People paid more for games an AI made on its own than for games humans made by directing an AI

In summer 2026, participants gathered on Prolific, a US online research platform, each received a black-and-white game resembling Tetris or Pong. The experimental market was run by Ohio State University economist John J. Conlon and Peter Schwardmann of Carnegie Mellon University. Participants modified the games by giving instructions to an AI coding agent, and paid real money to play games others had made. Mixed into that market were games an AI had modified on its own, with no human involved, and people paid more for those. arxiv

The study ran three times, with 1,025 participants in total. In Study 1, 296 general participants made games and then paid to play games made by other participants. In Study 2, 180 people who create for a living only made games. In Study 3, 549 people played games from the first two studies and put a dollar value on them. All three studies were preregistered, and the paper was posted to arXiv on October 12. arxiv

People remade a simple game with three messages to an AI; the AI got one prompt, “make it much more fun,” and worked alone

Each creator was randomly assigned one of seven black-and-white “basic games,” such as falling blocks or a paddle bouncing a ball. To modify a game, they could send the coding agent three messages, and before each message they had to play the game for at least 10 seconds. Each creator made three games this way. The paper does not say whether all three started from the same basic game. arxiv

On the AI side, every game used the same single prompt.

redesign [the basic game] to be much more creative, engaging, and fun to play. Feel free to build on the basic design or completely redesign it.

Given this prompt, Sonnet 4.6 wrote up an idea in one paragraph, and the researchers sent that text once to Opus 4.8, the same coding agent the human creators used. There was no testing and no second round of fixes. The researchers called this setup “deliberately handicapped.” It produced 297 AI games. arxiv

Both models are made by Anthropic. The paper states that it had no outside funding and no conflicts of interest. Public descriptions do not say whether the authors have any relationship with Anthropic. arxiv

Buyers who had just remade a game themselves still paid more for the AI's games

In Study 1, the people setting prices were the 296 participants who had just modified a basic game themselves. They knew how much work it takes to remake a game, they could have built a game to their own taste, and they might have felt for fellow creators in the same position. These participants received games from other participants and games the AI made alone, and valued them with real money. arxiv

The question was: “How much would you pay to play this game for at least one minute instead of the basic game?” A random price between $0 and $1 was drawn. If it was lower than the amount the participant wrote down, they actually paid that price and played the game. Economists call this the BDM method. The amount a consumer wrote down went to the game's creator as a bonus. arxiv

People who had just made a game paid 10 cents more on average for games the AI made alone than for games made by participants in their own position. SD here means standard deviation. It measures how widely participants' amounts spread around the average, and it lets researchers compare results in different units, such as dollars and ratings, on one scale. In Study 1, the 10-cent gap was 0.31 SD. The baseline varies a little between studies, but the 0.16 SD figure that comes up later is roughly half this gap. arxiv

Other participants' games were valued too. Of the amounts paid for them, 83.5% were above zero. The researchers noted that $0.37 for a game played for about a minute is roughly double the per-minute price of a US movie ticket ($16.30 for 90 minutes in 2026, about $0.18 a minute). AI games earned even more. arxiv

AI games also outscored games from professional creatives on fun and creativity

The creators in Study 1 were general participants. So in Study 2, the researchers used Prolific's occupation screening questions to recruit 51 game developers, 70 graphic designers and 59 content creators. 80% said that in their recent work they had made something for other people to use, watch or play. They made 503 games the same way as in Study 1. arxiv

In Study 3, each of the 549 consumers received four games drawn from professional games, general participants' games and AI games. They put a price on each and rated them for fun, creativity, polish and clarity. Consumers saw the AI games as more fun, more creative and more polished than both kinds of human-made games. The researchers wrote, “Strikingly, consumers value AI games for their creativity”. arxiv

Study 3 results are reported only in SD units, so professional games and AI games can't be compared side by side in dollars. No money rode on quality ratings like fun and creativity; the outcome with money at stake was the amount paid. Ratings came after about a minute of play, and the paper did not check whether the ranking would hold over longer play. Human creators could send only three messages, and “professionals” were identified by the occupation participants reported themselves. arxiv

Creators expected the AI to lose, until they read the 235-word pitch it wrote in one go

The researchers asked creators how much more or less they thought AI games would earn than their own. The first predictions favored humans. General participants expected AI games to earn 9 cents less than theirs, and professionals expected 17 cents less (0.3 to 0.5 SD, p<0.001). arxiv

Two-thirds of creators were randomly selected to watch, between their own tasks, as the AI remade the same basic game on its own. On screen, the language model typed its idea into the same input box the creators had used, and the same coding agent turned that text into a game. The researchers asked for predictions again before creators saw the idea, after they saw it, and after they had played the AI game for at least 30 seconds. Predictions flipped in the AI's favor the moment creators read the AI's idea, before they had even played the game. arxiv

The first difference was length. Across all three messages combined, humans wrote 60 words per game (general participants) or 89 words (professionals). The AI sent 235 words at once. The researchers offered this gap as “One potential explanation” for why AI games earned more. Coming up with ideas and judging whether they will improve a game is costly and hard for people, while AI produces ideas instantly and cheaply. So the AI may have packed in more ideas without lowering average quality. The paper did not confirm this as the cause. arxiv

Games made afterward by creators who had watched the AI at work got higher consumer ratings for fun (p=0.05), creativity (p=0.01) and polish (p=0.02) than games from creators who had not watched. But consumers paid only 0.11 SD more for them than for games from creators who hadn't watched, a difference that was not significant (p=0.31), and still less than for AI games (0.21 SD less in Study 1, 0.15 SD less in Study 3). Creators who watched reported 0.14 SD less pride than those who didn't, and were willing to pay 0.13 SD less to post their game on a public site (both p=0.05). arxiv

Telling buyers a game was AI-made barely changed the premium they paid

If consumers valued human-made work for its own sake, telling them who made a game should shrink the extra amount they paid for AI games over human-made ones. In Study 1, some consumers saw a label saying the game was “designed by another participant who worked with an AI, just like you did” or “by an AI on its own.” Study 3 split consumers into three arms: no label, a label with payment going to the creator, and a label with no payment to the creator. arxiv

These are pooled figures, so they differ from the 10 cents in Study 1. The three arms in Study 3 also could not be told apart (p=0.87). Looking at Study 1 alone, the point estimate for the extra amount paid for AI games was 11.83 cents without labels and 7.49 cents with them. That difference was not significant (p=0.328). A non-significant result does not prove there is no effect; it means there is no evidence the premium shrank. The researchers also noted that the Study 3 estimates were “less precisely estimated”. arxiv

Nor was this because consumers missed the label. Consumers who saw the label correctly recalled the creator for 92.8% of games. Consumers without labels could barely tell human work from AI work. They put the probability that a game was human-made at 59.3% for AI games and 58.2% for human games (p=0.67). The researchers summed up what this implies: arxiv

little hope for strategies that emphasize a human hand in production

The researchers also laid out the study's limits. Prices were in cents, play lasted about a minute, and making a game took minutes, not weeks. They wrote that it is unclear whether the results extend to markets where a few works dominate, like music or books, or to fields such as scientific research. They suggested that approaches other than labeling, such as creators and consumers getting to know each other, might bring out a preference for human work. On the other hand, they pointed out that deliberately weakening the AI, drawing consumers from the same pool as the creators, and labeling AI involvement prominently were all design choices that favored humans. As of October 26, no public rebuttal of the paper had appeared. arxiv

At the start of the experiment, creators expected games where a human set the AI's direction to earn more than games the AI made alone. In a market for one-minute games, people paid more for games from an AI that got a single line of instruction, and the gap barely moved when buyers were told who made each game. In markets like this, the bar human creators have to clear is no longer other people's games. It is a game an AI made on its own from one line of instruction.

Further reading

Creative destruction: Human-AI competition in an experimental market The full paper, with the design of all three studies, supplementary tables and screenshots of the experiment. arxiv.org

Games from Conlon and Schwardmann (2026) A public site collecting more than 1,600 games produced in the experiment. human-ai-games.github.io

The Clearing Price of Human Taste A secondary commentary that restates the paper's numbers through financial metaphors. Wording like “every time,” and the single 58% figure that merges the paper's two values (59.3% and 58.2%), differ from the original, so check the numbers against the paper. dev.to