ChatGPT: close, but no cigar

If there is one buzz word in the past three months, it must be ChatGPT.  This is the fastest, actively downloaded application ever, 1 million in 5 days, and 100 million in 2 months.  There have been AI tools, such as AI drawing, or even AI in photography for years now, but ChatGPT is different because it needs no training and is as simple as searching with google, while yielding answers more detailed, organized, and useful than google, in virtually every area.  Due to this unique nature, ChatGPT is going to be one of the watershed moments for AI in human history.

There have been a lot of hypes about ChatGPT, most justifiable.  Many educational institutions have now banned students using ChatGPT, since it will do their homework for them, often times quite adequately.  ChatGPT can give you answers to most if not all subjects, and those answers are at least plausible, in an intelligent and coherent manner.  Thus, people are crying that certain professions will lost all their jobs.  But this obviously depends on how good ChatGPT answers are.

At this moment, it is clear that ChatGPT is very good or proficient enough pertaining to questions with well defined rules or knowledge.  For example, coding, with well defined grammars and structures, and excel, which has finite formula and programming.  ChatGPT can also give you believable recipes, CVs, business plans, and steps on getting better at tennis, since there has been years and years of evolving yet consistent information available.

But ChatGPT is far from capable of doing a few other things, mostly involving creativity and new scientific knowledge.  ChatGPT may write you stories, even mimicking certain authors’ styles, which may be good enough for homework assignments or just amusement purposes, but the contents aren’t going to be anything special or usable for any novel publication.  Worse, if you ask scientific questions, not those kinds in the textbooks, but in active research areas, in hope for some literature synthesis and AI perspective, you will be utterly disappointed.  ChatGPT answers, language still presentable, are just a sequence of words that are generalized for similar question/answers but likely completely useless or wrong scientifically.  ChatGPT even made up references to support its answers: when I go to check its references, maybe 90% are nowhere to be found.  Some fake references don’t even support the claims by ChatGPT.  But more importantly, one has to wonder where ChatGPT gets these references from and how?  Does ChatGPT know how to cheat too?

So ChatGPT is not yet ready for prime time, and the current product is already showing worrying signs.  It will get better progressively, but whether it can truly overcome the above problems remains to be seen.  Another issue is: as there are fake news, there will be fake AI as well, “fake” in the sense that AI can be manipulated to be reliable on certain subjects but deceiving on others.  Fake AI will give fake answers that fool at least many people without the ability to know better or simply wanting to believe because AI looks so real!  A scenario is ChatGPT dominates the field, but it gives wrong answers while people act on them.  The other scenario is there are multiple AIs that feed diverging groups of users with different answers in certain subjects.  The same thing is already playing out daily, e.g., in the traditional and social media, and it is certain AI will be another, future divider.  How will AI such as ChatGPT will impact the world?  It won’t be tidy for sure.