Concerns about scams stemming from ChatGPT's ability to create realistic images.

ChatGPT "draws like a photograph"

On March 26th, the global and Vietnamese AI user communities were abuzz when OpenAI officially released Images in ChatGPT, a tool focused on image creation, for users of its intelligent model. Previously, ChatGPT also had image creation capabilities, but these were limited and relied on another well-known text-to-image model called DALL-E.

The tool's ability to transform text into drawings and perform "Photoshop-quality image merging" quickly attracted attention. Many users reported the output image quality as "astonishing," with most rating it as "superior" to other popular AI models on the market.

ChatGPT creates realistic coffee cup images with just a simple command.

According to Gabriel Goh, head of research at OpenAI, the significant improvement in ChatGPT's image generation capabilities using the GPT-4o model is "Binding"—a term referring to the degree to which the AI ​​image generator maintains links between attributes and objects. Goh asserts that GPT-4o can accurately link attributes for 15-20 different objects without error, significantly improving the accuracy and reliability of the generated content.

Not only does Images in ChatGPT excel in image creation, but it also addresses a critical weakness in other AI image creation tools: rendering text within images. In fact, many images containing text are rendered accurately, making it increasingly difficult to distinguish between real and fake images.

The tool's drawback is its relatively long image creation time, but this is considered reasonable because the delay is due to the fact that high-quality images with accurate details and matching the user's description always require patience. The research team stated that the delay issue will be improved in the future.

Many concerns have been raised about the possibility of image manipulation.

Beyond the initial enjoyment that ChatGPT and OpenAI offer users, concerns about information security, privacy, and the risk of exploitation have once again come to the forefront. "Soon, the images you see online may not necessarily be real," a Facebook user in Vietnam expressed his worry, as he could request ChatGPT to create completely fabricated images that look as realistic as photographs, for example, a picture of actor and former wrestler Dwayne Johnson (The Rock) advertising a chewing gum brand, or a calico cat sitting on a stove that looks exactly like a photograph.

With detailed commands provided, users can create surprising results without needing in-depth knowledge of traditional photo editing tools.

In fact, many people have even tried the new tool's "undressing" feature. TN ( Hanoi ) said she used a model's photo from the internet to request ChatGPT to change her clothes as requested, and almost all requests were fulfilled with high accuracy. "However, there were a few times when requests were rejected, mainly due to violations of the system's policy regarding the creation of copyrighted images or overly suggestive content," TN shared.

In tests conducted by Thanh Nien newspaper , ChatGPT refused to create images related to famous or copyrighted characters in the real world or in movies and comics. "I cannot create the image as requested because the reference image you sent clearly shows Doctor Strange – a copyrighted character of Marvel – and this violates our current content policy," the system responded when asked to create an image of Doctor Strange sitting... drinking iced tea in Vietnam.

Jackie Shannon, head of multimodal product at ChatGPT, affirms that Images in ChatGPT offers robust protection against deepfake, pornographic, or fraudulent content. Notably, the generated images incorporate standard C2PA metadata to identify them as AI products, but require a separate tool for verification.

However, many people still successfully "circumvent the rules" and use images of celebrities for entertainment purposes. For example, a picture of CEO Tim Cook holding a Samsung phone instead of an iPhone, or an image of actor Robert Downey Jr. chewing a vegetable candy, or a series of paintings of Cristiano Ronaldo and Lionel Messi competing for a ball in a small alley in Vietnam in front of many onlookers...

Source: https://thanhnien.vn/lo-ngai-gia-mao-voi-kha-nang-tao-anh-nhu-that-cua-chatgpt-185250327005110274.htm