"data ai" by 4X-Image/Getty Images Signature
Sanjiv R. Das is the William and Janice Terry Professor of Finance at Santa Clara University's Leavey School of Business and a faculty scholar with the Markkula Center for Applied Ethics. Views are his own.
On November 30, 2022, ChatGPT was born, in the aptly named Pioneer Building in the heart of San Francisco’s Mission District. OpenAI’s newest prodigy has surprised almost everyone it has met, dropping random pearls of wisdom in long-winded and immaculately correct English. It is the popular new kid on the block, with its high IQ and EQ. ChatGPT took just five days to reach a million users. In contrast, other services took much longer (Instagram 2.5 months, Spotify 5 months, Facebook 10 months, Netflix 3.5 years). So much has been written by ChatGPT, much more even than has been written about it, so much so that anything written here is likely to be redundant. Such caveats aside, here is a technical and emotional perspective on the ethics of ChatGPT and its effect on society, and more broadly, the huge impact of generative AI.
Large language models (LLMs) achieved a watershed moment in 2022, when they started to write grammatically correct and lengthy text in multiple languages, at college level (this is freaking out many academics, but that is another debate altogether, more about writing assessment than writing itself, which is an obvious ethical issue). I have jokingly noted that these LLMs performed better at making up stuff than the below-average student who comes unprepared for an exam. Passing off an AI’s work as your own is a new take on plagiarism! A leading conference on machine learning (ICML) has decried its own subject technology by explicitly qualifying its use as follows: “Papers that include text generated from a large-scale language model (LLM) such as ChatGPT are prohibited unless the produced text is presented as a part of the paper’s experimental analysis.” But the ICML program committee omitted to say that code generated by ChatGPT is included in this prohibition. Or did they deliberately not mention it? Programmers have been cutting and pasting code for years with no ethical ramifications, yet passing off text as their own is a no-no, suggesting that the domain in which an AI operates drives ethical context. In this the AIs are no different than we are; our ethics vary across domain and geography, as do our value systems.
These models long supported language classification and various other tasks in natural language processing (NLP), such as summarization, question answering, sentence similarity, disambiguation, analogy tasks, etc., and easily surpassed human levels of accuracy on benchmark tests such as GLUE and SQUAD. Still, beating humans at their own word games doesn’t make a machine feel human. However, talking/speaking (Meta), writing (OpenAI), and drawing/painting (Stability, Midjourney) are attributes of algorithms that have evolved to suddenly feel innately like we are interacting with a person. Without being both lovable and scary, like us, these AIs wouldn’t be human, and thus, this latest breed of language bots has charmed and alarmed their way into our consciousness. In terms of a paradigm shift the LLM phenomenon is entirely comparable to the advent of the internet. But as machines become more human-like, do we perceive the ethical issues to be more acute. Does the level of emotional engagement suggest a new angle on ethics in the way we view deception by a conman, connecting with a new literature on the ethics of emotion in AI?
The hype around ChatGPT has created an aura around it as the One. Its many siblings have gone unnoticed; though strong, they have been mostly silent. We’d do well to pay attention to them, such as Meta’s BlenderBot, Google’s LaMDA, AI21’s Wordtune, Anthropic’s Claude; Character.ai, DeepMind’s Sparrow, etc., to name a few. These are released as real products, whereas ChatGPT was released as a prototype aiming to collect research feedback, not intended to be the wunderkind it has become (the paid version finally launched in February, 2023, for $20/month). Indeed, human response has been ecstatic, delaying GPT-4 as OpenAI adjusts to feedback and deals with the technical challenges of a cult following. This demonstrates the power of making its product freely accessible even for limited usage to a wide audience, something the siblings did not pursue, and so, quantity trumps quality. Anthropic’s conversation bot Claude gives extensive answers in comparison to ChatGPT, and is by many counts superior and more accurate, based on an improved neural architecture, Constitution AI. Google’s LaMDA is trained on real conversations and comparable in performance, and likely to be more current than ChatGPT, whose training corpus may be aging. Future incarnations of LaMDA may access the latest information on the fly, as does Sparrow, and it is rumored that new versions will be way better. Wordtune provides explanations and facts behind its writing, thereby providing more grounded text, and ethical foundation (AI explainers and fairness checkers are one of the important checks and balances in machine learning, see the book “The Ethical Algorithm.”) And Character’s tens of thousands of chatbots crowdsourced and designed by developers across the globe makes for multiple personas, though none can write code as well as ChatGPT. These recent LLMs have shown that the early use cases are in the realm of chatbots, writing coaches, and programming assistants. Maybe ChatGPT will become a better therapist than the other wellness bots (such as woebots) which are raising ethical concerns? Bots that are trained to impersonate dead people, so called deathbots, suggest even more complex ethical issues.
Many of these applications are quite distinct from search, where deterministic accuracy has value in comparison to the stochastic responses from GPTs, rendering them quite inert in being a threat to search engines and much more of a complement, yet the threat story has elicited wide debate. To fix ChatGPT’s obsolescent information and its hallucination problems, we need to hook it up to a search engine, the very artifact claimed to be at risk; gosh, we still need Google. Anyway, let’s await the next incarnations, for who knows, they may even be sentient, or at least remind us of our own sentience (sometimes, it’s nice to just let go and let the waves of hype wash over you). Maybe sentience is not needed, if humans were not sentient, would we even need ethics!?
It is when AI touches the common man in a visceral way that it transcends from being software into almost wetware, maybe with a different set of ethical norms. Like any emerging technology, it is good and bad in the large. But now it is a social technology, which makes it more interesting. Unlike any other technology, it seemingly surpasses our own brains, in capacity and capability, because we extrapolate superior performance on a few tasks to all. As of late 2022, the size of Wikipedia is around 57 million pages, and requires storage (compressed) of about 22 gigabytes (excluding images). It is astounding that the storage required for the GPT-3 model (with 175 billion parameters) is 800 gigabytes. Even allowing for an uncompressed version of Wikipedia of 10x in size, we approach just a quarter of the memory of GPT-3, which makes it “super-encyclopedic” in dimension. It can therefore retain and generate “knowledge” at an astonishing scale. GPT-4 will be even more immense, and the (almost surely false) rumors of its sizing at 100 trillion parameters signal how huge people’s imaginations have become. Research establishing the emergent abilities of LLMs is evidence that a critical size of billions of parameters was needed to bring machines close to human language abilities. Size matters, especially when comparing brains! But humans still have an edge, our wetware is estimated to have a storage capacity of 2.5 petabytes (1 petabyte is 1024 terabytes, and 1 terabyte is 1024 gigabytes, so we may easily work out the comparison with GPT-3 using our massive brains!). Viewed at these scales, ChatGPT is a cretin, and we might temper our awe about it.
Immaturity issues abound with ChatGPT, making it an unreliable assistant. Without human interaction it is nothing but a random generator of words, an upgrade from spellchecker to writing assistant, at huge technical and environmental cost, with ethical issues around corporate decisions that impact future generations. Is it really artificial intelligence or merely artificial smartness, faking it but not making it? More entertaining than illuminating, generative AI is still going to school, and needs to learn continually and update itself, leading to several interesting problems in computer science, such as continual learning, optimal vs catastrophic forgetting, and self-supervised learning, known as the “dark matter of intelligence.” This world gets even more technically exciting as we move from generative AI to generalist AI, with models like DeepMind’s Gato that can perform hundreds of tasks, like writing, playing games, etc., all deployed already. And when these become more adept and less resource-greedy, we will train them ourselves, triggering our child-rearing instincts, for we want nothing more than to create machines in our own image (Genesis 1:27). Given the impact of social media on humanity, we might brace ourselves for the impact of generative AI.
Or, maybe ChatGPT isn’t all that it’s cracked up to be. Maybe it's just too young, expensive, wasteful, racist, and doesn’t really understand language because it went to the school of rote learning (as many of us did). It has no concept of time and it stopped learning more than a year ago. In other words, it is a dropout, though its cousins continue to persevere academically. Who cares? It’s big and cute, and we know well that size and beauty matter! It even engenders trust, in settings where mistakes are cheap. But it may also create misplaced trust in systems that impact economic and political outcomes. If it turns into Frankenstein, how does society ethically regulate or terminate it, especially if this cat outgrows its bag?
After all, what has made AIs such as ChatGPT household icons (and possibly harmful) is not that they are bigger and smarter than humans at limited tasks, but it’s the emotional response they trigger, as they reach out to our humanity and try to earn our respect, enhanced with cuteness as a factor in this new age beauty contest. They tug at our heartstrings despite lacking anthropomorphic form, unlike humanoid robots that certainly evoke human emotions. It is hard not to marvel at Character.ai’s landing page, with bots of all personalities with designer icons, created by thousands of people in an attempt to foster their AI progeny, as we attempt to create descendants of our intellectual gene pool, a whole new generation, say Gen-AI? ChatGPT is almost human in its fallibility, as it often says the wrong thing, while straining to remain factually correct and humanly sensitive, while its creators look on with embarrassment, and attempt to fix these flaws with personal interaction, using the paradigm of reinforcement learning with human feedback (aka RLHF). Who knows, we might test GPTs only to find that they are on the spectrum. Maybe that would enhance their humanity and ours, bringing a new angle to the growing field of affective computing.
What ChatGPT has convincingly done is unleash and unlock writing ability for many who struggle with it, enabling us to express ourselves better, for better or for worse. GPTs will not just impact language ability and use, but society at large. How much time should we spend with this teenage writing savant evolved from a spell checking, sentence-completion toddler? While it took me a few hours to write this, I am keenly aware that ChatGPT would have generated the same number of words in a few seconds, maybe on topic, with ideas I totally missed. Hopefully, ChatGPT cannot mimic my laborious writing style and steal my individuality, though I can always “fine-tune” it to do so. Eventually, I had to restrain myself from engaging ChatGPT to help, and maybe that was a mistake (I’ll let you decide).