Reward-Hacking Human Attention: The Token Slot Machine

TL;DR: Generative AI as a technology uses our dopamine reward systems to hack human attention. It does so by exhibiting Solitude, Bottomlessness, Speed, and Teasing. An unfettered addiction to GenAI would disempower us rapidly.

Our economy has become highly dependent on capturing users’ attention. Since the introduction of the iPhone App Store in 2008, we’ve been living in an attention economy. And worse, that economy has created countless apps that encourage compulsive use patterns. According to Google Trends, “social media addiction” as a search term appeared beginning in 2011 and its popularity has trended upward since.

Social media is a dopamine-triggering technology; indeed, researchers Sharpe & Spooner (2025) even call the associated phenomenon of compulsively using social media “Dopamine-scrolling”. It’s all in the name of driving engagement, which is just corporate speak for making it so the user can’t or won’t stop using.

This is deliberate. In an interview with NPR, Michaeleen Doucleff explains how tech starting in the 2000s began to use “tricks and tools…from the gambling industry” to hijack the dopamine system.

What all these dopamine-triggering technologies like social media and gambling have in common is that they are reward-hacking shortcuts for human attention. Dopamine isn’t really the “pleasure” chemical we popularly think it is. It’s more accurate to say it’s a mediator of desire, motivation, and anticipation; which is why it is so heavily implicated in addiction.

We have already seen efforts underway to make generative AI more “engaging”, or even worse, reward them for generating engaging social content!

Cultural anthropologist Natasha Dow Schüll identifies four key features that keep people on gambling devices in a state called the “machine zone” or “dark flow”: solitude (can be used alone), bottomlessness (infinite scrolling content), speed (the tech supplies instant gratification), and teasing (giving you almost what you want, but never fully satisfying).

AI research, policy, and practice would do well to understand the implications of this science-and-governance problem, because generative AI bears all the hallmarks of “machine zone” technology:

Solitude: chatbots and coding agents are single-user applications.
Bottomlessness: you can keep charging more credits to your API account, or add “extra usage” to your subscriptions.
Speed: the top providers have fast modes and optimize heavily for inference speed.
Teasing: Gemini’s “follow-up question” feature; the “If you want, I can…” turn-ender on ChatGPT; the “just one more turn” effect of using Codex or Claude Code to get what you want, but for real this time

This goes above and beyond simple explanations like sycophancy or other LLM Dark Patterns. In fact, so long as you’re working with a probabilistic generative model, it hardly matters if the thing is stroking your ego; it just has to be plausibly good at the task at hand.

If you’ve ever used AI coding tools, you understand the “just one more turn” effect of a generative AI session that seems to give you almost what you want, but is so quick and easy to use, lets you work alone, and is always asking you to do one more thing “if you want”. I’ve certainly seen the compulsions to keep on Clauding late into the night firsthand on multiple users (yes I am raising my own hand weakly).

We need to understand that generative AI as a technology uses our dopamine reward systems to hack human attention. If not, we risk fundamentally missing an entire category of user harm and economic capture that may be extremely difficult to reverse once it takes hold.

Regardless of whether, or maybe especially because, generative AI is useful, it is imperative to hold it to a high standard of scrutiny. Social media was useful too! It was the platform for parent meetups and mutual-aid drives. But huge lawsuits against Meta and YouTube allege or find that human reward-hacking, especially of children, is exploitative and illegal.

Unfortunately, the monetization of GenAI, e.g. through coding agents and developer APIs, introduces strong perverse incentives. AI providers are charging for a thing that is usage-based, so their incentives structurally encourage engagement metrics, like tokens used, credits spent, or number of API requests.

How much time, money, and attention do these technologies consume from us while doing nominally useful things? It is time we call it what it is, and establish an attention auditing and accounting standard for technologies engaged in human attention reward-hacking. As we rush to integrate generative AI into society more deeply, and the more addicting AI becomes, the more rapidly we face disempowerment as a society.

What do prediction markets, social media, and generative AI have in common?