Bottom comment is technically correct, you can bypass any dataset-related ethical concerns. You could very easily make a classifier model that just tries to find age and combine it with one of many existing NSFW classifiers, flagging any image that scores high for both.

But there are already organizations which have large databases of CSAM that are suitable for this purpose, using it to train a model would not create any additional demand, it would not result in any more harm, and it would likely result in a better model. Keep in mind that these same organizations already use these images in the form of a perceptual hash database that social media and adult content sites can check uploaded images against to see if they are known images while also not sharing the images themselves. This is just a different use of the data for a similar purpose.

The only actual problem I could think of would be if people trust its outputs blindly instead of manually reviewing images and start reporting everything that scores high directly to the police, but that is more of a problem with inappropriate use of the model than it is with the training or existence of the model. It’s very safe to respond like this to images flagged by NCMEC’s phash database because those are known images and if any false positives happen they have the originals to compare to so they can be cleared up, but even if you get a 99% accurate classifier model, you will still have something that is orders of magnitude more prone to false positives than the phash database, and it can be very difficult to find out why it generates false positives and correct the problem, because… well, it involves extensive auditing of the dataset. I don’t think this is enough reason to not make such a model, but it is a reason to tread carefully when looking at how to apply it.