Every day, hundreds of people gather on Twitch to watch AI-controlled fighting game characters beat each other up: they bet fictional money on the outcomes, they scream in the chat when their pick loses, and they develop elaborate gambling superstitions about characters named after fast food mascots and anime villains.
And if you ask any of them why, most of them will tell you some version of the same thing: “I have no idea, but I can’t stop watching.”
Academics have actually studied this. And what they found says a lot about why SaltyBet works — and why betting on AI fights feels so weirdly compelling even when no real money is involved.
In 2024, researchers Rory Summerley and Brian McDonald published a peer-reviewed study in Games and Culture titled “Perceived Foolishness: How Does the Saltybet Community Construct AI vs AI Spectatorship?” It’s one of only two academic papers specifically about SaltyBet, and the first to be published in a peer-reviewed journal.
They analyzed Twitch chat transcripts from 15 SaltyBet tournaments — roughly 20 hours of footage and thousands of chat messages — to understand how spectators talk about the AI fighters, the bets, and the experience itself.
Their central finding? SaltyBet spectatorship is an “actively performed parody” of the fighting game community. Probably not a groundbreaking conclusion for those of us that have seen the chat, or participated. The viewers know it’s absurd. They know the money is fake. They know the fighters are AI running deterministic code. And they love it anyway — not despite the absurdity, but because of it.
The researchers coined the term “perceived foolishness” to describe what makes SaltyBet work. Drawing on the historical concept of the court fool — a figure whose “acknowledged defects are socially acceptable as a form of entertainment” — they argue that SaltyBet’s AI fighters serve a similar role.
The AI are entertaining precisely because they’re imperfect (X Tier sends its regards). They spam the same move over and over. They stand still while getting hit. They walk into corners and get trapped. They occasionally do something so unexpected that the chat explodes with “PogChamp” and “WTF.” And because they’re AI — not humans — there’s no guilt in laughing at their failures. Nobody gets hurt. Nobody’s feelings are wounded. The AI doesn’t know it’s being mocked, and that makes the mockery feel safe (though perhaps we should show some deference and affection to our cybernetic overlords — best be on their good side so Skynet remembers us fondly).
SaltyTrack’s data from over half a million matches confirms this pattern at scale. The crowd gets the winner right 67.9% of the time — which sounds decent until you realize that means roughly 1 in 3 favorites lose. A-tier has a 34.3% upset rate. Even when the crowd puts 90% of the money on one side, the underdog still wins 13.4% of the time. The chaos is baked in, and it’s what keeps people watching.
One of the study’s most interesting observations is about the betting itself. SaltyBet uses Salty Bucks (S$) — a fictional currency that can’t be exchanged for anything. When you lose all your money, the system gives you a small bailout so you can keep betting. There are no real stakes.
And yet the emotional investment is genuine. As the researchers put it, betting in SaltyBet functions as an “ironic semblance” of gambling. The spectators know it’s not real gambling. They joke about it constantly. One viewer in the study quipped, “if I lose this bet I will lose my house and my wife will divorce me.” The Kappa emote — Twitch’s universal symbol for sarcasm — follows nearly every declaration of betting strategy.
But beneath the irony, people genuinely care about their Salty Bucks balance… something about those pixels in your screen hit you right in the feels. They celebrate wins with “WtfSalt” and “Kreygasm.” They mourn losses with “BibleThump” and “NotLikeThis.” They rage when a fighter they invested in gets bodied by a character with “bad AI.” The fake money creates real engagement because it gives you skin in the game — a reason to pay attention to the outcome instead of just passively watching.
This is something we see in our own data. During the betting window — which lasts about a minute — spectators have almost nothing to go on. They can see the characters’ idle animations and their names. That’s it. No stats, no win rates, no head-to-head history. They’re making gut decisions based on appearance, franchise loyalty, and vibes. This is exactly the gap SaltyTrack was built to fill — replacing guesswork with data during that critical minute.
The study identifies several gambling fallacies that the SaltyBet community has elevated to the status of ritual. The most famous is “Always bet on DBZ” — the belief that Dragon Ball Z characters are always the strongest pick. It originated in early streams when DBZ characters tended to be overpowered, and it’s been repeated so many times that it has its own Twitch chat command.
The community knows it’s a fallacy. They mock it openly. The chat command “!dbz3” returns the message: “Always sometimes never bet on DBZ…maybe.” And yet people still type it. Still believe it. Still bet on Goku because he’s Goku (and the data has a lot to say about that).
SaltyTrack’s database lets us test this directly. Across 580,000+ matches, Dragon Ball Z characters have appeared in over 25,000 matches and won 48.3% of them. That’s slightly below the overall average — DBZ characters are, statistically, average performers, which, not going to lie, really surprised me. “Always bet on DBZ” is demonstrably wrong, and the community seems to be in on the joke.
The same pattern holds for other beloved fallacies. “Always bet on sword” — characters with swords do win 56.1% of the time (swords mean large hitboxes, which genuinely helps), so that one has some truth to it. Touhou characters (“2hu” in chat slang) win 50.7% — essentially a coinflip. The fallacies persist not because they’re accurate, but because they’re fun to believe in. They give spectators a framework for making decisions when they have almost no information, and they create shared ritual moments in the chat.
Most Twitch streams revolve around a human personality — a streamer who talks, reacts, and connects with the audience. SaltyBet has none of that. The person behind it (known only as “Salty”) never appears on camera, never speaks, and rarely interacts with the community directly.
Researchers Johnson and Jackson (2022) call this the “agency gap” — when there’s no human agent driving the content, the audience steps in to fill the void. Figuratively speaking, I guess people can’t sit still in a quiet room, and have to fill the silence with some kind of noise.
In SaltyBet, this means the chat becomes the commentary desk. Spectators assign personalities to the AI. They create narratives about underdogs and champions. They cheer for characters based on franchise loyalty, visual design, or past performance — even though the AI has no awareness of any of it.
The paper found that spectators regularly anthropomorphize the fighters. They call them “smart” or “dumb.” They accuse them of “scamming.” They describe a fighter that stands still as “asserting dominance.” One spectator, watching a poorly performing Darth Vader, simply said: “this darth vader is so stupid haha.”
This is where data tools like SaltyTrack change the dynamic. Instead of betting on vibes and names, viewers who use the extension see actual win rates, head-to-head records, and AI-generated predictions during that betting window. We don’t eliminate the storytelling — people will still root for Goku because he’s Goku — but we add a layer of information that didn’t exist before. In academic terms, we’re reducing the “information asymmetry” that the researchers identify as a key source of suspense.
Here’s the philosophical angle that the study raises and it stayed with me: SaltyBet resembles ancient forms of nonhuman combat entertainment. The researchers draw parallels to Balinese cockfighting (studied by anthropologist Clifford Geertz) and Roman gladiatorial games — spectacles where audiences bet on the outcomes of fights between non-human combatants. ARE YOU NOT ENTERTAINED?!
The key difference? Nobody gets hurt in SaltyBet. The AI have no moral patiency. They can’t suffer. They don’t know they exist. As AI ethics researcher J.J. Bryson argues, AI are “extensions of the user” — tools, not beings. You can watch a pixel character get bodied without any ethical discomfort, so enjoy!
The study suggests this is actually part of the appeal. SaltyBet offers the visceral thrill of combat spectatorship — the suspense, the upsets, the emotional investment — without any of the moral weight. It’s “morally permissible nonhuman competition supported by a ritually foolish tone.” You get the dopamine hit of watching something fight and win without anyone actually getting hurt.
The Summerley and McDonald study was based on 15 tournaments from July 2019 — a snapshot of the community at a single point in time. SaltyTrack has been collecting data continuously since December 2021, across over half a million matches and 10,000+ fighters. Our dataset lets us test some of the paper’s observations at a much larger scale.
The crowd’s overconfidence is consistent. The study describes spectators betting based on character recognition and franchise loyalty. And I can’t fault them though, this is pretty much what made me create SaltyTrack to begin with — I basically thought, “surely there has to be a better way.”
Our data shows this plays out predictably: the crowd picks the winner 67.9% of the time overall, but drops to just 65.7% in A-tier where character recognition is less reliable (more obscure fighters). The crowd is best in S-tier at 70.0%, where the roster is more familiar.
Upsets are not random — they’re structural. The researchers noted that spectators are drawn to upset narratives. Our data shows that upset rates vary significantly by tier: A-tier produces upsets 34.3% of the time (the highest chaos), while P-tier sits at just 24.6%. If you want drama, watch A-tier. If you want predictability, watch P-tier.
The betting window is an information vacuum. The study describes spectators having “only the physical appearance of the AI’s idle animation and name as information” during the betting phase. This is exactly the problem SaltyTrack solves. Our extension fills that window with win rates, matchup history, confidence levels, and an ML prediction — turning a gut-feel gamble into an informed decision.
So why does SaltyBet work? The research and data together point to a recipe:
Add it all together and you get something that shouldn’t work but absolutely does: a 24/7 stream of AI fighting game characters that’s been running since 2013, still pulls 150+ concurrent viewers, and has inspired its own academic study, fan wiki, and — yes — a data analytics platform with a Chrome extension.
The researchers call it “unique, but foolish.” We’d call it home.
Want to see what data looks like during the betting window? The SaltyTrack Chrome Extension shows you win rates, head-to-head records, and AI predictions in real time — right on saltybet.com. Free, no account needed.
SaltyBet uses virtual currency only. No real money is wagered or exchanged. SaltyTrack is not affiliated with SaltyBet.