International investment firm Morgan Stanley downgraded price targets for a trio of poker-related stocks this week. The reason?
According to analysts, the industry faces a long-term threat from superhuman, poker-playing, artificial intelligence.
Bots are discussed frequently within both the online poker community and computer science circles, as the game poses the sort of difficult-but-defined problem that’s perfect for AI research.
They’re rarely mentioned elsewhere, however, so it’s curious to see gambling AI arise in a financial context.
Here are the three stocks analyst Ed Young and his team focused on:
Those three are the owners and operators of PokerStars, partypoker, and the iPoker Network respectively. Despite noting the potentially serious impact of AI on these company’s business, the downgrades were minor.
Analysts reduced price targets for GVC and Playtech by just a few points, while adjusting The Stars Group down more than 5%. TSG’s diverse portfolio has perhaps the most exposure to poker and the most upside to counteract the threat, including the recent launch of Fox Bet in New Jersey and Pennsylvania.
The report did not mention any other gambling operators with less-significant poker verticals.
This oddly includes 888 Holdings, despite the fact that 888 Poker remains the fourth-largest dot-com poker site in the world. Poker is performing so poorly compared to 888’s other verticals that its price targets perhaps already assume that the segment won’t be a critical part of the business in future.
The main cause for concern for Young and his team is the recent unveiling of Pluribus, a joint effort by Facebook AI and computer scientists at Carnegie Mellon University. Carnegie Mellon is one of two universities known for its research in poker AI, along with the University of Alberta in Canada.
Pluribus isn’t the first ultra-strong poker AI by a long shot, though.
The Alberta team announced in 2015 that its bot Cepheus effectively solved heads-up limit hold’em. The same year, Carnegie Mellon’s Claudico put up a good fight in heads-up no-limit against a team of top human pros. It lost narrowly, but its successor Libratus clobbered those same players less than two years later in January 2017.
That said, Pluribus differs from its predecessors in a couple of important ways. More than any bot before it, this new generation represents a concrete threat to the actual business of poker.
Previous bots focused solely on heads-up play, for instance, which accounts for only a small share of online poker traffic. Pluribus plays (and wins) the far more popular six-max format by default.
Both Claudico and Libratus additionally required the use of supercomputers, with Claudico taking three million core hours of computation to develop its strategy. For Libratus, that number was a whopping 15 million hours.
Pluribus, on the other hand, runs on a standard computer and required just 8 days and $144 worth of cloud-based training to develop its strategy. That makes the program easily within reach for a would-be bot user — even one targeting low stakes.
This may be the first time that the investment world has taken a serious look at whether AI poses an existential risk to the business of online poker. It is, however, a problem that players and operators have worried about since the boom years.
Virtually everyone agrees that the threat is real, but as for its exact nature and severity? That depends on whom you ask.
Some would agree with the Morgan Stanley analysts — that an accessible, flexible and superhuman AI that can beat any number of humans is a game-changer. To support that argument, we can look at online backgammon.
More to the point, we can look at its conspicuous absence. Backgammon was, at one time, a popular game among gamblers. Although 888 and others did try offering real-money online backgammon at one time, nothing ever came of it.
Unlike more complex games, you see, computers were already capable of beating humans at backgammon long before online gambling came along. It was more or less understood from the start that many players on real-money online backgammon servers were either bots or humans inputting computer-generated optimal moves.
The game was therefore pretty much dead on arrival for online play.
Poker sites and their security teams, on the other hand, will tell you that it’s not really the strength of a bots that’s the problem. Rather, it’s how easy or hard a bot is to detect that determines its threat level.
Because online backgammon was never profitable to begin with, there was very little investment in policy development and site security. Things are different for online poker, of course, which remains a profitable gambling business.
Any good poker player knows that you don’t have to beat everyone at the table to be a winner. You only have to beat the weakest opponents. And poker bots capable of turning a profit have existed for years at the lower stakes.
A Pluribus-style bot might be able to beat higher-stakes games than most, but such a bot would run a significant risk of detection. Traffic is low in those games, and the majority of players at that level know each other and their gameplay.
Operating a high-stakes bot also requires keeping a much larger bankroll on the site, meaning much bigger liabilities for the operator if the bot is detected.
The real technological battlefield is not the poker table itself, but the cat-and-mouse game of bot detection. The warring parties are not the bots and the players, but the bot developers and site security teams.
As for who’s winning that battle, it’s impossible to say. The parties involved are never going to fully lay their cards on the table.
Bot-ers won’t disclose the measures they use to avoid detection for obvious reasons, and site security teams similarly won’t talk about how they catch them. To do either would make the job of the other easier.
There has been a refreshing bit of transparency from GVC/partypoker of late, at least, in that it has begun disclosing the number of bot accounts closed each month. That tally reached a peak of 142 in March and has generally been coming down since then for the dot-com site.
On the surface, it seems like the team has the problem under control. However, the number of bots is unknown by definition. It’s possible that security has simply picked off the low-hanging fruit, and the bots that remain might just be better at avoiding detection.
Ultimately, perception may be more important than reality when it comes to bots.
Most players will lose less money against a theoretically perfect — but non-adaptive — bot than a strong human opponent seeking to exploit their specific weaknesses. Bots therefore don’t pose a huge threat to the ecosystem by themselves, no matter how strong they are.
According to Young, the real “plausible risk” is that fear of bots will drive players away.
Morgan Stanley is probably correct to weigh the artificial intelligence of Pluribus when setting its price targets. It’s not that the tables will be overrun by Pluribus clones any time soon. Rather, it’s the kind of story that stokes fear and threatens to drive fence-sitting players out of the game for good.