We can probably do better than this
right?
AI tools are pretty amazing. You provide some information in the form of a prompt and context, and they return some content. Drop a penny in the slot and you get a prize. Unfortunately, the default is slop because the machine does not require meaningful input, nor does it evaluate whether the output is true, valuable, or worth creating in the first place.
We are drowning in a sea of slop because we were given the power to create content easily without necessarily acquiring the wisdom to determine if it deserved to be created in the first place. When it was difficult to create, the friction stopped a lot of us from doing it. Now that it’s easy, most of us are throwing off slop the way Goyle set fire to the Room of Requirement in the Harry Potter movie. We quickly conjured something that we did not know how to control, and it is now consuming the space where we kept important things. For examples, check the internet, streaming platforms, or social media.
I should clarify that I do not think AI creates slop by necessity, only by default. It is possible to do quality work with AI tools in your workflow, but it requires attention to the inputs and an understanding of what the prediction machine contributes. (And if we have proven one thing as a society, it’s that we all love attention to detail and a deep understanding of how things work.)
LLMs are not meaning-creation machines. They are language-expansion machines. They use patterns learned from past meaningful human expression to produce new expressions that look as though they emerged from understanding, intention, and judgment. Humans can use those expressions to do meaningful work, but the meaning enters through human observation, purpose, interpretation, and selection.
When those things are absent, the model still produces the form. That gap between the appearance of meaning and the substance behind it is where slop comes from.
Meaningful content is the more difficult path. It requires an understanding of what the machine contributes, of what it cannot contribute, and of what must still come from the human. In order to regain control of this beast, we need to understand the difference between meaningless content (slop) and meaningful communication. Claude Shannon’s original distinction between information and meaning gives us a useful place to begin.
When Shannon established information theory in 1948, he purposefully excluded meaning. It was irrelevant to the engineering problem he was trying to solve. Information theory was concerned with how messages could be represented and transmitted, not with what those messages meant.
Shannon described the bit as the basic unit of information. A bit measures information without describing its significance, intention, or value. A communication system can transmit a love letter or an appliance manual without understanding either one. Meaning exists outside the transmission mechanism.
This relationship made intuitive sense because we did not expect our telephones to contribute to the conversation. We expected them to accurately carry a message from one person to another.
Seventy-five years later, AI tools, one of which shares Shannon’s first name (Claude), attempt to challenge that division. They use large-scale prediction algorithms to generate sequences of tokens, computational fragments of language, that form coherent expressions. Because these expressions resemble the products of human thought, we infer that understanding, intention, and judgment must exist behind them.
The way these systems are marketed amounts to an implied theory of meaning: that sufficiently sophisticated prediction can bridge the gap that Shannon deliberately left outside information theory. That our devices can actually contribute to the conversation.
Shannon showed that predictable structure could be removed from a given message so that it could be transmitted more efficiently. To say that another way, Shannon-style compression reduces the volume required to transmit an existing message by exploiting its predictable structure. The predictable bits weren’t necessary to get the message across. LLMs metaphorically move in the opposite direction. They use learned predictability to add linguistic structure to a limited input. LLMs love adding back in the predictable bits, it’s their thing.
The AI machine treats the prompt as though it were compressed meaning, as though a few supplied words contained a supernova of meaning waiting to be released. But the model is not actually recovering hidden meaning from the prompt. It’s just supplying plausible associated language from patterns learned elsewhere.
LLMs have learned the forms of language that usually accompany meaningful human thought. They can reproduce those forms without reproducing the observation, experience, intention, or judgment that originally produced them. The output may actually still be useful. It may help a person explain, organize, translate, summarize, or explore an idea. In many situations, the distinction between actual meaning and perceived meaning really doesn’t matter very much. We are predictable communicators for the most part.
There are many uses for a tool that can predict a meaningful-sounding statement when what needs to be said is already highly patterned. A work email about setting up a meeting, a summary of a commonplace idea, a translation, or a familiar explanation may not require original observation or consequential judgment.
The boundary appears when we ask prediction engines to supply meaning, something that prediction cannot provide. When the work depends on a new observation, an unstated purpose, a heartfelt sentiment, a consequential choice, or responsibility for deciding what matters, the machine has nothing substantive to expand unless a human supplies it.
By meaning, I do not mean merely that a sentence can be understood. I mean that it is grounded in some combination of observation, intention, context, purpose, and judgment. Coherence alone is not evidence of meaning. Coherence without substance can present as meaning even when there is very little underneath the form.
This is where slop is born.
When Charles Babbage proposed the Difference Engine (his proposed mechanical calculation machine) to the British Parliament, he reported being asked whether the machine could produce the correct answer if someone entered the wrong figures. Babbage regarded the question as nonsense. A calculating machine could produce reliable answers to well-constructed problems, but it could not repair a defective question.
But the question also revealed something about what people had come to expect from intelligent human counterparts. A person did more than process the inputs. A person could question the premise, interpret the intention, identify missing information, and help determine whether the problem made sense.
A machine that automatically answered well-constructed questions was difficult to understand because many people still needed help constructing the question.
The Difference Engine may have been ahead of its time, but the idea eventually took hold. As we learned the mechanics of calculating machines, we built tools that used their power under the direction of human programmers. We developed a clearer division of labor. The machine performed the calculation. The human defined the problem, selected the method, interpreted the result, and accepted responsibility for its use.
I believe the same process will occur with AI. As we understand its strengths, we will learn to use them. As we understand its weaknesses, we will learn what must remain under human control.
The lesson is not simply that we need better prompts or refined context. We do need those things, but more importantly, we need to understand the division of labor. The machine can predict, expand, reorganize, translate, and imitate. The human must supply the intention, the purpose, the standards; ultimately the responsibility for meaning is still ours.