or How to work with AI that “suffers from hallucinations”
An author from the Pitch Avatar team explains how science fiction writers, as early as the mid-20th century, found solutions to the problem of “machine delirium.”
Since the 19th century, a clear pattern has existed: long before civilization encounters problems related to technological progress in practice, science fiction writers manage to “invent” them first. And not only invent them, but also propose solutions.
AI hallucinations — or, as they are sometimes called, machine delirium — are no exception. Let’s clarify right away: we are not talking about malfunctions or failures caused by physical damage to hardware or, for example, computer viruses. Nor are we talking about a hypothetical “machine uprising,” where AI “consciously” refuses to obey humans and decides for itself what to do.
What interests us are situations in which artificial intelligence generates false, incorrect, or outright absurd information. For example, non-existent quotes by non-existent people from non-existent books. Or links to web pages that do not exist. Or a kind of “visual salad” in images. We assume that every reader has encountered such AI-generated outputs at least once. Sometimes, the AI’s response to a prompt is so far removed from what it was asked to do that an entirely illogical urge arises to declare it “crazy.” This is especially frustrating when AI starts to “glitch” while solving a truly important task that requires a high degree of accuracy. As a result, some people become so disappointed that they stop using AI-based tools altogether, preferring to solve their problems “the old-fashioned way.”
So, let us reassure you. Masters of science fiction have been thinking about this topic since the middle of the last century. Why not look to them for a solution to the problem of “machine delirium”?
The Terrifying Terminus
Let’s take, for example, Stanisław Lem’s short story “Terminus.” The plot, in brief, is as follows: the captain of a spacecraft inherits an old maintenance robot named Terminus, which, while performing its duties, constantly taps something out in Morse code. It turns out that the robot witnessed a spacecraft disaster and is reproducing the final communications between the crew members who died. What is most interesting is that when someone addresses one of the people, on whose behalf the robot is “conducting the conversation,” using Morse code as well, Terminus begins to reply on their behalf, as if they were still alive and aboard the doomed ship. Otherwise, Terminus performs its duties perfectly well, and when spoken to in normal human language, it does not recall either the disaster or the deceased crew members. The ship’s captain, being a pragmatic person, uses the robot until the end of the mission and then scraps it.
Despite the fact that this story does an excellent job of “sending chills down the spine,” what we see here is a clear example of a machine hallucination. For some reason, during the catastrophe, the AI robot self-trained to imitate the communication between crew members desperately trying to save themselves. Why this happened is, by and large, not important. One of the story’s key ideas is precisely that when creating intelligent machines, it is not always possible to predict in which direction their self-learning process might deviate. What is important is something else. When the hallucinating robot was given a direct, clearly and unambiguously formulated command, it executed it quite effectively. Personally, I have always felt that the captain in this story acted too radically. It would have been far more logical to hand Terminus over to specialists for study.
Apparently, Lem wanted to show that when difficulties arise in working with AI in the future, the best solutions will often be simple ones. Namely, simply replacing a “glitchy” tool with another.
Simple Logic
Personally, I much prefer the idea underlying the short story “Reason” by another classic, Isaac Asimov. In this story, an AI robot appears, at first glance, to stage a typical “rebellion.” The action takes place at a remote space station with a crew of only two people. The robot in question is a complex system designed, essentially, to manage other robots. It considers itself the “crown of creation,” deifies one of the station’s main machines — the energy converter — and refers to humans as beings of a lower order. Moreover, it begins spreading its “teachings” among the other robots.
However, it soon becomes obvious to the humans that the “rebellious” AI still obeys their commands and performs its duties flawlessly. Yes, it constantly explains its actions according to its own worldview, but its hallucinations have no effect whatsoever on the quality of its work. As a result, the humans decide to leave the robot’s “harmless madness” without attention.
The Key Is Formulation
Take note: in both Lem’s story and Asimov’s story, hallucinating AIs perform their intended functions perfectly when they are given clearly and precisely formulated commands. Thus, already in the “golden age” of science fiction, the main solution to the problem of AI hallucinations was identified. In most cases, it is enough to formulate the task for AI correctly. If you want to encounter “machine delirium” less often, formulate your prompts as clearly, precisely, unambiguously, and concisely as possible.
