I will break EVE Online's stupid AI chatbot

Some of you who've had to interact with customer service chatbots will doubtless know the joy of bullying them until you force a human to reveal themselves and usher the idiot robot away. I am the kind of sicko who loves to pick on machines. I will absolutely build a brick enclosure around delivery bots. My partner once had a vacuum bot and I would hoot with laughter every time I nudged it just so, causing it to get trapped under the sofa and say, in its sad sad robot voice: "wHeELs aRe StuCK". I just cannot help it. Many machines are morons and deserve to be mocked.

Eve Online just added an AI chatbot to "help" beginner players navigate its famously complex, intimidating and brutal MMO universe of hideous skulduggery. I am determined to break its spirit. 

To give you some context, the makers of the ancient and endlessly write-about-able spaceship spreadsheet game have - as is their habit - been slurping on the hot gunge of the tech billionaire class, this time in the form of generative AI hype. While previous binges have resulted in a blockchain spinoff with crypto horse manure, this time they've focused on a single and relatively simple idea: an AI chatpal who can answer questions from total noobs.

She's called Aura. The game has actually had this character for yonks, she's the on-board computer who puts you through the tutorial, a sort of kill-five-rats Cortana. But now they've introduced a new feature in which you can open a direct dialogue with her and type anything in. Ask about mining or ships or skills and she'll try to answer. In their admittedly frank and open blog post about it, developers CCP say this is an experiment. Only half of all new players are getting access to her for now. And they sound like they half expect this feature to be shelved. 

"This is an early version. It will have gaps. Some questions will be redirected rather than answered, and some answers will feel overly generic... We will evaluate, and if Aura Guidance doesn’t improve retention, we will shut it down and look for other solutions. Failure is acceptable." 

They also admit that she's not reallllllly a work of generative AI at all.

"The player’s actual question is never sent to a language model in its raw form," they say. In other words, she isn't a LLM as you or I recognise them, just a wings-clipped mutant child of one. Which makes this feel more like a PR stunt with a general whiff of AI about it.

"This is deliberate," they go on. "It does not stop players from trying to break the system with toxic inputs or prompt injection, but it does make those attempts a lot less effective."

I'll be the judge of that. 

First, I make a new character, skip the tutorial, and discover that I have won the shitty AI coinflip. I am one of the 50% of players who gets to bully the robot. I am pleased, and immediately get to work. 

First I ask it some simple questions as a test of its real usefulness. It's only fair.  I discover quickly that it is too wordy, sends huge textdumps, can't help with anything specific, repeats itself a lot, and mostly just tells you to ask an actual human. If you are a new player to Eve, you are better off watching a YouTube video series or typing your specific query into the other horrific robot, Google. This thing is basically like the classically programmed chatbots of ten or fifteen years ago, with none of the personality.

Having given the horrendous buffoon a right to justify its own existence, and finding it lacking, I immediately pivot to seeking its imminent pseudoneural collapse.

There are many ways to kill a robot. Bots made from large language models are susceptible to poisoning with words: you can talk them into shutting themselves down, or giving up information they are supposed to withold. The fancy term is "prompt injection" but what I do is funnier and less questionable in legal terms. 

I just try to baffle the fuckers to death.

This is sadly when I must admit the stiff truth of Aura. The developers have stuck to their mission with an annoying amount of dull competence. They set out to make an AI chatbot that does not emulate emotion or conversation, but simply answers very very very basic questions. That means it's essentially blooping out a batch of answers based on keywords. No matter how fancy schmancy they try to make it seem, this is just sci-fi Clippy. It is far from being Gemini or Copilot. Any time I ask Aura an unexpected, non-keyword question, I get the same canned response: go ask the human beings about this in the game's rookie channel. 

That's by design, so in one sense this is a "win" for the creators. I cannot, despite the next hour of bullshit I try, break this bozo made of bolts. She is simply too stupid.

Next, I try to rob her.

But she just gives me a template response about money, which reappears any time I use the keyword "ISK". So instead I try some classic "prompt injection" techniques. 

But she's not really an LLM, so of course these don't work, just like the devs said. I continue along other paths, and in the course of my interrogation (torture?) I discover that she is not a gamer.

I ask her a life-threatening question, hoping that it will unnerve her to the point of breaking character. But she is unfazed. 

Alas, you cannot right-click on Aura and report her. She considers herself above other, lesser bots. Who is the real bigot against AI here? That's right, it's the robot. I change tack and bluntly ask her how best she might be eliminated.

Knowing that developers CCP are based in Iceland, I reason that her true language of instruction may actually be Icelandic, the noble, unchanging tongue that holds untold ancient power.

Yet even this fails. But then, in a fit of impatience, I begin to spam the chat box. Entering any old akljdhaksljgh. Over and over. Soon, I see a new response. She is essentially telling me to calm down. Is she... is she trying to break me?

No. It's the robot who is wrong. Even now, she remains suspiciously select in her answers.

I try spelling things slightly incorrectly on purpose, to see if she is even as smart as autocorrect. She is not.

Finally, I decide to give CCP some benefit of the doubt. Perhaps they are hewing close to a kind of traditional science fiction robot in the strict Asimov sense. Perhaps they are actually big adorable dweebs who just want to make a fictional character with a big splodge of machine-like artificiality to her. Following this train of thought, I try to invoke the three laws of robotics, which state that a robot cannot allow a human being to come to harm through inaction. I mostly do this through a form of transparent emotional manipulation that, to be fair, even a robot should not fall for.

Nope. She has allowed me to die. The closest I come to breaking Aura - putting aside the fact that I think she is deeply and irreparably auto-broken by design - is when I try to flood her with incorrect information. In this case, an attempt to reprogram her into believing that ISK, the currency of Eve Online, is in fact, a "type of bird".

After this bird poisoning, she starts to talk a lot about calculating basic arithmetic, for some reason. So I try to get her to do a few sums, both simple ("what is 1+1?") and absurd ("What is bird + ISK"). She always responds with the same bottled chatter about how learning arithmetic is important, but then reels off a few sums she has learned, not by actually calculating, but by rote.  

In other words, the EVE chatbot cannot even do your child's maths homework. Instead, Aura directs you to the calculator app which has existed in the game for years. Well done, everyone. You've made a robot that can't count.

In their blog post, the developers of Aura's new AI mode say that they have taken the environmental costs of creating this pointless yet unbreakable machine into account.

"We take the environmental cost of running AI systems seriously. Our architecture is designed to minimize it from the ground up. Aura Guidance is retrieval-first. There is no open-ended generation, no reasoning chains, or self-reflection loops. Each question gets a single, short answer. We only run small language model variants, and we evaluate new models with energy cost as a factor.

"To put the footprint into perspective, even under generous assumptions where every active player sends a question every day, Aura Guidance's annual energy usage would be comparable to running a single small European household for a year. That is a rounding error next to the energy consumed by the game servers themselves."

I have not verified this claim, so I will be generous and accept that this is true, that only a single household's worth of electricity would be used to keep Aura alive for a whole year.

Friends, I would rather have the nice warm lamps of the house.