It has often been pointed out that ChatGPT can produce incorrect text: it will output incorrect dates, assign people to incorrect occupations or disciplines, create fictitious academic citations, etc. Because of this, it may be easier to share DALL-E with people trying to understand these technologies. When DALL-E produces an image, we don’t generally focus on whether its image is “right or wrong.”

ChatGPT and DALL-E work on technically similar underlying foundations (“generative” machine learning models). Briefly, this class of AI tools uses source data culled from the Internet (ChatGPT’s source data are text; DALL-E’s source data are images), then produces an original output (ChatGPT produces text; DALL-E produces images). The mind boggling thing to know is that, as with DALL-E, ChatGPT  is a machine that automatically creates unique, original work.

However, assessing ChatGPT with a “right vs. wrong” lens misses out on what it is doing. At this stage of its development, ChatGPT’s aim is to produce human-like text. Getting things “right” is not necessarily its primary goal.

The technical marvel is as if your sewing machine suddenly start mopping the floor. While it may not be very good at mopping the floor (yet), the fact that the machine does it at all is the thing to marvel about because the machine has learned to do something without you teaching it what to do:

large-scale language models trained on large amounts of multi-lingual data and source code are capable of improving the state-of-the-art across a wide variety of natural language, translation, and coding tasks, despite never having been trained to specifically perform those tasks (From: “Google Research, 2022 & beyond: Language, vision and generative models” – “Google Research, 2022 & beyond: Language, vision and generative models” – Emphasis added.)

ChatGPT does not understand what it produces: it can make a joke, but does not understand humor. Thus, focusing on “right vs. wrong” is looking past its achievement. Imagine the following verse never existed and it came out of ChatGPT:

Hey diddle-diddle
The cat and the fiddle,
The cow jumped over the moon.

Overlooking what ChatGPT is doing might lead someone to say, “Well, that verse is wrong because cats can’t play the violin and cows cannot jump over the moon.” At this stage of the technology, ChatGPT is not intended to be factually correct: we should instead be impressed that–given a prompt–a machine can create new, original, human-passable text.

U Wisconsin describes this distinction: “ChatGPT can write with correct grammar and confident flow, but cannot create accurate citations or write with much depth. This is because it creates word patterns, with some variability or randomness, but does not generate meaning (Warner, 2022).” (https://idc.ls.wisc.edu/guides/using-artificial-intelligence-in-the-classroom/)

With time, this will all change. These technologies are changing rapidly. Google and Microsoft are building generative AI into their search engines and in the near future, as their aims of factual accuracy improve and their search engines provide more conversational output, holding the companies responsible for errors will make more sense.