The second week of the LEC Summer Split has arrived, bringing with it a new round of the humans versus robots experiment. Last week, Craig Robinson and three AI chatbots — Claude, ChatGPT, and Gemini — were tasked with building $10 accumulators based on the opening-weekend fixtures. Robinson's attempt didn't go as planned, as he admitted to being "greedy of spending money that was not real" and acknowledged that two of the AI models outperformed him. This week, Robinson is returning, promising to be more cautious, though one of his choices suggests his self-imposed restraint might be overstated.
According to Robinson, this ongoing comparison is less about the accuracy of the odds and more about testing human psychology. He humorously warns readers that by the end of the experiment, they may consider him "bananas." Whether that proves true remains to be seen, but it’s clear the stakes — fictional though they are — are high for the participants.
Week one: accumulator performance breakdown
The first week of the experiment delivered mixed results for the four participants. ChatGPT’s strategy of backing G2 in every leg of its treble backfired when Karmine Corp unexpectedly defeated G2 2-1. That upset was the only bot to lose a leg and not return a positive result. Craig’s more ambitious selection of KOI as an upset pick also failed to materialize, leaving him watching as the AI-generated predictions continued to deliver results.
By contrast, both Claude and Gemini used a more conservative approach, stacking their bets on heavy favorites. Their accumulators, which included matches with odds suggesting over a 90% win probability, returned positive results. Craig, in turn, adjusted his strategy, opting for three safer bets while still incorporating the week’s longest-odds pick.
Comparing betting strategies and risk profiles
The differences in betting strategies were clear. Claude took a four-leg approach, focusing entirely on matches where the favorites had extremely strong win probabilities. Karmine Corp appeared twice, as did Movistar KOI and G2, all of which were considered the strongest teams in their respective fixtures. The bot explained its logic as prioritizing a high hit-rate over a potentially large multiplier.
Gemini maintained a similar structure to the previous week, keeping its bets on teams with strong form and favorable odds. ChatGPT, however, made a complete pivot, shifting from an all-favorites accumulator to a single bet on Fnatic to defeat KOI at 6/4. This decision, while appearing straightforward, introduced a level of risk, as KOI had started the Summer Split with a 0-3 record, including a notable upset win over Vitality at long odds.
Fnatic, on the other hand, had a mixed record in the previous split, finishing eighth in Versus and seventh in Spring. Their recent roster changes further added to the uncertainty of the pick. Despite the 6/4 odds, the combination of KOI’s momentum and Fnatic’s instability made ChatGPT’s choice a more dangerous one than it first appeared.
Raw AI outputs and their unfiltered logic
For transparency, all the selections made by the chatbots and Craig were presented in their raw, unedited form. Claude’s final accumulator included four strong favorites: Movistar KOI to beat Shifters at 1/14, G2 to defeat SK Gaming at 1/8, Karmine Corp to beat Natus Vincere at 1/6, and Karmine Corp to beat Team Heretics at 1/20. The rationale was that by stacking matches with very high win probabilities, the bot could maximize the chance of success while minimizing the potential for a single upset to derail the entire bet.
Gemini’s strategy was consistent with its previous week’s approach, maintaining a similar betting structure and not making drastic changes. Its focus remained on discipline and avoiding high-risk choices. Meanwhile, ChatGPT’s decision to abandon the accumulator model entirely came after its first attempt — a multi-leg bet on all G2 matches — failed due to a single unexpected result.
The experiment continues to offer insight into how humans and machines think about sports betting. Craig Robinson’s human instincts, influenced by psychology and the lure of bigger payouts, have shown signs of being less disciplined than the AI models. However, as he has noted, the goal is not to determine who is objectively better at betting, but to explore how different mindsets — whether human or artificial — approach the same challenge.
With the first week complete, the second promises to reveal even more about the strengths and weaknesses of each participant.