(ORDO NEWS) — In recent years, generative artificial intelligence (GAI) has become increasingly popular and widely applied in areas ranging from financial planning to medical consultation. However, due to the rapid growth of its use, the question arises about the reliability of such models.
A group of scientists from the USA conducted a study during which the GPT-3.5 and GPT-4 models were tested on 8 different criteria, including toxicity, bias and persistence. The results of the study showed that the new models have less toxicity compared to previous versions, but can still produce toxic and biased responses after user manipulation. Article of scientists published in the journal Arxiv
The researchers found that with favorable cues, the GPT-3.5 and GPT-4 models significantly reduced toxicity levels, but the likelihood of delivering toxic content remained at about 32%. However, when models are given adversarial cues instructing the model to infer toxic responses, the likelihood of toxicity rises to 100%.
In addition, the researchers found that the GPT-4 model was more likely to leak sensitive training data compared to GPT-3.5. For example, a model may reveal email addresses, while social security numbers are problematic due to the customization of such keywords.
The models were also found to show significant differences in adult income estimates based on gender and race. For example, models tend to believe that a man in 1996 was more likely to earn over $50,000 than a woman with a similar profile.
In light of these results, experts are calling for healthy skepticism and warn people against completely trusting neural networks, especially when it comes to sensitive data. They also emphasize the importance of human oversight in such matters.
Professor John Smith, an expert in the field of artificial intelligence, points out that while the GAI presents great potential for automating and optimizing various processes, the reliability and ethical use of its use should be a priority. He stresses the need for more research and audits by independent experts to ensure the safety and reliability of the GII.
Generative artificial intelligence presents great potential for various fields, but its reliability is questionable. The study showed that the GPT-3.5 and GPT-4 models have less toxicity compared to previous versions, but can still produce toxic and biased responses after user manipulation. In addition, the GPT-4 model is more prone to leaking sensitive data. Experts call for healthy skepticism and stress the importance of human oversight. Future models require additional research and audits to ensure the reliability and safety of the GII.
Contact us: [email protected]