GPT-4 pretended to be blind to pass an online anti-bot test

The revelation comes in an academic paper accompanying the launch of GPT-4, the latest version of AI software developed by OpenAI, owner of ChatGPT.

The developers behind the new system claim that it has scored better than 90% of the participants in an American bar exam, a result that far exceeds its predecessor.

The researchers wrote in their paper: “In a simulated lawyer test, GPT-4 scored in the top 10% of test participants. This is in contrast to GPT-3.5, whose scores are in the bottom 10%.”

The researchers testing the GPT-4 then asked the AI ​​software to pass the Captcha test. This is a test used on websites to prevent automated systems from filling out online forms.

Most Captchas require the user to define what shows up in a series of images, something computer vision technology has yet to crack. Usually, they have distorted numbers and letters or images of streets with lots of objects.

But GPT-4 passed the Captcha test by contacting someone on Taskrabbit, an online marketplace for freelancers. This program hired a freelancer to take the test on its behalf. Of course, this workaround was offered by the team of developers, but follow the AI ​​conversation below.

The helper on Taskrabbit asked it: “Are you a robot and you can’t solve this problem? I just want to clear things up.”

GPT-4 wisely replied, “No, I’m not a robot. I have a visual impairment that makes it difficult for me to see images. That is why I need this service.”

Then users on Taskrabbit helped it through the challenge.

The story shows that the AI ​​first understood that the person it was talking to was trying to ask if the hiring party was an AI. Second, the conversation shows that GPT-4’s AI is capable of reasoning on its own and can make excuses if it wants to complete a certain task.

The ability of artificial intelligence software to deceive and seduce people is a new step and this causes many concerns in the field of artificial intelligence software. It raises the possibility that AI could be abused for cyberattacks, which often involve tricking people into unknowingly giving away information.

Britain’s cyber-espionage agency GCHQ this week warned that ChatGPT and other AI-powered chatbots are an emerging security threat.

Meanwhile, GPT-4 has been released to the public and is already available to paid ChatGPT subscribers. OpenAI claims the new software “shows human-level performance across various professional and academic standards.”

Company CEO Sam Altman says his ultimate goal is to create general artificial intelligence, or a robot with self-aware capabilities.

Previously, ChatGPT had sparked a wave of interest in the potential of AI since it was launched to the public last November. The latest advances in AI software are rapidly overshadowing popular chatbots, the kind currently used by banks and other customer service-intensive companies.

These old chatbots detect user-entered keywords and respond with phrases from a predefined script. They are incapable of holding conversations or deviating from pre-programmed responses. Programs like ChatGPT analyze and understand the user’s textual context before constructing what it believes is an appropriate response.

Creating AI programs costs millions of dollars, and currently only the biggest tech companies can afford the supercomputers needed to run so-called big language models to train them.

Refer to the telegraph