No, GPT-3 Is Not Superintelligent. It’s Not Tricking Humans, and It‘s Not Pretending to Be Stupid.

There is an unhealthy level of GPT-3-powered snake oil being sold to non-technical readers.

There is a lot of hype surrounding OpenAI’s incredibly powerful machine learning algorithm GPT-3. The algorithm can generate convincing pieces of text with very little input required. Give it a title such as “Feeling unproductive? Maybe you should stop overthinking,” and it can generate an entire full-length cohesive article that could go viral (and it did). It can be used to perform a wide range of tasks. For instance, GPT-3 can summarize articles, generate fan fiction, and even write programming code.

There is no doubt that this is an outstanding machine learning algorithm. Some, myself included, might rightfully call it a game-changer. However, as is often the case with riveting technologies, some people jump to the most absurd conclusions.

Let me give you some examples.

One author, Bernard Mueller, wrote a popular article titled “I asked GPT-3 for the question to ‘42’. I didn’t like its answer and neither will you.” In its conclusion, the author hypothesized: “Perhaps it [GPT-3] knows the question perfectly well, but considers humans as too immature and spoiled to tell: In its [GPT-3’s] opinion, we shouldn’t even bother to find questions to answers we can’t possible understand” (Bernard Mueller).

On Twitter, Eliezer Yudkowsky, who at the time of writing has quite a reach with close to 40,000 followers, wrote a series of tweets on GPT-3 saying: “So I don’t want to sound alarms prematurely, here, but we could possibly be looking at the first case of an AI pretending to be stupider than it is” (Eliezer Yudkowsky).

Now, let me be clear.

This is complete and utter nonsense. GPT-3 does not consider humans immature. It does not consider humans to be spoiled. It does not consider us stupid. It does not pretend. It isn’t trying to trick us. And while these two particular authors may not claim so, I do want to make it clear that GPT-3 is neither conscious nor superintelligent.

GPT-3 can do one thing and one thing only: predict what word comes next in any given sequence. And it’s really, really, really good at it.

As a matter of fact, GPT-3 doesn’t even know what humans are. If you ask GPT-3 a question such as “What do you think about humans?” or “What are humans?” it will surely come up with a great answer, but it doesn’t actually comprehend the answer itself. The algorithm is simply trained on various human-written text pieces and tries to imitate them (usually successfully).

Another important point to mention is that both of the aforementioned authors drew their conclusions based on conversations had with GPT-3 through an application called AI Dungeons. See, OpenAI has been very restrictive with handing our licenses to use GPT-3. Most of us, therefore, have to interact with GPT-3 through applications made by other developers. AI Dungeons is one such application. It’s a free-to-use text-based role-playing game where the algorithm has been trained to create stories. Obviously, if you ask GPT-3 a question through AI Dungeons, the algorithm will have a bias towards responding in a way that creates a good story. Not so surprising, then, that it gives thrilling answers to philosophical questions.

Again: I don’t want to downplay GPT-3. It’s a massive leap in word-generating AI. But I also think it’s unhealthy to perceive it as a Hollywood-esque AI that attempts to trick humans. I understand that selling fear to non-technical readers is great for clicks, but I strongly disagree with it. As if the public perception of AI isn’t already distant enough from reality, having GPT-3 snake oil going viral only further instills fear in artificial intelligence.

There are certainly reasons to be afraid of our present-day machine learning algorithms. However, AI isn’t deliberately trying to trick us, and it isn’t trying to act like it’s stupider than it really is.

Well, not yet, anyway.