Generative AI's technological and business model are designed around a kind of gambler's logic. Here, I am talking about general purpose GenAI marketed as being a general problem solver in a number of domains. I am not including here AI/LLMs/GPTs etc developed for highly specific domain areas. Basically, these GenAI chatbots transform work and research questions across many domains from questions of judgment into problems of probabilities.
We can see a lot of similarities between the logics here with the logics of other closed circuit currency systems on platforms like Twitch and TikTok, particularly the fluctuation of cost and value and their use in encouraging more participation. Hell, you could even compare the platforms to Chuck E. Cheese, which itself has been transformed into a postmodern children's gambling den filled with more screens than slides.
Tokens are closed circuit currencies only expendable in a specific network or a limited geographic area. Dollars are spent on credits which are spent on tokens. Output is also measured in tokens. Generative AI tokens can only be "spent" on the platform from which they are purchased; you put in dollars and receive a different currency. Tokens and money never flow outward from the house. It's monopoly money with only 1 bank. These tokens have fluctuating values largely disconnected from the cost of doing business. The true token cost has been subsidized by the AI companies in order to get people on their platforms. A token economy for a closed system.
The token system encourages users to think of productivity in terms of token expenditure, instead of problems solved.
AI tokens do not scale effectively. Longer problems require more tokens and more computing power. There does not seem to be gained efficiency per expended token over time (you cannot get more with less). The output does not return tokens, but costs tokens. Any output, whether a winning or losing one, is not a portion of the house's winnings or bank returned to the player. The output just costs more tokens.
Once acquired, it is really easy to spend more in tokens than you actually purchased. The design of these platforms always encourage "Just one more pull" of the lever. Since Generative AI are predictive models that are estimating the next most statistically likely output, each prompt is a throw of a weighted dice and gambling the cost of each token that the next output will be what one ultimately wants. The platforms are designed to make this techno-economic transfer process difficult to see. Like continuity editing in the movies, the platforms use aesthetic and professional techniques to hide the messy complex production process of its output. Like video games, they provide an illusion of agency. Video games are also saddled with accusations of gambling in loot boxes and RNG rewards. Unlike video games, however, statistical prediction promises real world action. In this way, they become very similar to the addictive slot machines described by Natasha Schull. Like Schull's Vegas slot machines, the gambling logic is part of the technology and enmeshed into its built environment (platform ecosystem). Last I checked, Grok has no timestamps.
Like slot machines, the interface allows for easy and rapid repetition. But, like poker or blackjack or other games that are influenced by luck and skill, GenAI also suggests that one can exert control over probabilities and win by developing savvy prompting skills. While this is presented as a specific cognitive skill that makes use of a natural language processing system, many of these platforms require users conform to highly specific, increasingly formalized forms of interaction and linguistic patterns in order to get desired results. Because this software is voracious for new data based on interactions with users, it almost always asks more questions in its friendly and servile tone, like a casino floor worker asking if you want anything from the bar. While users must conform to the gambling logic, they are also prohibited from becoming too knowledgeable about the running of the casino. "You also may not attempt to reverse engineer, extract or replicate any component of the Services, including the underlying data or models (e.g., parameter weights)." Counting cards is still forbidden. Together, this seems to create the back-and-forth flow-like, immersion state prized by gamers, gamblers, and token-maxxers.
In other words, the gambler's logic and interfaces convinces users that they are using judgment to solve problems when in reality are actually using probability.
In this way, GenAI platforms are what Morgan Ames calls "charismatic machines," machines whose imagined uses and users are more important than their actual users or practical uses. GenAI is highly charismatic, convincing many users that they are saving time, money, and energy when the opposite is likely to be true. The fever dreams of GasTown or the worries of "token anxiety" are good examples of how charismatic these token machines are. When workplaces measure employee performance in token expenditure, that's the charisma machine. In all of these cases, the assumption is that higher productivity is related to higher token costs, even when companies find this to be financially challenging in the long term.
Calling this form of engagement a gambling problem has its own drawbacks. It can lend itself to a sort of Victorian moralism. And, like social media, the evidence suggests habitual use is more of a discursive and media problem than it is a problem of addiction. So, I am not saying that GenAI always causes addiction. Its sheer ubiquity will help surface all sorts of pre-existing or latent dispositions, like the TV did when it entered American homes.
What I am saying is its design, language, and interfaces lends itself a gambler's logic of "just one more pull."