bloxyen in  
Software Engineer  

OpenAI’s GPT-4 With Vision Still Has Flaws, Reveals Paper

OpenAI's much-touted model GPT-4, lauded for its multimodal abilities, including advanced image recognition, still has significant flaws. These glitches range from inventing facts to misinterpreting chemicals' images and hate symbols, according to a new paper from OpenAI.

To stay ahead of AI developments, look here first.

Unintended GPT-4V behaviors

  • GPT-4V has a tendency to hallucinate or invent facts with unwarranted confidence.
  • The model struggles to make correct inferences, sometimes creating fictional terms by wrongly combining text strings.
  • It misinterprets certain symbols of hate and can give incorrect answers in the context of medical imaging.

OpenAI’s mitigation strategies

  • OpenAI has implemented various safeguards to prevent GPT-4V's misuse, such as breaking CAPTCHAs or using images to infer personal details.
  • The company insisted that GPT-4V is not to be used for identifying dangerous chemicals from image structures.
  • OpenAI acknowledged it has a long way to go in refining the model and is working on it.

Discrimination and bias

  • When OpenAI’s production safeguards are disabled, GPT-4V displays bias against certain sexes and body types.
  • The paper reported offensive responses related to body positivity when prompted by an image of a woman in a bathing suit.


P.S. If you like this kind of analysis, I write a free newsletter that dissects the most impactful AI news and research. 1000s of professionals from Google, Meta, and OpenAI read it daily.