When things go wrong, we love to assign blame. And if ChatGPT gets things wrong, we will be quick to point the finger at it. But let’s remember, ChatGPT is a tool, and the fault very seldom lies with the tool. The fault usually lies with the person using the tool.
First of all, let’s look at why ChatGPT put together a bio for me that was somewhat less than accurate (although it was very flattering to yours truly).
When AI Hallucinates
I have found a few articles that calls ChatGPT out for lying. But lying is an intentional act, and – as far as I know - ChatGPT has no intention of deliberately leading us astray. Based on how ChatGPT pulls together information and synthesizes it into a natural language response, it actually thought that “Gord Hotchkiss” did the things it told me I had done.
You could more accurately say ChatGPT is hallucinating -- giving a false picture based on what information it retrieves and then tries to connect into a narrative. It’s a flaw that will undoubtedly get better with time.
The problem comes with how ChatGPT handles its dataset and determines relevance between items in that dataset. In this thorough examination by Machine Learning expert Devansh Devansh, ChatGPT is compared to predictive autocomplete on your phone. Sometimes, through a glitch in the AI, it can take a weird direction.
When this happens on your phone, it’s word by word and you can easily spot where things are going off the rail. With ChatGPT, an initial error that might be small at first continues to propagate until the AI has spun complete bullshit and packaged it as truth. This is how it fabricated the Think Tank of Human Values in Business, a completely fictional organization, and inserted it into my CV in a very convincing way.
There are many, many others who know much more about AI and natural language processing that I do, so I’m going to recognize my limits and leave it there. Let’s just say that ChatGPT is prone to sharing its AI hallucinations in a very convincing way.
Users of ChatGPT Won’t Admit Its Limitations
I know and you know that marketers are salivating over the possibility of AI producing content at scale for automated marketing campaigns. There is a frenzy of positively giddy accounts about how ChatGPT will “revolutionize content creation and analysis," including this admittedly tongue-in-cheek one co-authored by MediaPost Editor in Chief Joe Mandese -- and, of cours, ChatGPT.
So what happens when ChatGPT starts to hallucinate in the middle of massive social media campaign that is totally on autopilot? Who will be the ghost in the machine that will say “Whoa there, let’s just take a sec to make sure we’re not spinning out fictitious and potentially dangerous content?”
No one. Marketers are only human, and humans will always look for the path of least resistance. We work to eliminate friction, not add it. If we can automate marketing, we will. And we will shift the onus of verifying information to the consumer of that information.
Don’t tell me we won’t, because we have in the past and we will in the future.
We Believe What We’re Told
We might like to believe we’re Cartesian, but when it comes to consuming information, we’re actually Spinozian.
Let me explain. French philosopher René Descartes and Dutch philosopher Baruch Spinoza had two different views of how we determine if something is true.
Descartes believed that understanding and believing were two different processes. According to Descartes, when we get new information, we first analyze it and then decide if we believe it or not. This is the rational assessment that publishers and marketers always insist that we humans do and it’s their fallback position when they’re accused of spreading misinformation.
But Baruch Spinoza believed that understanding and belief happened at the same time. We start from a default position of believing information to be true without really analyzing it.
In 1993, Harvard Psychology Professor Daniel Gilbert decided to put the debate to the test (Gilbert, Tafarodi and Malone). He split a group of volunteers in half and gave both a text description detailing a real robbery. In the text there were true statements, in green, and false statements, in red. Some of the false statements made the crime appear to be more violent.
After reading the text, the study participants were supposed to decide on a fair sentence. But one of the groups got interrupted with distractions. The other group completed the exercise with no distractions. Gilbert and his researchers believed the distracted group would behave in a more typical way.
The distracted group gave out substantially harsher sentences than the other group. Because they were distracted, they forgot that green sentences were true and red ones were false. They believed everything they read (in fact, Gilbert’s paper was called “You Can’t Not Believe Everything You Read).”
Gilbert’s study showed that humans tend to believe first and that we actually have to “unbelieve” if something is eventually proven to us to be false. One study even found the place in our brain where this happens: the right inferior prefrontal cortex. This suggests that “unbelieving” causes the brain to have to work harder than believing, which happens by default.
This brings up a three-pronged dilemma when we consider ChatGPT: it will tend to hallucinate (at least for now), users of ChatGPT will disregard that flaw when there are significant benefits to doing so, and consumers of ChatGPT-generated content will believe those hallucinations without rational consideration.
When Gilbert wrote his paper, he was still 3 decades away from this dilemma, but he wrapped up with a prescient debate:
“The Spinozan hypothesis suggests that we are not by nature, but we can be by artifice, skeptical consumers of information. If we allow this conceptualization of belief to replace our Cartesian folk psychology, then how shall we use it to structure our own society? Shall we pander to our initial gullibility and accept the social costs of prior restraint, realizing that some good ideas will inevitably be suppressed by the arbiters of right thinking? Or shall we deregulate the marketplace of thought and accept the costs that may accrue when people are allowed to encounter bad ideas? The answer is not an easy one, but history suggests that unless we make this decision ourselves, someone will gladly make it for us. “
What Gilbert couldn’t know at the time was that “someone” might actually be a “something.”