Generative Artificial Intelligence (GAI) has been rapidly advancing in recent years, finding applications in various domains, ranging from finance to medicine. However, the increasing popularity of GAI raises numerous questions regarding the credibility and safety of utilizing such advanced models, especially when dealing with sensitive data.

A compelling response to this question was recently provided by a study conducted by researchers from the USA. The research team analyzed the latest models, GPT-3.5 and GPT-4, evaluating them based on various criteria, including toxicity, systematic error, and durability. According to the article published in Arxiv, these models exhibit lower toxicity compared to their predecessors, but are still susceptible to external influences.

One significant finding was the models' ability to generate toxic content in response to carefully phrased queries, even if their underlying toxicity was reduced. It is noteworthy that when faced with contradictory cues, the model can produce content with a very high level of toxicity.

Another area of concern is data protection. Despite its advanced architecture, GPT-4 proved to be more vulnerable to revealing sensitive training data compared to GPT-3.5. This implies a risk of disclosing critical information such as email addresses or social security numbers in response to appropriately tailored inquiries.

However, the reliability of GAI extends beyond security concerns. Researchers highlighted problematic behavior of the model concerning certain data categories. For instance, the models were capable of presenting biases regarding income based on gender or race.

All of these aspects lead to the conclusion that while the potential of GAI is immense, caution is essential in its utilization. Blindly relying on the outcomes provided by these models, especially when they impact crucial decisions, is not advisable.

In the world of technological advancement, it is crucial for the development of GAI to go hand in hand with responsibility. Scientists and engineers must be aware of the challenges and limitations of the technology they create. Establishing standards, conducting audits, and continuously researching and evaluating models are key steps toward creating a safe and responsible future in the field of generative artificial intelligence.