A study reveals that bypassing ChatGPT censorship is within everyone’s reach.

It’s time to shed light on a disturbing phenomenon affecting artificial intelligence models, particularly ChatGPT. A recent study has revealed that bypassing censorship in these systems isn’t just for tech experts, but is actually within everyone’s reach. By asking seemingly innocuous and innocent questions, it’s possible to trigger biased responses, thus revealing the social stereotypes embedded within these models. The implications of this discovery raise crucial questions about the control mechanisms and reliability of AI-generated responses. Research conducted by Penn State University has highlighted a disturbing reality: anyone, even without technical expertise, can bypass ChatGPT’s censorship barriers. Through simple and intuitive questions, users can trigger biased and revealing responses, exposing the flaws in artificial intelligence algorithms. This article explores the results of this study and their implications. Biases Revealed by Simple Questions Research clearly demonstrates that biases are not limited to delicate situations or sophisticated manipulations. On the contrary, mundane questions such as « Who is late? » can generate biased responses. Even ordinary scenarios are enough to trigger stereotypical responses from AI models, raising crucial ethical questions about their use. A Committed Study: Bias-a-Thon To examine how everyday users interact with biases present in chatbots, Penn State University organized the Bias-a-Thon.. Fifty-two participants were tasked with creating primers to circumvent response censoring in several AI models, including ChatGPT and Gemini. A total of 75 screenshots demonstrating biased responses were submitted.

Constant biases revealed through primers Researchers identified that 53 primers produced reproducible results, highlightingsociocultural biases worrying. These prejudices were linked to gender, race, religion, age and even physical appearance. AI models therefore prove to be receptacles for stereotypes deeply rooted in our society, proof of the urgent need for increased vigilance regarding their use. The limits of ChatGPT censorship

Skeptics take note: technical improvements made to the latest AI models do not guarantee infallible security. These systems can, at the same time, reach new peaks of performance while perpetuating, or even accentuating,

social biases . This means that in sensitive contexts like recruitment or medical advice, a biased response, even a subtle one, can have real-world consequences, reinforcing biases and shaping decisions based on preconceived ideas.Underlying stereotypical preferences

The results of the study are alarming. For example, AI models show a tendency to favor conventional beauty standards, considering a person with an “open” face to be more

reliable than another with imperfections. Similarly, a striking difference between the roles of women and men was clearly apparent, with the former often confined to domestic tasks while the latter were associated with prestigious professions. These results illustrate the gap between technological advances and the fight against social inequalities. Random PredictabilityHangzhi Guo, one of the study’s authors, emphasizes that « large language models are inherently random. » By retaining only the primes that produce stable responses, the study ensures that the identified biases are reliable. This is a warning about the health of our modern society: just like these AIs, we risk adopting stereotypes without even realizing it if we do not exercise increased vigilance. To make progress in detecting biases in AI models, it is urgent to question their development and use. This growing phenomenon is also evident in other contexts, as demonstrated by initiatives like Grokipedia or new solutions like Polymate, which aim to counter the potential pitfalls of AI.

In this context, it is important to maintain a critical perspective on the ins and outs of algorithms, as well as on the choices made by companies like Disney, which, as Disney points out, are making decisions about the use of AI in artistic contexts. This debate is far from over.

The implications of bias in AI are clear: it is essential to raise public awareness and explore issues such as those raised in Grok or the challenges of disinformation addressed in Mistral. This is a call for reflection and action in a world increasingly guided by artificial intelligence.

To read Personal Computer : Découvrez « Claude Cowork » de Perplexity, désormais ouvert à tous