Generated with Adobe Firefly AIIn the last few years, editors of Sanskrit texts, and presumably also others, have been asked one question again and again. The question is a safe one, one can ask it without knowing too much about the topic on which one asks it, and the person asked will never want to indicate that this question was anything else than ingenious. The question can be asked by people in lectures who have only the slightest idea what the lecturer is talking about, but still one can be sure to appear capable of
asking relevant cutting-edge questions. Needless to say, it is the perfect question for assessing projects on any level and our Light on Haṭhayoga project was no exception.
The question can be phrased in different ways, but basically, it runs like this: “Would it not be better if you had relied more on digital tools?” It can also be phrased in a more sophisticated way: “Would the result not have been much quicker/better/more reliable/statistically valid/reproducible if you had used AI combined with computer stemmatics/large language models[…]?”
If the question is asked when you defend a multi-million project funding, giving a simple “no” in response is not advisable, even in cases where this would be the obvious answer. With a “no”, however heartfelt this answer may be, you are giving a whole set of wrong signals. And, of course, you cannot win the argument, even when you win in substance, since digital is better in an almost suprarational way. And if it is not yet, it will be, which places the burden of proof not on the
programmers who have not been able to do what expects them to be able to, but on the timing of the hapless researcher who should have waited until conditions would be ideal.
Another way one should not answer this question is by saying artificial intelligence and everything digital cannot compete with human intelligence or analog technique, for no one will believe you. The only excuse you can make is to say that digital is much better, but it is not yet ready right now. This is something everyone will immediately understand because it describes large segments of digital promises, from supposedly intelligent administration webpages to car software that can theoretically drive alone, but in actual life often seriously disturbs the driver or simply blacks out most of the time. It all depends on whether you read press releases by the companies selling the products, or critical reviews, or indeed use it yourself. An Indological example would be the OCR of handwritten Indian scripts, which has been promised for decades, but still produced too many mistakes to be helpful.
As a hobby musician, I should add that nowadays guitarists rediscover analog signal production for guitar amplifiers and start to see the tonal limitations of digital modeling or profiling amps, which have already passed their zenith. The natural reaction of many readers is surely the end of digital tone. Surely we have not yet reached the end of digital tone! Maybe so, but the musician who gets on stage and wants to sound good needs the ideal tool now, and this is often, simply because of its tonal quality, a tube amp, a technology so ancient that the components (i.e. tubes) have to be imported.
Of course, we would have preferred to produce an automatic digital analysis of our 200 manuscripts, which some questioners, who never had to try, must have thought to be within reach, but even ignoring the question of who would transcribe them in the time we had, we would have to solve the problem that present-day software is not able to deal with this amount of data.
How can this be? Are we not reading about the stunning advances of AI? See, for instance, the headline in nature.com “AI reads text from ancient Herculaneum scroll for the first time”? (or
https://www.heise.de/news/Verkohlte-Papyrusrollen-der-Antike-KI-entziffert-schon-ganze-Absaetze-9620197.html)
The crucial question in this type of research, one which was voiced in this contest, is: How can we know that AI has really recognized a word and not invented one? The phenomenon of AI hallucination in Large Language Models, which, according to some researchers, occurs more often than has been realized, should guard against placing too much trust in software that is like a black box, at least to those who did not program it. But even if AI can understand correct language, can it understand error and its motives, which are so important for criticism, and at what point would the program tell us why it selected this but not that? Should the AI write the philological commentary, for no one else will know why a reading was selected?
When it comes to textual criticism, I tend not to trust anyone and surely not any editor, this is precisely why one has invented critical apparatuses and a philological commentary, and this is why I have difficulties trusting software that works in the dark and presents its results without explaining how it came up with it. If this were not the digital age, we might even consider calling this approach philologically unscientific.
Our solution in the project, by the way, was not digital but pluralistic: we tried as many stemmatical programs as we could. Unfortunately, we could not follow the advice of one expert, who with good arguments thought that an old version of MacClade would be the best choice. However, unlike him, none of us had an old Macintosh computer that was still working (the program does not run anywhere else); we had to use what was working. In the end, we had as it were many digital
expert opinions produced by the various programs as well as our own observations, hopefully the best of both worlds, but at least the middle way avoiding the extremes.
