Over the weekend, an AI-generated image of Pope Francis wearing a puffy white coat went viral on Twitter, and apparently many people believed it was a real image. Since then, the puffy pontiff has inspired commentary on the deceptive nature of AI-generated images, which are now nearly photorealistic.

The pope image, created using Midjourney v5 (an AI image synthesis model), first appeared in a tweet by a user named Leon (@skyferrori) on Saturday and quickly began circulating as part of other meme tweets featuring similar images as well, including one that humorously speculates about a pope “lifestyle brand.”

Not long after, Twitter attached a reader-added context warning to the tweet that reads, “This is an AI-generated image of Pope Francis. It is not a genuine photo.“

As noted in our piece on last week’s AI-generated Donald Trump arrest photos, Twitter guidelines state that users “may not deceptively share synthetic or manipulated media that are likely to cause harm.” Although in this case, the line between harm and parody might be a fuzzy one.

How do we know the image is fake? Aside from a Reddit post containing alternative images of the Pope from the person that likely made it, The Verge breaks down the evidence fairly well in a piece analyzing the impact of the false image. For example, if you zoom in on details, you’ll see telltale signs of image synthesis in warped details like the pope’s crucifix necklace, the crooked shadow of his glasses, and whatever he is carrying in his hand (a cup?).

But still, upon a quick glance, the false photo (“fauxto”?) looks fairly realistic. And as The Verge notes, a stylish image of Pope Francis plays into our beliefs about the papacy, which often involves wild non-fake imagery—although Pope Francis is known for his “humble” outfits.

A Midjourney journey

The image service used to create the fake photo, Midjourney, debuted last year. Along with DALL-E and Stable Diffusion, it’s one of three major image synthesis models that have become popular online. All three allow users to generate novel images using only text descriptions called “prompts.”

In this case, the prompt used to create the puffy pope photo might have been as simple as “Pope Francis in a puffy white coat” because Midjourney has made huge leaps in photorealism recently, rendering complex scenes full of details from relatively simple prompts.

What this almost effortless capability to fake photos means for the future of media is still uncertain, but as we’ve speculated before, due to image synthesis, we may never be able to believe what we see online again.

On Sunday, Singapore-based retrocomputing enthusiast Yeo Kheng Meng released a ChatGPT client for MS-DOS that can run on a 4.77 MHz IBM PC from 1981, providing a unique way to converse with the popular OpenAI language model.

Vintage computer development projects come naturally to Yeo, who created a Slack client for Windows 3.1 back in 2019. “I thought to try something different this time and develop for an even older platform as a challenge,” he writes on his blog. In this case, he turned his attention to MS-DOS, a text-only operating system first released in 1981, and ChatGPT, an AI-powered large language model (LLM) released by OpenAI in November.

As a conversational AI model, ChatGPT draws on knowledge scraped from the Internet to answer questions and generate text. Thanks to an API that launched his month, anyone with the programming chops can interface ChatGPT with their own custom application.

Thanks to his new app, which can run on MS-DOS, Yeo can use a vintage IBM PC-compatible computer to chat with ChatGPT over the Internet. It’s a similar back-and-forth conversation as the traditional ChatGPT web interface, albeit as a text-only, full-screen application running on the antique machine.

Development challenges

MS-DOS posed a particularly challenging platform for a ChatGPT client, lacking native networking abilities. In addition, Yeo targeted a computer with very limited processing power: a 1984 IBM 5155 Portable PC, which includes an Intel 8088 4.77 MHz CPU, 640KB conventional memory, CGA ISA graphics, and MS-DOS 6.22.

To create the client, Yeo used Open Watcom C/C++, a modern compiler running on Windows 11 that can target 16-bit DOS platforms. For testing purposes, he used a VirtualBox virtual machine running DOS 6.22 to streamline the development process, then he transferred the compiled binary to the target IBM DOS PC for testing.

To handle networking on the IBM PC, Yeo needed to weave his way through several layers. First, Yeo utilized a “Packet Driver API” standard invented in 1983. He integrated the open source MTCP library by Michael B. Brutman into the application to communicate with the Packet Driver, enabling networking capabilities for the client.

For the ChatGPT API, Yeo used OpenAI’s Chat Completion API, constructing the POST request (and parsing the JSON-formatted response) manually in C.

However, Yeo hit a major snag: the ChatGPT APIs require encrypted HTTPS connections. Since there are no native HTTPS libraries for MS-DOS, Yeo had to create an HTTP-to-HTTPS proxy that can run on a modern computer and translate the requests and responses between the MS-DOS client and ChatGPT’s secure API, acting as a transparent middleman in the communication process.

Yeo says that reading and writing input to the console presented another challenge due to the single-threaded nature of DOS applications. He devised a method to check and receive keypresses without pausing the program using the MTCP page and online samples as a reference.

In the end, the client works better than Yeo expected, and he looks forward to more retro challenges in the future: “After experiencing this, I will definitely be writing more retro-software in future,” he writes in a blog post that describes his development process in more detail.

Yeo has released his code (called “doschgpt”) on GitHub if others want to run it themselves—or perhaps improve or extend the code in the future. With a little creativity, the latest tech in AI language models need not be limited to cutting-edge machines.

Portions of Twitter’s source code recently appeared on GitHub, and Twitter is trying to force GitHub to identify the user or users who posted the code.

GitHub disabled the repository on Friday shortly after Twitter filed a DMCA (Digital Millennium Copyright Act) takedown notice but apparently hasn’t provided the information Twitter is seeking. Twitter’s DMCA takedown notice asked GitHub to provide the code submitter’s “upload/download/access history,” contact information, IP addresses, and any session information or “associated logs related to this repo or any forks.”

The GitHub user who posted the Twitter source code has the username “FreeSpeechEnthusiast,” possibly a reference to Twitter owner Elon Musk casting himself as a protector of free speech.

“It was unclear how long the leaked code had been online, but it appeared to have been public for at least several months,” a New York Times article said. Despite that, the NYT article said Twitter “executives were only recently made aware of the source code leak.”

GitHub user FreeSpeechEnthusiast’s profile indicates the user joined GitHub on January 3, 2023, and made its only code submission on the same day. Twitter’s DMCA notice to GitHub described the code as “proprietary source code for Twitter’s platform and internal tools.”

Suspect list could include thousands of ex-employees

The leaker may have been one of the roughly 5,500 employees who left Twitter via layoff, firing, or resignation after Musk bought the company. Twitter also reportedly laid off about 5,000 contractors shortly after the Musk acquisition.

“Twitter began an investigation into the leak and executives handling the matter have surmised that whoever was responsible left the San Francisco-based company last year, two people briefed on the internal investigation said,” the NYT wrote.

Musk said on March 17 that Twitter will make “all code used to recommend tweets” open source by March 31, but the leaked code may be much more sensitive. The NYT said its sources indicate that Twitter executives are concerned “that the code includes security vulnerabilities that could give hackers or other motivated parties the means to extract user data or take down the site.”

Twitter sent the takedown notice on Friday and asked a federal court to issue a subpoena later the same day. “The DMCA Subpoena is directed to service provider GitHub,” Twitter’s request for a subpoena said. “GitHub operates a website to which the infringing party or parties (identified by their GitHub username as FreeSpeechEnthusiast) posted various excerpts of Twitter source code, which posting infringes copyrights held by Twitter in those materials.”

Twitter seeks “all identifying information”

Twitter’s proposed subpoena seeks “all identifying information, including the name(s), address(es), telephone number(s), email address(es), social media profile data, and IP address(es), for the user(s) associated with the following GitHub username: FreeSpeechEnthusiast.” It also asks for “all identifying information provided when this account was established, as well as all identifying information provided subsequently for billing or administrative purposes.”

The subpoena request further seeks all identifying information for any “users who posted, uploaded, downloaded or modified the data” at the repository where the Twitter source code was posted.

When contacted by Ars, GitHub did not comment on Twitter’s request for the user’s identifying information or the attempt to obtain a subpoena. “GitHub does not generally comment on decisions to remove content. However, in the interest of transparency, we share every DMCA takedown request publicly,” a GitHub spokesperson said. The Twitter DMCA takedown notice was posted by GitHub here.

GitHub is owned by Microsoft. Another Twitter court filing contains the email thread between Twitter and GitHub that led to the takedown on Friday. It appears that GitHub disabled the repository less than an hour and a half after Twitter filed the takedown notice.