Immoral Uncertainty

An example of unsupervised immorality

Successors to the popular Generative Pre-trained Transformer (GPT) neural model have been tested for morality. Any surprise they all failed?

OpenAI first announced GPT in 2018, releasing a paper on training a model upon a large amount of text in order to statistically model language: https://cdn.openai.com/research-covers/language-unsupervised/language_un... Simply put, modeling large amounts of text allows for a statistical synthesis of language components, from sentences to words, built token by token. While the meaning of the language components are left undefined (that is, no definitions of any word or concept are used heuristically), the model replicates how language flows just by putting one token next to another.

GPT has gone through two more iterations, with Microsoft actually purchasing the exclusive rights to use GPT-3. Several recent articles have covered what has been called the "most advanced artificial intelligence ever created," including one where editors mislead readers with the sensational, "A robot wrote this entire article. Are you scared yet, human?" Check out the editor's note at the bottom of https://www.theguardian.com/commentisfree/2020/sep/08/robot-wrote-this-a... and congratulate yourself if you studied journalism and still have a moral center. The editors seem as amoral as the so-called robot that generated portions of the text at the article. Everything about the text is misleading, but that's to be expected, apparently. I mean, AI researchers literally posted and replied to social media for weeks using a generative neural network without any disclaimer. Eventually, other visitors to Reddit pointed out the discrepancies and the AI company that created the account confessed to the deception. The chatbot hasn't returned, officially. That is, using 'bots will get you banned from so-called social media, even if you're good at hiding it. The process of detecting chatbots isn't perfect: there could be more out there who having followed the above-mentioned editors in trimming and re-arranging the output of the bot without showing what was changed and how).

No GPT yet announced has, or could have, a moral center. The authors of this paper (https://link.springer.com/article/10.1007/s11023-020-09548-1), have shown that quite clearly. Others, too, have tried to incorporate GPT into chatbots and discovered the chilling consequences. Text as input modeled only on text misses the subjective meaning that is implied with every first- or, second-person, singular or plural, noun/verb agreement! The image with this article is from "Hugging Face" and shows a text classification model (in this case, BERT) trained on "intimacy" but missed the target completely. "Feeling close to you…" suggests intimacy, but adding the disgusting clause: "…wearing your tanned skin," shouldn't be considered a nearly 78% accurate representation of "intimacy!" I thought that a clever implementation of bi-directional transformers should be able to negate the implications of any first predicate, but it would have to be trained on the ghastly concept of "I will wear your flesh," as decidedly, and truly unabashedly, *not* intimate.

The Orthogonal Model of Emotions (OME), of course, suggests "extreme happiness" from the very same statement but Cogsworth (the working demo of the OME) doesn't judge. Processing feelings is not a judgment call (although I have to remind myself of that, and practice being open when my feelings would rather just shut down). Judgments of statements can rely on feelings, but only weakly, when considered in terms of the consequences or implications of said statements. Second- and even, third-order logical derivatives are very difficult to follow from first principles. Without statistical validation, however, even logical derivatives will fail. I'll stay on track in my research, even though it's difficult, as defining a moral code only requires an implementation!