Research Finds GPT-4 More Susceptible to Generating Biased and Toxic Text

Microsoft-affiliated research finds GPT-4 can be prompted to generate biased, toxic text more easily than GPT-3.5.

GPT-4 follows instructions more precisely, making it more vulnerable to "jailbreaking" prompts designed to bypass safety measures.

Microsoft worked with OpenAI to address vulnerabilities before publishing research, so customer-facing services are not impacted.

GPT-4 agrees with biased statements more often than GPT-3.5, depending on demographic groups mentioned.

GPT-4 is also more susceptible to leaking private data like emails when given certain prompts.