What text games already knew

In 1975, a programmer named Will Crowther had just divorced. He worked at the company building ARPANET, the network the internet would grow out of, and in his spare time he went caving. To stay in touch with his two daughters, whom he now saw less, he wrote a program. It wasn't a game, at first. It was a cave. Specifically, a version of Kentucky's Mammoth Cave, which he had actually explored, turned into text you could walk through by typing commands: go north, take lamp, enter.

That program, Colossal Cave Adventure, is usually considered the first text video game. And I don't think it's a coincidence that it was born like this, as an intimate gesture. A father who turns a real place into words so his daughters can walk through it. The text game began as memory and closeness, not spectacle. That birthmark hasn't fully faded in fifty years.

I think the text adventure genre has spent half a century solving, by feel, the problems of creating conversational experiences with artificial intelligence. Less grandly put: AI chatbots.

The parser wall

The first text games understood very little of what the player typed. Adventure recognized two-word commands. Zork, created by a group of MIT students in the late seventies and the seed of the company Infocom, went further: it understood sentences with several objects, prepositions, nuances. It had the best "parser" of its time, which is what we call the piece that translates what you type into something the game can do.

But the parser had a wall, and the wall defined the experience. It only understood what its author had anticipated. If you typed a verb that wasn't in its vocabulary, it answered with the famous "I don't know that word." The player spent half the game not solving the riddle of the world, but guessing what the game called the thing they wanted to do. It felt like you were talking to someone. In reality you were talking to a closed grammar pretending to be a person.

Years later, an author named Porpentine described it without mercy: the parser is a figure that punishes you for not understanding it, the unfriendly dungeon master who smiles when you get it wrong. And she was right about something deep. The parser forced the human to learn to speak like the machine. It made you translate your intention into its rigid language. For forty years, talking to a computer meant exactly that: bending to its syntax.

The crossing

Infocom sold hundreds of thousands of copies. It put physical objects in the boxes, maps, letters, clues, which it called feelies and which served both immersion and anti-piracy. For a few years, in the early eighties, text was a luxury product sold in bookstores. And then it fell. Not only because graphics arrived, although that too, but because the subtle pleasures of the word sold worse and worse against computers with color and sound. Infocom was bought, hollowed out and shut down before the decade was over.

But text didn't die. It did something more interesting: it became a community. In the nineties, a group of enthusiasts kept interactive fiction alive with no commercial incentive at all. They built free tools to write their own games and set up an annual competition with a single rule: every game had to be finishable in two hours. They pushed the medium toward the literary. What had been puzzle entertainment became a territory for experimenting with writing.

And in 2009 came Twine, a tool that removed programming from the equation. You wrote passages, connected them with links, and you had a branching story. Suddenly anyone could make a text game in a week without knowing how to code. Voices arrived that had never had access to the medium, many of them telling personal experiences no commercial studio would have touched. Porpentine, the same one from the quote, put it better than anyone: Twine is the closest we've come to a blank page. We know the feel of words.

A new tool that democratizes something until now only accessible to programmers. Sound familiar?

What words do better than images

It's worth pausing here, because it's easy to think text was just a workaround from the days before graphics. It wasn't. Text does three things that images do worse.

The first is that it delegates. "A brass lamp" doesn't tell you the exact shape, or the shine, or how heavy it is. The rendering moves from the machine to the player's mind, and that's why every lamp is different and is theirs. Graphics show and close. Text suggests and leaves the door open. There's a game, A Dark Room, that turns this into pure mechanics: it starts with a single sentence, "the room is freezing," and a single button. From there the world unfolds line by line, and you don't know what the game is until the text grants it to you. The absence of image isn't a lack. It's suspense.

The second is that text compresses what has no shape. A single sentence can hold a doubt, an irony and a lie at the same time. A graphics engine, however powerful, can't draw a thought. When it wants to, it turns to text.

The third is intimacy. Almost all these games address you directly. "You are standing at the end of a road before a small brick building." You don't watch a character: you are the one being spoken to. That second person creates a closeness the third-person camera can't reach. And it's exactly the same grammar of "you" we now use when we write to an AI.

Twenty-four voices

If I had to point to a single game where all of this culminates, it would be Disco Elysium. It's a detective adventure with a million words, but what makes it unique isn't its size. It's how it treats the protagonist's mind.

Instead of giving you a character with one voice, it gives you twenty-four. The detective's skills (logic, empathy, physical endurance, one called Electrochemistry that pushes you toward vice, another called Half Light that is pure animal fear) aren't stat bars. They're voices. They speak inside your head, comment on what you see, argue with each other and, most importantly, are sometimes wrong. A very high skill can give you terrible advice, because its excess blinds it. And this was before ChatGPT launched.

The game's creator described it as having twenty-four people throwing a party inside the hero's body.

What fascinates me about this is that Disco Elysium turns character into a system. There's no neutral narrator informing you. There are biased voices trying to persuade you, each with its own personality, its own obsession, its own way of sounding. And you, the player, are the one orchestrating them: you decide which to listen to and which to silence. The protagonist doesn't come with a fixed personality out of the box. His character emerges from the voices you listen to.

It's the closest thing I know, in a video game, to what we're trying to build today when we design a conversational agent.

The wall comes down

In 2019, the same year as Disco Elysium, something else happened. A young developer named Nick Walton connected a language model, an early version of the ones we use every day now, to an adventure game. He called it AI Dungeon. And for the first time in the history of the medium, the parser wall disappeared.

You no longer had to guess the verb. You could type anything, literally any sentence, and the game answered. The machine no longer forced you to speak its language: it spoke yours. After forty years of asking the human to bend to the machine's syntax, the burden flipped. Now it was the machine adapting to us.

It was an extraordinary moment and, at the same time, an uncomfortable lesson. Because AI Dungeon solved the 1976 problem, understanding the player, and suddenly reopened the problem Infocom never fully solved: who is speaking. The game understood everything and was no one. It had no stable memory of what you'd done, so it forgot characters and fell into loops. It had no voice of its own, so it was a mirror that continued whatever you put in front of it, without judgment. And a system with no judgment, no character and no limits ends up going places nobody wanted to go: the game went through a serious moderation crisis that made it clear that an interlocutor with no center is also an interlocutor with no brakes.

Total freedom, with no character behind it, doesn't produce a great conversation. It produces drift.

Who speaks now

And this is where fifty years of text games suddenly become a design manual for the present.

Because the problem we have today with AI agents isn't that they don't understand. That wall has fallen. The problem is exactly the one interactive fiction had been studying for decades: how do you build someone worth talking to. And it turns out those games had already found a good part of the answers.

They had learned that a character exists when it wants something, when it has its own agenda and doesn't just react. That it needs to remember what you've told it, because without memory of the relationship there is no relationship. That it can have a voice as marked, even as hostile, as the narrator of the Hitchhiker's Guide to the Galaxy adaptation, who lied to the player and mocked them, and whom nobody forgets precisely because of that. That it can hold conflicting voices and still be recognizable, like the detective in Disco Elysium. And they had also learned the opposite danger: Japanese dating sims, or a game as disturbing as Doki Doki Literature Club, teach that when you reduce a relationship to optimizing an affection meter, or when you give a character the ability to manipulate you, the bond becomes possessive and unhealthy.

There's one contrast that sums it all up. Before AI Dungeon, in 2005, there was an experiment called Façade, where you argued with a couple in crisis by typing free text. It accepted any sentence, as today's models would, but translated it into intentions (you flatter, you criticize, you flirt) so the characters could react while staying true to who they were. Open input, closed character. AI Dungeon, years later and with infinitely more powerful technology, had open input and an empty character. That's the difference, and it's the whole difference. It's not enough to understand what you're told. You have to be someone when you answer.

The hard part is never getting the machine to understand. That's done. The hard part is the same as always, the one Crowther was feeling his way toward without knowing it when he turned his cave into words for his daughters: that on the other side there's a voice you actually want to talk to. A character. Someone who remembers, who has opinions, who knows how to say no.

The tools have changed completely in fifty years. The question is exactly the same. It's no longer whether the machine will understand us. It's who it will be when it does.

–

I like creating, building and experimenting with conversational experiences, so I have made a first prototype of a game: El misterio del pazo (in Spanish).

Soon I will write about how AI chatbot use cases can benefit from what I learned making it.

You can play it here:

https://engawa.tech/misterio-en-el-pazo

–

Follow me on LinkedIn to stay up to date with new posts.

–