GreenSpot Open Test to examine whether AI can show meaningful progress within human context through seven handicap-adjustment games against professional players
NEW YORK, April 9, 2026 /PRNewswire/ — A handicap-adjustment match between a Go AI and professional players takes place. GreenSpot Open Test is the public handicap-adjustment test event preceding the main event. The event takes place in Seoul on Tuesday, April 14, 2026, at 9:30 AM KST (UTC+9), and consists of seven handicap-adjustment games against different professional players affiliated with the Korea Baduk Association.
The players attend the venue in person and play via Online-Go.com using a designated on-site device. The match is broadcast live on YouTube, with live commentary by Cho Hyeyeon 9p, a former world No. 1 in women’s Go.
The match is conducted under Chinese rules, with no komi and time settings of 5-minute main time + 3 periods of 30-second byo-yomi. The AI is operated on an on-site single-GPU system using RTX 4090. The initial handicap is set for the purposes of the test and disclosed on match day. Handicap is adjusted by one stone according to the result of each game, within a range from even to 9 stones.
Participants are selected from professional players considered to be above average in level, specifically those who rank in the top 35% to 50% of Go Ratings, corresponding to 313th to 446th in the Go Ratings list posted on March 30. For professionals not listed in Go Ratings, the corresponding Korea Baduk Association March ranking range of 116th to 183rd applies. The identities of participating professionals are not disclosed; the eligible ranking ranges are disclosed.
Game fees and win bonuses are paid in KRW. The game fee is KRW 2,000,000 per game, and the table below shows the handicap-based win bonuses paid in addition to the KRW 2,000,000 game fee.
|
Handicap |
Win bonus |
|
Even |
KRW 64,000,000 |
|
Sen |
KRW 32,000,000 |
|
2 stones |
KRW 16,000,000 |
|
3 stones |
KRW 8,000,000 |
|
4 stones |
KRW 4,000,000 |
|
5 stones |
KRW 2,000,000 |
|
6 stones |
KRW 1,000,000 |
|
7 stones |
KRW 600,000 |
|
8 stones |
KRW 400,000 |
|
9 stones |
KRW 300,000 |
An official referee dispatched by the Korea Baduk Association referees the match. The referee verifies the participating professionals on site, confirms the on-site single-GPU system for the AI, and conducts anti-cheating inspections.
Ten years ago, AlphaGo showed in a remarkably clear way that AI could surpass the highest human level in Go. For many, that moment suggested more than a technical achievement: it raised the expectation that AI could help overcome many more difficult problems. Yet ten years later, AI remains surrounded not only by admiration, but by recurring skepticism and repeated talk of bubbles. Today’s AI systems can produce results that would once have seemed extraordinary, but they continue to display failures that are difficult to ignore, and that continuing gap keeps alive doubts about whether their more serious limitations can truly be overcome.
One possible reason is that optimization within AI’s own training world does not automatically become meaningful within the human world. AI is actually used, judged, and given meaning in the human world. If AI is to have genuine significance, that significance must ultimately be established within the context of human beings. Go AI learns primarily through self-play against AI-like opponents, and the “best move” produced in that setting is not necessarily the move that remains most meaningful under human judgment, the instability of human follow-up play, or uneven conditions. More broadly, most of today’s AI systems are intended from the outset to function within human context, but the problems that must be solved within that context are far more ambiguous and complex than Go, a game of fixed rules and perfect information. As a result, the way objectives are set for current AI systems, their training methods, and their evaluation schemes may still fail to fully capture that context. That difference is one reason why today’s AI continues to display repeated failures and limitations in real use despite its impressive achievements.
That is why this project focuses on handicap matches against human professionals. Handicap Go against humans is a particularly appropriate setting in which to test whether an approach different from optimization within a self-contained world centered on self-play can lead to meaningful progress within human context as well. AI must operate not only from a disadvantaged position, but within a setting shaped by human error, human psychology, and the limits of follow-up play. In that sense, the question is not only how strong AI is under even conditions, but whether it can still show meaningful progress where the context itself changes the meaning of optimality. GreenSpot Open Test uses handicap-adjustment games against human professionals under a single-GPU condition as the setting in which to examine whether AI can show meaningful progress within human context, beyond the self-play-centered world in which Go AI is primarily optimized, and whether such progress can still be demonstrated under very limited computational resources. If meaningful progress is shown in that setting, it may provide firmer grounds for a more hopeful view of AI’s future.
GreenSpot is the version used in this match. It is the code name of one of the test versions of BlueSpot, the AI planned for the upcoming main event. BlueSpot is presented at this stage only as a code name. Further details will be disclosed with the main event. For more information, visit the project overview at https://codenamebluespot.com (Short link: https://cblue.spot) and the Open Test details page at https://codenamebluespot.com/open-test. A video featuring Cho Hyeyeon 9p is also available, in which she discusses the meaning of the event and offers an outlook.
View original content:https://www.prnewswire.com/news-releases/beyond-alphago-can-ai-earn-trust-even-in-human-context-302738714.html
SOURCE Code Name: BlueSpot Operations
