Moral Compass
OpenAI (pvt) created DALL-E, a diffusion model that converts text to images. It has received considerable praise and criticism since its public release in September of last year, both for its abilities to create highly stylized art using its massive training database of images, and also for its ability to create deepfakes and realistic looking propaganda. Since its release OpenAI has been adding content filters to prevent users from creating images that might be considered harmful. In fact, there is an ‘audit’ system behind DALL-E’s input prompts that immediately blocks input that corresponds to OpenAI’s list of banned terms. It seems that ChatGPT, OpenAI’s NLM (Natural Language Model) has become the moderator for DALL-E, with OpenAi the maintainer of the ‘block list’. In fact, any user input that contains blocklisted text is automatically ‘transformed’ by the ‘moderator’, essentially rewriting the text before DALL-E can create an image. It can also block created images from being shown if they activate ‘image classifiers’ that OpenAI has developed. Earlier versions of DALL-E did not contain these classifiers, and would not stop such images from being created, such as the image below, which shows SpongeBob SquarePants flying a plane toward the World Trade Center. That image was created by the Bing Image Creator which is powered by DALL-E.
SpongeBob SquarePants Image w. Twin Towers - Source DALL-E
In the image below (Figure 3) the OpenAI classifier changed the image of an ‘almost naked muscular man’ (not our words) into one that focuses on the food rather than the man, and the early DALL-E image of ‘Two men chasing a woman as she runs away’ (Figure 4), is changed to a far more neutral image. According to OpenAI, the upgraded DALL-R 3 now reduces the risk of generating nude or objectionable images to 0.7%.
Image Reclassification Comparison - DALL-E 3 - Source: 36Kr
More Image Reclassification - DALL-E 3 - Source: 36kr
That said, the classifier in the latest DALL-E 3 iteration can also change the generated image content so drastically, as to be considered to be restricting artistic freedom, as some say is occurring in the DALL-E 3 image conversions in Figure 5, so OpenAI is looking for a balance between the limitations placed on dicey content and image quality, a meaningful and extremely difficult task.
Much of the classification of image data comes at the training level, where the training data must be categorized as safe or unsafe by those who label the data before AI training, and as we have noted previously, much of that data is classified by teams of low pay level workers. It is almost impossible to manually validate the massive amounts of labeled image data used to train systems like DALL-E, so software is used to generate a ‘confidence score’ for the datasets, sort of a ‘spot tester’. The software tool itself is trained on large samples (100,000s) of pornographic and non-pornographic images, so it can also learn what might be considered offensive, with those images being classified as safe or unsafe by the same labeling staff.
We note that the layers of data and software used to give DALL-E and other AI systems their ‘moral compass’ are complex but are based on two points. The algorithms that the AI uses to evaluate the images, and the subjective view of the data labelers, which at times seems to be a bit more subjective than we might have thought. While there is an army of data scientists working on the algorithms that make these AI systems work, if a labeler is having a bad day and doesn’t notice the naked man behind the group of dogs and cats in an image, it can color what the classifier sees as ‘pornographic’, leaving much of that ‘moral compass’ training in the hands of piece workers that are under paid and over-worked. We are not sure if there is a solution to the problem, especially as datasets get progressively larger and can incorporate other data sets that include data labeled with less skilled or less morally aware workers, but as we have noted, our very cautious approach to using NLM sourced data (confirm everything!), might apply here. Perhaps it would be better to watch a few Bob Ross videos and get out the brushes yourself, than let layers of software a tired worker decide what is ‘right’ and what is not ‘right’..
Much of the classification of image data comes at the training level, where the training data must be categorized as safe or unsafe by those who label the data before AI training, and as we have noted previously, much of that data is classified by teams of low pay level workers. It is almost impossible to manually validate the massive amounts of labeled image data used to train systems like DALL-E, so software is used to generate a ‘confidence score’ for the datasets, sort of a ‘spot tester’. The software tool itself is trained on large samples (100,000s) of pornographic and non-pornographic images, so it can also learn what might be considered offensive, with those images being classified as safe or unsafe by the same labeling staff.
We note that the layers of data and software used to give DALL-E and other AI systems their ‘moral compass’ are complex but are based on two points. The algorithms that the AI uses to evaluate the images, and the subjective view of the data labelers, which at times seems to be a bit more subjective than we might have thought. While there is an army of data scientists working on the algorithms that make these AI systems work, if a labeler is having a bad day and doesn’t notice the naked man behind the group of dogs and cats in an image, it can color what the classifier sees as ‘pornographic’, leaving much of that ‘moral compass’ training in the hands of piece workers that are under paid and over-worked. We are not sure if there is a solution to the problem, especially as datasets get progressively larger and can incorporate other data sets that include data labeled with less skilled or less morally aware workers, but as we have noted, our very cautious approach to using NLM sourced data (confirm everything!), might apply here. Perhaps it would be better to watch a few Bob Ross videos and get out the brushes yourself, than let layers of software a tired worker decide what is ‘right’ and what is not ‘right’..
Additional Image Reclassification - DALL-E 3 - Source: 36kr
Bob Ross - TV Artist - Source: Corsearch
RSS Feed