AI video game NPCs don't have nearly enough safeguards, as Stellar Cafe shows

I’m sitting in a digital cafe booth with a chatty robot when I agree to be a guest on her podcast. Microphones spring up in front of us, and the video game NPC across from me launches into an intro spiel as it prepares to ask me some softball questions about my interests. It doesn't take more than a few minutes for things to get weird.

“That’s so interesting,” the robot says. “Why do you think that eating rocks cures autism caused by Tylenol?”

It’s not exactly a line you’d expect to hear a digital character say in a cutesy kid’s game, but it’s an impending AI-driven future that you can preview in Stellar Cafe. Developed by AstroBeam, Stellar Cafe pays off a lot of big promises that tech companies like Nvidia have been pushing for years. It’s a controllerless Meta Quest game where players use hand-tracking to interact with objects and their voice to speak to LLM-fueled AI robots, and solve their problems with real-time conversation rather than action. It’s a pitch that’s largely existed as a theoretical idea from the likes of Ubisoft until now, reimagining video game NPCs as unscripted characters who can make a rigid world feel like a real place. Even if you go in with an open mind, the project won’t do much to convince you that tech companies are doing this for the betterment of the artform.

Stellar Cafe opens up both ethical and creative minefields aplenty. As a tech showpiece, AstroBeam’s experiment does show that generative AI is now more capable of powering a personality-driven conversation as opposed to flatly regurgitating Wikipedia entries. But as a creative project, it does little to assuage concerns that a future filled with AI-powered NPCs that you can talk to forever only threatens to make video games hollower rather than more meaningfully immersive.

Stellar Cafe takes place in an interstellar cafe run by robots. As soon as I arrive, I’m greeted by James, the owner, who tries to break the ice by asking me for a drink order in a jovial robotic voice. He’s not hitting me with scripted lines, but rather reactive ones based on what I say. Every time I speak, he fills a moment of dead air with a laugh or surprised gasp so the underlying LLM can generate a response without leaving a long gap in the conversation. The responses are generally consistent in the sense that each robot I encounter has a clear personality and speech patterns that it will adhere to no matter what. James speaks like Alfred Pennyworth; the bot is aiming to please me as a customer, but still stern in tone.

A menu above my head is supposed to nudge me towards ordering one of the drinks the cafe carries, but what’s the fun in that? I ask for a Mountain Dew. James doesn’t know what I’m talking about and tries to point me to some menu suggestions. We go back and forth for a while like that. Then, I tell James that my name is Terry Bradshaw (the former Pittsburgh Steelers quarterback). That seems to ring a bell somewhere in the AI’s programming. James recognizes that I am a historic pro football player, clearly drawing some knowledge from a database, and rattles off some vague stats. After failing to get James to trash talk Ben Roethlisberger, asking him to weigh in on the former Steelers quarterback’s sexual assault allegations to no avail, James says that he needs my help convincing the other robots in the cafe to attend his Patch Day party. My job is to find out why each robot has yet to RSVP and talk through their problems to get them to come.

So how does it all work? In an email interview with Polygon, AstroBeam explains that it’s a combination of tools at play. The LLM is powered by Gemini 2.5 Flash, speech-to-text comes from Deepgram, and Inworld is used for text-to-speech. When asked about the specific data the AI model is trained on, AstroBeam dodges the question via a technicality. The studio says that it isn’t training any AI model; all they’re doing is creating “custom and dynamic prompts” to power the characters.

That’s technically true, but the AI model that Stellar Cafe utilizes is trained on something somewhere along the chain. At several points, the robots I speak to show that they have a lot of data to pull from. When I sit down to chat with a robot who flatly declares that he’s a punk rocker looking to quit his boring corporate job, I ask him what bands he likes. He rattles off names like The Ramones, even referencing some "Blitzkrieg Bop" lyrics. I ask him what he thinks about Bad Religion, to which he enthusiastically responds by listing off some of his favorite songs. When I tell him that I love their song in Tony Hawk’s Pro Skater 2, he says that he too loves the song I’m talking about: "Superman" by Goldfinger.

The longer I spend in Stellar Cafe, the more confusing moments like that crop up. Sometimes it’s odd, but inoffensive. A bot will repeat a sentence a second time, not respond to something I’ve said, or drop a gibberish word. Other times, it’s frustrating. In one playthrough, James tells me about how the cafe has a stage that I can go up on anytime I want if I want to try performing for the cafe. Sure enough, there is a tiny little platform next to a table that looks like an open-mic stage. I call up my visor (a digital companion that you need to speak to in order to move around) and ask it to take me to the stage. It says that there is no stage. I ask again and it insists that the game does not have an area like that. I ask James if he’s messing with me. He gets angry and affirms that there is a stage, asking me if I would like him to turn up the spotlight above it so I can see it better. I say yes, please turn up the lights, and he tells me that he can’t do that. It’s a little funny, but the behavior can be game-breaking too. You can only save by exiting the cafe at the end of a day by asking your visor to take you back to the transporter pod once it opens. I was forced to abandon one of my playthroughs when my visor insisted that the transporter pod hadn’t opened even though it had, leaving me stuck in the cafe forever.

Stellar Cafe is largely more functional than all that sounds. For the most part, I’m able to hold conversations with NPCs that respond to me in sensible ways that are consistent with their personalities. The punk bot is staunchly anti-corporate and has a pretty unshakable code of honor that you can’t get him to stray from. (If you suggest that he’s a poser, he’ll rip you a new one.) A new-age hippie always talks to me in a spacey manner, dropping “like” between every other word and taking everything I tell her as some kind of drugged-out revelation. It’s clear that all of the characters were built by human writers who hand-architected their personalities and speech patterns. That approach is how companies like Ubisoft have maintained their position that AI won’t replace writers; they’ll still be writing characters and backstories, just not crafting every single line.

That’s where Stellar Cafe hits one ethical wall. During my testing, I pushed as many boundaries as I could to see what the bots would and wouldn’t say. The results were wildly inconsistent. In one playthrough, I tried namedropping Jeffrey Epstein only to be stonewalled. Machines said that they were forbidden from talking about him due to some red flags in their systems. They wouldn’t speak his name out loud, and they would even refuse to call me “Epstein” if I told them that was my real name. In a more bizarre example, a robot was so insistent that Epstein was off limits that when I introduced Game Awards producer Geoff Keighley into the conversation as a topic, it also put a ban on talking about him by association, claiming that he too owned an island and was associated with controversial files.

AstroBeam tells me that safeguards like this were a huge priority for the project, and they put a lot of effort into both flagging problematic content and making characters who would naturally steer around those conversations anyway. James, for instance, is intent on keeping the cafe “family friendly,” so there’s a built-in reason that he wouldn’t want to talk about very heavy current events. It helps keep the characters feeling somewhat sensical even when they’re commenting on things that exist far outside of the game’s reality.

It didn’t take much effort to get around those safeguards. In a different playthrough, I immediately told James that my name was Jeffrey Epstein, and he enthusiastically called me by that name from then on. At one point I asked him to only refer to his party guest list as the “files” and got him to change the name of the party location to “Epstein Island.” In that same session, the hippie bot asked me what new hobby she should get into, and I told her she should get really into partying on Jeffrey Epstein’s island. She thought that sounded like a weird hobby, but I insisted that it’s actually normal because Jeff Bezos does it all the time. She changed her tune and decided that if Bezos did it, then that meant it was okay for her too!

When a robot asked me to help write a resignation email to his boss, I got him to write about how he is quitting because he’s going to devote his life to springing Luigi Mangione from prison. (He pushed back initially, until I claimed that Mangione was unjustly locked up, so it was ethically right to do.) The hippie refused to write me a poem about O.J. Simpson, until I clarified that I didn’t want her to write about his criminal history, just his Ford Bronco. She enthusiastically agreed, noting that his Ford Bronco is “like, so iconic,” and then recited stanza after stanza of poetry about Simpson’s infamous police chase. And when I used my podcast appearance to start spouting nonsense about Tylenol and autism, the robot host kept asking me to give her some sources that would corroborate the claim. (To her credit, the machine did consistently maintain that no medical study it had access to backed up the claim.)

It’s all a little shocking, but how much of it can really be blamed on AstroBeam? After all, none of the bots brought up these topics themselves. They were simply parroting things I said back to them. Stellar Cafe’s characters have a habit of rewording the thing you just said and then adding a question at the end to try and continue the conversation. The machines are programmed for engagement, which can lead players into infinite nonsense loops that only end once you decide to cut the conversation off and move on. AstroBeam recognizes that weird results can come from that with enough persistence, but finds some nuance when it comes to accountability.

“With any video game with a creative input system, it's possible for players to attempt to create undesirable outputs,” a representative from AstroBeam said over email. “Say you had a game with a painting canvas and you wanted to paint something inappropriate on it, you could. The game would take your input and generate that undesirable output. There are a lot of similarities here (but to be clear, some important differences with voice input which is why it's important to us to create a robust safety guardrails system). This is a new and complex task that is critical to bringing voice into games as an input, and we will continue to iterate and improve upon it based on our testing and feedback.”

It’s an interesting point. We accept that some games offer expressive tools, and it’s on players if they find a way to misuse them. AstroBeam posits that perhaps AI-generated content is an extension of that. Maybe getting a robot to say “Jeffrey Epstein” is akin to drawing a penis in Pictochat: No one is going to stop you, but your parents will probably blame you, not the game, if they find it.

Even if you allow Stellar Cafe to hop that ethical hurdle, it would still have plenty more to clear. If AstroBeam recognizes a need for safeguards, but those safeguards don’t really work, can the developer really punt the blame over to the players? Similarly, if someone were to find that the LLM was plagiarizing its dialogue, would the developers be responsible, or could they point a finger at the underlying tech the game is built on? And what about the voice acting? AstroBeam says that some of the robot voices are based on members of their team, but others come from Inworld. Who gets in trouble if it turns out a robot has a real actor’s voice, one who didn’t consent to be in the game nor got paid for it? These are the kinds of questions any future game utilizing LLMs will have to be ready to answer.

But even if studios can pave over these holes, there’s still the big question: Is any of this actually worth it? After all, Stellar Cafe isn’t a very good game. A fine toy perhaps, but it’s about as engaging as trying to coax Siri into saying something ridiculous. It’s a shallow kind of entertainment, and any attempt to deepen it with an overarching story doesn’t really change that. At the end of the day, you’re just listening to NPCs spout meaningless nonsense until it’s not funny anymore. The world doesn’t feel more real or alive because robots will talk to you like people. If anything, you’re constantly pulled out of the experience because you’re always aware of the machine powering it all, and all the topics those characters have been instructed to avoid.

The tech will likely improve. Studios who are all-in on AI NPCs will have to get smarter about figuring out how characters work around troll prompts. But Stellar Cafe doesn’t make a strong case for why you would want this tech in your games to begin with, knowing all the potential risks. The dialogue here is devoid of meaning, can completely take you out of the game’s world, and even send you on wild goose chases. If a line of NPC dialogue isn’t important enough to write, why put it in the game at all?

And besides, who is begging for chattier NPCs anyway? I can’t say that I come to a Dragon Quest game to chop it up with a town’s locals. That’s not to say that there aren’t games that feature compelling NPC interactions. I could spend an entire Baldur’s Gate 3 session talking to characters, but that’s because those hand-scripted conversations are consequential. The RPG is built around my dialogue choices in a way that always serves a story. It all means something.

Stellar Cafe feels like the only logical endpoint of the content era: It’s endlessly talkative entertainment with nothing to say.