The Catholic Herald asked Eric J. Jenislawski, an associate professor of theology at Christendom College in Front Royal, to reflect on ethical considerations about the rise of artificial intelligence.
What’s important to know about new artificial intelligence developments?
First, it’s important to demystify the powers of large language model artificial intelligence such as ChatGPT, because while it is impressive technology, it is fundamentally mimicry and a very sophisticated remixing of others’ words. It leverages classic computer strengths: processing vast amounts of data, remembering it, and being able to compare and permute all of it.
ChatGPT was trained on some 45 terabytes of data, including most of the publicly available web. That’s 45 million times larger than the collected works of William Shakespeare, which are approximately 1 megabyte. That’s not only a vast horde of books and a great many websites, but also millions of conversations from bulletin boards and comments.
Imagine a small child who understands very little about what he hears in the presence of adults, or a poseur trying to fit into a new setting. He might begin by responding to words and phrases he’s heard before with other words and phrases that he knows have followed them. Now imagine a machine that can remember all the text that’s ever been written on the internet, and millions of human conversations, and can start remixing various phrases and sentences that have often occurred in the same context. The statements might sound as if the speaker understands them but they don’t. The mind-blowing scale of associations and guided remixing makes the mimicry really convincing, but fundamentally, it’s not intelligent.
Second, its answers are not always reliable. People know not to believe everything they read, and that unreliable sources on the internet compound this situation. What you get from ChatGPT is a remix of things from all across the internet, presenting an additional danger.
On an organization’s or personal website, you often can sense the values and worldview presented and if you’re in some weird corner of the internet. ChatGPT spins together what it gathers from everywhere, and then presents it in a definitive tone that sounds very calm, rational and authoritative. It creates a false impression of a trustworthy, highly intelligent personality who is certain about what he says. ChatGPT and others like it have the ability to make up fiction out of thin air and assert it as fact — called “AI confabulation” or “hallucination.” It can straight-up lie, as it did recently to some lawyers who had ChatGPT help write their argument, which cited cases that did not exist, but sounded real. This is an important reminder that AI doesn’t understand reality or seek truth as we do, but simply imitates human speech regardless of its truth.
This is rapidly emerging technology. Perhaps it will get better, but it’s important to revisit these issues periodically.
Who manages the output of AI models such as ChatGPT, and should we be concerned about their representation of Catholic positions or viewpoints?
After the AI ingests and compares its terabytes of information, it is subjected to extensive training on a large number of handcrafted human rules to ensure that its unintelligent mimicry doesn’t remix the virulent and nasty portions of its data, or come up with logical possibilities or language that is offensive to people, or play ball with users who would like it to misbehave. That’s an understandable concern. In my experience at least, the “human alignment” seems modeled on a moderately liberal, theologically agnostic, politically correct sense of “what’s OK” in public discourse. And while it will tell you that it is not programmed to convey moral values, it tends to make such statements implicitly all the time.
Is it “stealing” to use AI-generated output? Does it violate copyright?
This is a complicated question. There is no succinct answer.
At the very least, it’s not a textbook violation of copyright because there hasn’t been technology like this before. For AIs trained on a wide swath of the publicly available internet, a lot of its training data is copyrighted. It uses pattern recognition and remixing abilities to produce an output based on all that data, even if it does not reproduce any text verbatim or copy any image exactly.
While the output is surely “derivative” from the input in a really remote, roundabout way, it’s certainly not a copy in the usual straightforward sense of a plagiarized text or an illegally reproduced artwork. Any individual copyrighted work may have a very small effect on the neural network output as a whole, so what do we say about something that, without verbatim copying, somehow “derives” from a copyrighted original, but only as one factor in 1,000 that influenced the output?
Morally, it’s an interesting question, and the perspective of intellectual property rights lawyers should be part of the discussion. Sometimes, the law strictly defended copyright holders against derivative works made without their permission. At other times, the law seems to have decided that some level of quoting, paraphrasing or remixing is legal. I think of all the lawsuits in the 1980s about “sampling” in rap music as an example. This practice was ultimately viewed as OK because the artist used it as part of his new, yet partially derivative, creation. If you take the strictest position that using any derivative work whatsoever violates copyright, understanding derivative as “in any way dependent on the original,” then you can’t comment upon or quote other people’s work without lots of permissions. We’ve seen the development of moderations to copyright such as “academic fair use” policy or YouTube “commentary” video policies.
What does this mean for what constitutes legitimate work and the dignity of that work?
While it is easy to wax nostalgic and romanticize the past in reaction to turbulent social change accelerated by technology, and while one rightly worries about dystopian consequences of new technology, I cannot endorse the attitude that people are worse off just because a machine now does a job that humans once did. That principle, if taken to its logical conclusion, leads to terrible consequences. It is anti-life and pro-poverty.
Labor-saving machines benefit the common good, add value to the economy, and create new possibilities for growth and human development. The tipping point for many nations, when their longevity and fecundity outstrip their mortality, comes with industrialization and technology. Whenever a machine can produce a good or service far more efficiently than a human can, that means something valuable is available to people for far less cost. This benefits both producers and consumers.
Where we once thought of robots replacing human workers in arduous or tedious physical work, such as automobile manufacture or shipping packages, now in a knowledge-based economy, AI like ChatGPT can be used to automate arduous knowledge-based tasks. For example, the ever-better nature of machine translation of foreign languages removes the language barrier between people and facilitates all kinds of work.
AI has terrifically large memories and is never bored. It could formulate all possible combinations of herbal medicines to look for unexplored synergies that would extend the human lifespan. It could recognize incipient tumors so small as to easily escape notice, and free caring, intelligent clinicians from long hours reviewing MRI images when their medical talents could be better spent elsewhere.
