Have the user label highlighted objects in video that a state of the art classifier cannot solve
Create a state of the art video classifier. Might as well train it on Google's YouTube-8M video training data. But you will want to continually feed it original video as well.
Have the classifier label as many objects as it can. Have it isolate which objects it can recognize as objects but which it is unable to label.
Have it output videos that outlines the objects. Preferably GIFs, which can be easily embedded in forms.
For 100 of these, ask 100 users what the object is. If 90% of the users agree on the name of an object, add that video to the captcha-set. Call this the pre-trained set.
Every time a user needs to authenticate, show them one of the highlighted objects in a video not from the pre-trained set. If the image has less than 100 showings, record the label and give the user another one from the pre-trained set. If they get it right, let them through, if not, give them another from the pretrained set.
Once the non-pre-trained video has more than 100 showings and more than 90% of the captcha-users agree, add that video to the post-trained set.
Over time, slowly remove the pre-trained set. Put expirations on each video in the post-trained set and remove them after expiration, so that they don't get used too many times.
Ideally, this process would constantly improve the video classifier, keeping it state of the art and slightly ahead of other classifiers. Perhaps it could also favor less common words and objects and more esoteric things, so as to specialize this classifier against other classifiers.
The same could be done for image labeling, but the utility of the video classifier will probably last longer, given advances in AI.
Strictly speaking, though, short of some quantum trickery, there is no captcha system that will not one day be solved by external AI systems.
(edit: oh, I just noticed you specifically said "textual captcha." If that's what you mean, then no I don't think text classification has much mystery left in it. Computers can probably glean text from pictures better than humans now. But techically, the input in the above described captcha system is textual.)