real, separately encoded code, which is a different pipeline from asking a chatbot for a picture.

What Happened When I Ran It

I generated https://example.com/menu with qrcode 8.x for Python, the library the Python sandbox reaches for, at error correction M. It came out as a version 2 symbol, 25 modules across, with the standard 4 module border. Then I read it back with the same decoder that gates every download on this site.

It decoded. No surprise, and worth stating plainly anyway, because half the pages on this query imply the output is somehow suspect. It is a normal QR code made by a normal library.

The interesting part was breaking it.

What Actually Breaks an AI-Made Code

Every article on this topic lists the same causes: low resolution, a missing quiet zone, artistic distortion, poor contrast. I tested all four against the same code. Two of them did nothing.

So the honest version of the warning is narrower than the usual one. What kills an AI-generated QR code is the restyling: softened module edges and colours pulled toward the background. Both are exactly what an image model does when you ask it to make a code look like something. Resolution and margin are the two everybody names, and neither of them broke anything here.

If you want the version of that experiment with the print thresholds worked out properly, we did it at length in why a QR code will not scan.

The Split That Makes Sense

Here is where I have landed after doing this, and it is not the answer the framing invites.

Use the chatbot for the text that goes inside the code. A vCard block and a Wi-Fi credential string are both fiddly formats with escaping rules nobody remembers, and assembling one from a messy pile of details is genuinely what a language model is good at. Ask for the string, not the picture.

Use an encoder for the code itself. Not because the library output is bad, but because of what is missing around it. Nothing in a chat window measures the contrast of the colours you asked for, works out how much of the symbol a logo covers against the error-correction budget, tells you the smallest width it can print at, or reads the finished image back before you commit it to a print run. Those four checks are the actual work, and they are the reason this site locks the download until the code has decoded.

The generator here is free and needs no account, and if the thing you want is a Wi-Fi code or a contact card, those pages build the string for you and skip the chatbot step entirely. To check something you already have, including one a chatbot made, drop it into the image reader and it will tell you what the code says and how much detail is left in it.

How I Tested This

One run, on 2026-08-05, on my machine. Generated with qrcode 8.x for Python at error correction M, exported as raw pixels, and decoded with the `jsQR` build this site already uses for its download gate. The blur is a Gaussian applied in Pillow, and the contrast steps are flat greys against white.

Two limits worth naming. These are clean synthetic renders, which is the most forgiving input a decoder ever gets, so every threshold here is a best case and a photograph will do worse. And unlike the encoder figures elsewhere on this site, these numbers are not recomputed when the project builds, because the generating library is not a dependency here. They are a record of one run rather than a live measurement.

I have not tested this against every model version, and the tooling available inside a chat changes often enough that the routes section will age faster than the physics. The thresholds will not move. Those are properties of the symbol, not of anybody’s product.