The image generation portion of this is not the biggest long term problem because there genuinely very dumb. Good training data can mediate this a lot but more importantly.
Image generation does not reason like llms can.
Once the tech is properly matured where fine tuning of details is possible i expect a true llm reasoning component build in that will always specify to the image generation module exactly how the intended image is supposed to look. Including gender and age if those where not user specified.
This does not solve the problem of bias in llm but i want to highlight that the bias in llm-reasoning module of ai is the single most important part that needs to be bias aware, image generation will smooth itself out.
This is somewhat a reactionary rant on “researchers” and people addressing image gen and text gen under the same rules. And as a worst offender judge gpt models based on the output off dalle outputs. However faulty and hallucinatory they all are they are not the same thing.
Thanks for reading.
Congrats, you setup an algorithm that aims to trigger real people to behave a certain way by tricking their subconscious trough language.
You might call this business, i call it enshitification.
When people need an app they are able to look for a fit and compare optioned tailored to their needs, ideally using factual neutral information about similar apps.
Customers who are lured into clicking based purely on a ad are not making a deliberate decision for themselves and while you are not responsible for their actions you are preying on it.
To me, and this is an opinion: All software should Be open source and free, all software should exist to the benefit of all people and preferably be well documented, and public advertisement of any kind should be contained within very strict limits.