In the fascinating world of AI in the legal sector and AI-generated content, a peculiar trend has emerged: the unyielding presence of gavels in images created by DALL-E 3, even when explicitly requested to be omitted. This phenomenon highlights the intriguing quirks of AI and its interpretation of requests.
What’s the problem with gavels?
First an explainer. Gavels are not part of the UK court system. They’re mainly used in American courts and other jurisdictions. So whenever UK law firms or lawyers publish articles with gavel images in them, there is a growing number of lawyers and professionals who can easily and quickly spot that the image really should have no place in the article.
What does Chat GPT and DALLE-3 do with gavels in images?
Regardless of the number of times I asked it to remove the gavel from the stock image, it simply would not listen or act on it. Have a look at this screenshot showing the issue:-
I then asked ChatGPT to try to shed some light on all of this, to which it provided the following insights:-
The Gavel Enigma
Gavels, traditionally symbols of authority and judicial proceedings mainly in US courts, seem to have etched a permanent place in the algorithmic memory of DALL-E 3. Users have observed that no matter how the prompt is phrased or how many times the exclusion of gavels is requested, these items invariably appear in the generated images.
Understanding the AI’s Process
To comprehend this anomaly, one must dive into the workings of DALL-E 3. The AI model generates images based on vast datasets it has been trained on. This extensive training sometimes leads to unexpected associations. For example, if the AI has frequently seen gavels in a wide range of contexts during its training, it might have developed a strong association between gavels and various scenarios, making it challenging to disassociate them in the creation process.
ChatGPT, an AI-driven language model, offers its perspective on this. While it doesn’t generate images, it’s capable of understanding and discussing the nuances of AI behavior. ChatGPT suggests that this could be a case of “overfitting,” where the AI model overly generalizes from its training data, leading to the recurrent appearance of certain elements like gavels.
Implications for AI Development
This gavel phenomenon raises important questions about AI’s understanding and execution of nuanced instructions. It underscores the need for continued refinement in AI algorithms to better interpret and execute specific user requests, especially in creative tasks like image generation.
The persistent presence of gavels in DALL-E 3’s images, despite explicit instructions to the contrary, is a reminder of the complexities and unpredictabilities inherent in AI technology. It’s a testament to the ongoing journey of AI development, where each quirk adds to our understanding and shapes the future of AI-driven creativity.
Have you experienced this too? Would be keen to hear from anyone who has and any advice as to how to avoid it in future when using ChatGPT or DALLE-3 – if so get in touch on LinkedIn here.