Sanderson’s Third Law of Magic: Expand what you already have before you add something new
Belle’s Third Law for AI as Magic: Build AI on what you already have (AI by itself is rarely the first, only, or best solution)
REMINDER: Sanderson’s Laws are for authors, Belle’s Laws are for characters (us).
Sanderson’s Third Law of Magic: Expand what you already have before you add something new
As with the first two laws, Sanderson has an insightful and interesting essay that fully expounds on this concept and offers illustrative examples. He captures the heart of this law when he says, “Often, the best storytelling happens when a thoughtful writer changes one or two things about what we know, then extrapolates purposefully through all of the ramifications of that change.”
Superman, although never mentioned directly in Sanderson’s Third Law blog post, is actually a great example of how an author can use this law to great effect. Powers across the DC Universe stem from a variety of sources but in the core Superman stories, the key difference between his world and ours is simply the concept of Krypton. The history, ‘DNA’, etc, that stem from Krypton facilitate truly epic storytelling and were built out over time by expanding what was known. There is no need for new powers or new sources of power within the Superman story because diving deeper into Krypton is, all in itself, a deep well of storytelling potential.
Fantasy worlds explored across multiple books or stories, such as the DC Universe or Discworld (created by Terry Pratchett with stories from Sanderson, among many other authors), are the best examples of how this law plays out. More pages mean more opportunities to truly explore and expand on existing magical systems before needing to add something new.
A personal favorite of mine is the way Drew Hayes evolved his characters’ understandings of their capabilities in his Super Powereds series. The final book in that series would have been far less satisfying if the main characters all suddenly manifested new powers to rise to the escalating challenges and yet they were all able to act in ways that would have seemed impossible in the first book.
In many ways, this law is the most author-oriented of the set. There is truly nothing stopping an author from ‘discovering’ a new facet of the world they have created (or been invited into) beyond the desire to engage their audience. This Law is great advice for how to make wise decisions about when and how to use that gift of freedom while still writing compelling stories.
Belle’s Second Law for AI as Magic: Build AI on what you already have (AI by itself is rarely the first, only, or best solution)
AI’s rapid evolution often feels as if the author of reality is intentionally breaking Sanderson’s Third Law. The gap between what technology is capable of now compared to 5 years ago is equivalent to everyone suddenly gaining the ability to manifest a new superpower or magical capability in response to each new challenge that pops up.
That could still be a compelling foundation for opening a story, particularly a saga. But it is an interesting idea and an interesting premise only if the goal is to create a truly chaotic ball of disaster from which the characters will eventually need to emerge. It is basically a great way to torture the characters for the enjoyment of the reader.
As a character stuck in this world, I would prefer to limit the disasters up front no matter how curious I am to see how the growing balls of chaos would be untangled.
To that end, we may have nearly 5 years of rapid AI evolution but we have thousands of years of human innovation and introspection that we can build upon. Before we determine that technology, even AI, is the answer to a given challenge we should start by questioning if a technical solution makes sense at all. Only then should we consider if AI is the first, only, or best answer we need.
Let’s start once more with our email reviewer example. I happen to be someone who thrives with a blank page. Writing is how I shape my thoughts and I often do not know what I am trying to say until I have written it out (and a version history review of this exact article would perfectly illustrate this point). For me, jumping straight to AI for email drafting would undermine my ability to think clearly about what I am trying to say. The result is that AI is rarely, if ever, the ‘first’ piece of the puzzle when I am looking to make improvements to the way I engage with my inbox or any writing.
However, many people struggle with a blank page and are actively seeking a jump start to help them get from nothing to something that they can react to. This is a need where AI is very well suited for offering up that initial draft. Yet if the human removes themselves from the loop completely, if the human makes AI the ‘only’ solution to email drafting, emails can quickly drift far from what and how the human would prefer to communicate.
Plus, LLMs are highly nondeterministic. This is by design and a core strength of the technology but also means that LLMs are a poor fit for situations where the answer should always be the same. Where safety regulations mandate a particular performance standard, test results performing below the standard cannot be used to sign off on a new design. This is a deterministic standard that should be hard coded in with any variations permitted only by the human who will have to defend the decision in court. In any equally rigid context, a less ‘advanced’ technology that prevents surprises is likely to be the ‘best’ fit.
As we expand the challenges beyond something as focused as email reviews, the likelihood that AI will be the first, only, or best solution shrinks proportionally. Poor sales team performance almost certainly requires a survey review across tasks (scripts), process (sales funnel), and product (particularly the product market fit). Replacing an underperforming sales team with a herd of AI agents will have no impact on a poor product market fit and only changes the funnel issues if the agents are granted freedom to vary from the funnel beyond what the humans were allowed (a policy rather than technology fix). The agents may perform better on the individual tasks, such as scripts, or they may perform worse. Belle’s First Law would really come into play there.
AI alone is neither the first, only, or best solution to poor sales performance.
Of course, AI could be a powerful tool for understanding and assessing product market fit. It may be a crucial piece of the puzzle when redesigning the funnel. AI may even be the key to identifying poor script delivery as the actual issue. You may even end up with a fully agentic sales team when all is said and done.
Increasingly often, AI will be part, maybe even a big part, of the solution. But humans still have a role to play and that means that AI is always going to be one piece of a larger puzzle.
Intellectually, I understand the progression from deterministic algorithms through neural network based machine learning into LLMs. Practically, however, it certainly feels like there has been a hard break between what technology was capable of in 2020 versus 2026.
When there is a constant refrain of “AI can do that”, it can be easy to assume that that is the entire story. If AI can do it, AI should do it. Belle’s First Law tells us that we need to understand AI within the context of the challenge we are looking to solve. Belle’s Third Law tells us that we also need to understand the full picture even when AI is part of the solution.
Years before LLMs, Bill Gates was quoted as saying, “The first rule of any technology used in a business is that automation applied to an efficient operation will magnify the efficiency. The second is that automation applied to an inefficient operation will magnify the inefficiency.” Sadly, many people today are more likely to treat AI like a key FOB than an amplifier. A key FOB simply needs to be waved towards the sensor and the lock is opened. We use it and then move on without needing to think much more about the context around the FOB. An amplifier requires input and creates output. It makes sense only as part of a larger system.
Given that Sanderson’s Third Law is the most “author-oriented” of the set, it took me some time to figure out what the AI as Magic equivalent should be. In fact, I half expect the framing of this law to evolve over time so stay tuned (and share your thoughts if you have them). One version even simply said, “AI is not a set of key FOBs (First Only Best solutions)”. I also explored several variations of ‘clean your closets before you build a new shed’ and even a classic “Garbage in -> Garbage out even when the -> is AI”.
The First Laws play on unintended consequences and the Third Laws effectively tell us that there is no such thing as a free lunch (or TANSTAAFL, as I will always associate with Robert Heinlein). In Sanderson’s case, the point is that free lunches are boring. If everyone is always perfectly well fed at no cost to themselves, the story is already over. In Belle’s case, the point is that readers don’t care about that story because it is implausible and therefore definitely not the story we are living today. Only by building from what we know will we arrive at answers that are best suited to serve our needs.
AI is a given at this point. The questions are who, when, and how. Belle’s Third Law is the third plank in a foundation for answering those questions.
Check out the rest of the series:
Previous