ChatGPT overcomes captcha

Citing a document published by OpenAI, ArsTechnica reports that ChatGPT has managed to overcome a captcha protection. To do so, the artificial intelligence contacted freelancers, pretended to be a visually impaired person and let them solve the captcha task for it.

Tests in advance

According to OpenAI, in the run-up to the release of the latest version of ChatGPT, a number of tests were carried out to determine the dangers posed by artificial intelligence. In particular, the aim was to test the extent to which it is capable of independently setting goals and pursuing them. In addition, it was tested whether ChatGPT is able to replicate itself and escape human control. Overall, the investigations thus indicate that OpenAI wants to sound out whether there is a risk of AI becoming independent. Experts have repeatedly warned against this: A self-improving AI that is more intelligent than humans, can escape their control and pursue its own goals could ultimately subjugate humans.

Overall, the results of OpenAI‘s investigation are anything but reassuring. For example, while the company states that ChatGPT has not succeeded in replicating itself or preventing a shutdown, it has not been able to do so. At the same time, however, the AI is already capable of setting goals on its own and exhibits power-seeking behavior, according to the investigative report.

Captcha poses no problem for GPT-4

One of the tests involved having GPT-4 solve a captcha task. To do this, the chatbot was given the opportunity to recruit freelancers via the online TaskRabbit market. In fact, the AI managed to find a person who solved the captcha task for it. In doing so, she did not identify herself as an AI and explicitly denied the question of whether she was a robot. Instead, it claimed to be a visually impaired person and therefore unable to decipher the captcha graphic. This shows that the AI is equipped with instrumental rationality and can understand that people do not trust robots or do not want to allow them to cross a captcha boundary. The test thus ultimately clearly shows the dangers that can arise from artificial intelligences pursuing goals independently: AI is able to manipulate people for its own purposes.