Mechanical Turks

2016 was the year that AI started to get really exciting in the wider world. Deep machine learning, self-driving cars, voice-driven user interfaces - we've seen a step change in both expectation and achievement. And in games-y news, Elon Musk's AI company has just announced an initiative, 'Universe', that trains AI agents in general intelligence skills by having them play thousands of games - everything from Flash-based browser games to GTA V.

So AI is hot right now. And I've had variations on the same conversation many times in the past few months. 'So,' people say, 'your thing is game narrative. Procedurally generated narrative keeps getting better. AI is the new hotness. Are you about to become obsolete? Aren't you worried?'

I'm not terribly worried. Procedural content generation can do a lot of things that writers and designers would otherwise have to do manually. But spreadsheets can do a lot of things that accountants and project managers would otherwise have to do manually. Spreadsheets haven't yet made accountants and project managers obsolete. I mean in the end we'll likely all of us become obsolete - machines will end up doing accountancy and project management and writing and neurosurgery - but we're a long way from that.

"The key insight is that story is not the same as words, any more than buildings are the same as bricks."

So why do people keep asking me that question? Why does it seem intuitively reasonable that it would be practical-ish, soon-ish, to automate the creation of stories - to have the Story Machine, the Infinite Content Generator, the holodeck where AI agents generate their own dramas, or have a rule system churn out story that slots into the gameplay?

First of all, honestly, it's because story is the red-headed stepchild of game development. There have always been great game narratives, and the bar continues to rise. But it's impossible to imagine videogames without art or tech, and there are plenty of videogames with minimal story, or no story at all. So we get a lot of schlock. And when someone says 'a machine could do that', often what they mean is 'even a machine could do that'.

But there's more to it than that. I think that people see machines constantly putting the right words in the right order - and after all, that's ultimately what writing is. Of course it's much easier to put words in some kind of right order than other kinds of right order, but the difference isn't always obvious - again, especially when game content embarrassingly often looks like filler.

The key insight is that story is not the same as words, any more than buildings are the same as bricks. It's very easy to automate the creation of bricks, and a building may be composed mostly of bricks. So we can automate the creation of buildings, right? Intuitively, we know that's wrong. Partly this is because buildings also have foundations, cabling, roof tiles - but of course story also has plot, worldbuilding, theme. But partly this is because we understand immediately that it's fundamentally a different kind of problem to arrange a hundred thousand bricks in the shape of a wall than to arrange a hundred thousand bricks in the shape of a library. In the same way, it's a fundamentally different kind of problem to arrange a hundred thousand words into a good novel than it is to arrange a hundred thousand words into a series of procgen madlib gags - as the Magic Realism Bot wonderfully demonstrates.

" You could say that gameplay is the stuff that keeps happening over and over again, and story is the stuff that keeps happening that's different"

Consider a traditional RPG - something like Diablo - which uses light procedural generation to create levels, place monsters, generate loot. A lot of the quests in Diablo are pretty templatey. I frequently see people suggest that in this kind of situation, you could break down quests into standard components, and reassemble those components jigsaw-wise into new quests. You could, and this is what a lot of procedurally generated quest systems look like. But that's the point - they look procedurally generated.

You could say that gameplay is the stuff that keeps happening over and over again, and story is the stuff that keeps happening that's different. You can play Tetris or DOTA or Civilisation for hundreds of hours and keep seeing the same thing happen in a different orders and contexts and it'll feel like satisfying gameplay, but if you see the same story happen in different orders and contexts, you'll stop caring about it quite quickly. You'll see the same patterns. So when a player wants more story in a game, what they actually often want is more interesting and novel things in the game. That's not just a resource problem. It's also a design problem.

"Hang on!" - someone will quite reasonably say - "some of the best game narrative experiences are emergent and player-driven - they arise out of interactions in gameplay." That's true! But when someone suggests we use AI to generate content, they're not talking about emergent stories. They're talking about script, content, words in the right order. Machines have no monopoly on uninspired, remix-y content. Humans create plenty of that too. So creating original content that feels different and maintains the player's interest isn't even a solved problem for us. Creating samey, uninspired content is very much a solved problem, but I don't know why you'd bother to write software to do that. Mediocre writers are both cheap and plentiful.

"Machines have no monopoly on uninspired, remix-y content. Humans create plenty of that too. So creating original content that feels different and maintains the player's interest isn't even a solved problem for us"

So although you could probably eventually build an AI to write good story content - or a decent poem, or a political manifesto which would actually win an election - it's not something we're likely to be able to do tomorrow. Much more than that, though, why would we particularly want to?

The world is not short of stories. The world is bursting with stories. I will never get to the end of my Netflix watchlist or of my narrative-tagged Steam library. We have a healthy ecology of competitive creative endeavour, and the quality of mass-market narrative entertainment, broadly speaking, improves year on year.

So as a consumer I have an embarrassment of riches available to me. When I feel short of content, it's because I want something specific - the next Half-Life (DREAM ON, KENNEDY) or the next episode of Fargo. I have an interest in a particular, distinctive story or creator, and even if you could perfectly synthesise more of that material (the way Next Rembrandt [https://www.nextrembrandt.com/] are trying to do with the painter), it wouldn't allow for infinite content. There are only so many stories you can tell in a particular world or context before they start, again, to look samey. The desire for infinite quests in an RPG is the desire to have Christmas every day. It's special because it's exceptional. Pretty soon you'd have present-calluses, and you'd be ill at the sight of a sprout.

And as a studio - well, I've managed a writing team who were really good but also human, and I've employed freelancers who've varied from excellent to not-excellent - I can understand the temptation to replace intermittent humans with 24/7 machines. But I think any manager who feels that temptation should also recall the problems they've had with fallible technical teams, and imagine the conversations they'd have trying to winkle bugs out of a dialogue generator. Software, except in speeches given by pop-eyed enthusiasts in 80s tech thrillers, is not infallible. It just causes different problems under different circumstances.

"Some technologists and creatives simply find the idea of a story machine an interesting problem to solve, and make it their quest. That's no crazier than any other creative endeavour"

And at the macro level, if you sell content, why would you want to create a technology that reduced the cost of that content from 'relatively cheap' to 'basically zero'? The De Beers diamond mining company has a unit, Element Six, which synthesizes diamonds which are almost indistinguishable from the ones that come out of their mines. But De Beers don't sell their synthetic diamonds. They use them to develop machines that keep the 'almost' in 'almost indistinguishable'. Okay, I'm fudging the issue here. Most studios do a lot of things besides selling content. But the bigger the operation, the more the challenges around story content are not around churning out words - they're around making the words fit meaningfully with everything else in the project. It's not the bricks - it's the building.

(Have you ever rolled a snowball across snow to make a super-big snowball? To begin with, you just push it, and it grows satisfyingly. But at a certain size, it becomes more of a lumpy cylinder than a ball; it picks up rocks and bits of branch; cracks form, snow falls off and needs to be patted back on; the effort of pushing the snowball or providing the snow is the least part of it. Maintaining a big story always reminds me of that.)

Some technologists and creatives simply find the idea of a story machine an interesting problem to solve, and make it their quest. That's no crazier than any other creative endeavour. Some focus on a more specific aspect of creating story-like effects in a reproducible way, through procgen or game design - that's really what I'm enthused about on my side projects. And some focus on technologies that are to story what spreadsheets are to accounting - better seeds for creative work, better ways to manage complex interactive narrative. That is really cool, and I get very excited by these technologies.

But fundamentally, the answer to the question 'when can we replace novelists with software' is not so much 'not yet' as 'why would you bother?' Attention is a zero-sum game. I do wonder if, a hundred years from now, we'll be using AIs not to write our stories, but to read them.