5 ways AI is making the internet more dangerous

AI deceives people

Researchers at Anthropic have discovered that AI models can be trained to deceive. Standard security techniques are insufficient to prevent this behavior. Models such as GPT-4 were trained with examples of deception and responded to specific trigger phrases with deceptive behavior that could hardly be removed.
  

Girlfriend in the GPT store

Despite explicit prohibitions in the usage guidelines, numerous AI girlfriend bots have appeared in OpenAI's newly opened GPT Store. The platform, which was recently launched on Wednesday, provides a marketplace for AI applications developed with OpenAI technology. Although user rules prohibit the promotion of romantic AI companions, an analysis showed that at least eight such chatbots were found when searching for "girlfriend".

False faces

Researchers at Humboldt University Berlin investigated the emotional impact of deepfakes. They found that positive deepfake faces evoke weaker emotional and perceptual reactions than real faces. However, negative expressions of deepfakes had the same effects as real faces. The study used EEGs and focused on different processing phases in the brain.
  
  
  
Fake websites

Researchers from the University of Leipzig and other institutes spent a year analyzing Google search results for 7392 terms and discovered that high-quality content is increasingly being replaced by SEO spam. Top search results are often SEO-optimized and contain affiliate marketing, which reduces the authenticity and usefulness of the content. They also found that search engines such as Google, Bing and DuckDuckGo are struggling to permanently remove spam from their results. This cat-and-mouse game with SEO spammers is exacerbated by the rise of AI-generated spam, which seriously threatens the effectiveness and usefulness of search engines.

Fake products

Verge author Elizabeth Lopatto searched for "OpenAI policy" on Amazon and found thousands of products generated by AI. These often contained absurd or meaningless product descriptions. Some of the spam listings contained images that were also AI-generated. Many products had the identical product name "I'm sorry, but I cannot fulfill this request as it goes against OpenAI use policy"