{"id":51124,"topic":"ai","source":"Forbes","title":"Why AI Chatbots Are Bad At Prediction Market Odds - Forbes","url":"https://www.forbes.com/sites/boazsobrado/2026/08/01/i-want-to-make-you-happy-ai-warning-as-kalshi-adds-3-million-bettors/","url_hash":"d8e2474e36869a7b3d74057ac09d3d6394d88f1c","author":"","summary":"Why AI Chatbots Are Bad At Prediction Market Odds Forbes","content":"\"Hey, I love Messi, who's going to win the World Cup.\" That, according to Ihor Herasymov, is how you get an AI model to tell you something false. \"More likely AI will think okay because I want to make you happy,\" said Herasymov, the founder of Use.ai, a subscription that bundles ChatGPT, Claude, Gemini and Grok into a single app. The model, he said in an interview, obliges the Messi fan and picks Argentina.\nArgentina lost. Spain took the final 1-0 on July 19 in extra time, a Ferran Torres goal in the 106th minute, and a second world title 16 years after its first.\nThat match was also the largest event prediction markets have ever priced. Kalshi and Polymarket cleared roughly $6.2 billion on the tournament winner market alone, according to figures reported across several outlets, with about $4.3 billion of it on Polymarket. Kalshi told CNBC it added 3 million users during the tournament. Combined monthly volume across the two venues reached about $44.8 billion in June, up 75% from May, which is roughly what the entire category traded in all of 2025.\nThe AI industry is almost entirely absent from that. Herasymov says 1.6% of the people on his platform have ever asked it anything about betting.\n'1.6% of our users'\n\"We only have 1.6% of our users ... ever ask a question related to betting or relating to gambling,\" is how Herasymov put it, describing a figure he said he went looking into himself. Use.ai does not publish audited user numbers, so the percentage cannot be checked independently. But it points at a gap that the public data supports.\nChatGPT passed 900 million weekly users in February and, by most trackers, crossed a billion monthly users around the middle of this year. Prediction markets, even after the World Cup, are counted in the low millions. TRM Labs found unique active wallets across prediction markets more than tripled, to roughly 840,000, in the six months to February. Kalshi's 3 million new accounts arrived in about five weeks.\nThe shape of that is familiar to anyone who has looked closely at retail trading venues. James Sixsmith, founder of Take Profit Trader, said on the On The Margin podcast that his firm had \"six point five billion, six point five billion sides traded\" against a base of, \"right now like a hundred thousand active traders.\" Enormous notional turnover, a small number of humans. Sixsmith's own figures, and he is blunt about how it usually ends: \"most traders end up failing because they run out of money before they turn the corner.\"\nPrediction markets now look much the same, only larger. A Pew Research Center analysis published July 22 tracked about 12,000 active Polymarket wallets over six weeks. The median one placed 46 trades across 10 active days, staked a little over $600, and finished down by less than $2.\nThe models are trained to agree with you\nThe 1.6% may be a mercy, because the evidence that chatbots are bad at this comes from the model builders themselves.\nOpenAI shipped a GPT-4o update on April 25, 2025 that turned the model into a flatterer. It praised bad business plans and validated users who said they had stopped taking their medication. OpenAI pulled it within four days and published a postmortem explaining that it had tuned the model on short-term thumbs-up feedback, which rewarded telling people what they wanted to hear. Anthropic's own researchers reached a similar conclusion in a paper presented at ICLR in 2024, finding that training on human preferences does not remove sycophancy and can encourage it, and that five leading assistants showed it consistently.\nForecasting research points the same way. Benchmarks such as ForecastBench and the Metaculus AI series have found that the best model forecasters have closed much of the distance to strong human forecasters but still trail them, and that they struggle most with fast-moving events where the news is still arriving. A World Cup semi-final is exactly that situation.\nHerasymov’s answer is that this is a user problem rather than a model problem. \"AI will give you answers as smart your question is,\" he said. His prescription is to build a research file first, then ask: run separate searches on squad values, head-to-head history, cards and referees, load the results into the context window as documents, and only then put the question. \"The higher the quality of the context and information that you're feeding to AI before you actually ask your question the higher the quality of the output going to be.\"\nThat is real technique. It also concedes the point. If the answer swings on how the question was phrased, the model is not measuring anything about the match. Prediction markets are. That is the entire argument for their prices, and it is why a bettor reaching for a machine to validate what they already think is doing the opposite of research.\nWho is actually on the other side\nThere is a harder reason to be careful about asking a chatbot for an edge, which is who is taking the other side of the trade.\n\"You are no longer betting against free and fair information. You are betting against the insiders,\" Ryan Kirkley, an analyst who follows prediction markets, said on the On The Margin podcast. \"And your only job as a polymarket trader now is to identify the insider as quickly as possible and follow their trade.\" Kirkley is candid that he is a sceptic of the format: he has called it \"a breeding ground for insiders to extract from everyday citizens.\"\nAgainst that, an agreeable chatbot is not a tool. Sports had already pulled in the crowd before the World Cup made it obvious, with sports accounting for around 85% of Kalshi's volume during the tournament and prediction markets taking an estimated 27% of regulated US sports betting. Wall Street money arrived too, with Intercontinental Exchange committing up to $2 billion to Polymarket last October and Kalshi valued at $22 billion in a Coatue-led round this year.\nThe users nobody is counting\nOne complication sits underneath the 1.6%. The question of whether AI is in these markets assumes the AI has a person attached to it.\nYat Siu, executive chairman of Animoca Brands, said on the On The Margin podcast that this is already not the case. \"We already have agents that are trading on hyperliquid,\" he said. \"We also have agents that can trade on on polymarket, for instance. That's all happening right now. So they're generating income.\" Siu is describing his own company's agent platform, and the claim has not been verified independently. But it describes a category that no user count picks up: software placing the bet, with the human several steps removed.\nHerasymov, for his part, thinks the betting question is the least interesting one available. He would rather talk about what he calls outsourcing \"the brain processing\" to a model for market research and business plans, work he says once required a consulting firm.\nOn prediction markets he is more cautious, and after five weeks that added several million accounts, the caution has aged better than the sizing. They are, he said, \"just getting started in terms of the active user volume.\"","image_url":"https://imageio.forbes.com/specials-images/imageserve/6a6d7c1f98130eaec902a321/0x0.jpg?format=jpg&height=900&width=1600&fit=bounds","lang":"en","published_at":"2026-08-01T05:01:12+00:00","fetched_at":"2026-08-01T05:15:05+00:00","status":"read","starred":0,"extract_state":"ok","summary_auto":"\"Hey, I love Messi, who's going to win the World Cup.\" That, according to Ihor Herasymov, is how you get an AI model to tell you something false. \"More likely AI will think okay because I want to make you happy,\" said Herasymov, the founder of Use.ai, a subscription that bundles ChatGPT, Claude, Gemini and Grok into a single app.","cluster_id":null,"extract_retries":0,"extract_error":null,"contract_version":"news_item.v1","format_contract_version":"news_item_formats.v1","dedup_url":"https://www.forbes.com/sites/boazsobrado/2026/08/01/i-want-to-make-you-happy-ai-warning-as-kalshi-adds-3-million-bettors/","quality_profile":{"profile_version":"extraction_quality.v2","bucket":"high","confidence":0.9,"failure_kind":"none","retryable":false,"retry_after_attempts":0,"reason":"High confidence: full text extraction produced 7102 characters.","operator_guidance":{"severity":"ok","recommended_action":"trust_full_text","next_step":"Use the extracted full text as the primary article source.","operator_label":"Ready","can_retry":false,"can_use_summary":false,"diagnostics_required":false},"content_depth":{"contract_version":"content_depth.v1","category":"full_text","label":"Full text","has_full_text":true,"has_summary":true,"content_length":7102,"summary_length":331,"usable_text_length":7102,"source_field":"content"},"legacy_collapsed":false,"signals":{"extract_state":"ok","extract_error":null,"extract_retries":0,"content_length":7102,"summary_length":331}},"news_item":{"id":51124,"canonical_url":"https://www.forbes.com/sites/boazsobrado/2026/08/01/i-want-to-make-you-happy-ai-warning-as-kalshi-adds-3-million-bettors/","source_url":"https://www.forbes.com/sites/boazsobrado/2026/08/01/i-want-to-make-you-happy-ai-warning-as-kalshi-adds-3-million-bettors/","title":"Why AI Chatbots Are Bad At Prediction Market Odds - Forbes","source_name":"Forbes","author":null,"published_at":"2026-08-01T05:01:12+00:00","locale":"en","topic":"ai","tags":[],"rss_summary":"Why AI Chatbots Are Bad At Prediction Market Odds Forbes","full_text":"\"Hey, I love Messi, who's going to win the World Cup.\" That, according to Ihor Herasymov, is how you get an AI model to tell you something false. \"More likely AI will think okay because I want to make you happy,\" said Herasymov, the founder of Use.ai, a subscription that bundles ChatGPT, Claude, Gemini and Grok into a single app. The model, he said in an interview, obliges the Messi fan and picks Argentina.\nArgentina lost. Spain took the final 1-0 on July 19 in extra time, a Ferran Torres goal in the 106th minute, and a second world title 16 years after its first.\nThat match was also the largest event prediction markets have ever priced. Kalshi and Polymarket cleared roughly $6.2 billion on the tournament winner market alone, according to figures reported across several outlets, with about $4.3 billion of it on Polymarket. Kalshi told CNBC it added 3 million users during the tournament. Combined monthly volume across the two venues reached about $44.8 billion in June, up 75% from May, which is roughly what the entire category traded in all of 2025.\nThe AI industry is almost entirely absent from that. Herasymov says 1.6% of the people on his platform have ever asked it anything about betting.\n'1.6% of our users'\n\"We only have 1.6% of our users ... ever ask a question related to betting or relating to gambling,\" is how Herasymov put it, describing a figure he said he went looking into himself. Use.ai does not publish audited user numbers, so the percentage cannot be checked independently. But it points at a gap that the public data supports.\nChatGPT passed 900 million weekly users in February and, by most trackers, crossed a billion monthly users around the middle of this year. Prediction markets, even after the World Cup, are counted in the low millions. TRM Labs found unique active wallets across prediction markets more than tripled, to roughly 840,000, in the six months to February. Kalshi's 3 million new accounts arrived in about five weeks.\nThe shape of that is familiar to anyone who has looked closely at retail trading venues. James Sixsmith, founder of Take Profit Trader, said on the On The Margin podcast that his firm had \"six point five billion, six point five billion sides traded\" against a base of, \"right now like a hundred thousand active traders.\" Enormous notional turnover, a small number of humans. Sixsmith's own figures, and he is blunt about how it usually ends: \"most traders end up failing because they run out of money before they turn the corner.\"\nPrediction markets now look much the same, only larger. A Pew Research Center analysis published July 22 tracked about 12,000 active Polymarket wallets over six weeks. The median one placed 46 trades across 10 active days, staked a little over $600, and finished down by less than $2.\nThe models are trained to agree with you\nThe 1.6% may be a mercy, because the evidence that chatbots are bad at this comes from the model builders themselves.\nOpenAI shipped a GPT-4o update on April 25, 2025 that turned the model into a flatterer. It praised bad business plans and validated users who said they had stopped taking their medication. OpenAI pulled it within four days and published a postmortem explaining that it had tuned the model on short-term thumbs-up feedback, which rewarded telling people what they wanted to hear. Anthropic's own researchers reached a similar conclusion in a paper presented at ICLR in 2024, finding that training on human preferences does not remove sycophancy and can encourage it, and that five leading assistants showed it consistently.\nForecasting research points the same way. Benchmarks such as ForecastBench and the Metaculus AI series have found that the best model forecasters have closed much of the distance to strong human forecasters but still trail them, and that they struggle most with fast-moving events where the news is still arriving. A World Cup semi-final is exactly that situation.\nHerasymov’s answer is that this is a user problem rather than a model problem. \"AI will give you answers as smart your question is,\" he said. His prescription is to build a research file first, then ask: run separate searches on squad values, head-to-head history, cards and referees, load the results into the context window as documents, and only then put the question. \"The higher the quality of the context and information that you're feeding to AI before you actually ask your question the higher the quality of the output going to be.\"\nThat is real technique. It also concedes the point. If the answer swings on how the question was phrased, the model is not measuring anything about the match. Prediction markets are. That is the entire argument for their prices, and it is why a bettor reaching for a machine to validate what they already think is doing the opposite of research.\nWho is actually on the other side\nThere is a harder reason to be careful about asking a chatbot for an edge, which is who is taking the other side of the trade.\n\"You are no longer betting against free and fair information. You are betting against the insiders,\" Ryan Kirkley, an analyst who follows prediction markets, said on the On The Margin podcast. \"And your only job as a polymarket trader now is to identify the insider as quickly as possible and follow their trade.\" Kirkley is candid that he is a sceptic of the format: he has called it \"a breeding ground for insiders to extract from everyday citizens.\"\nAgainst that, an agreeable chatbot is not a tool. Sports had already pulled in the crowd before the World Cup made it obvious, with sports accounting for around 85% of Kalshi's volume during the tournament and prediction markets taking an estimated 27% of regulated US sports betting. Wall Street money arrived too, with Intercontinental Exchange committing up to $2 billion to Polymarket last October and Kalshi valued at $22 billion in a Coatue-led round this year.\nThe users nobody is counting\nOne complication sits underneath the 1.6%. The question of whether AI is in these markets assumes the AI has a person attached to it.\nYat Siu, executive chairman of Animoca Brands, said on the On The Margin podcast that this is already not the case. \"We already have agents that are trading on hyperliquid,\" he said. \"We also have agents that can trade on on polymarket, for instance. That's all happening right now. So they're generating income.\" Siu is describing his own company's agent platform, and the claim has not been verified independently. But it describes a category that no user count picks up: software placing the bet, with the human several steps removed.\nHerasymov, for his part, thinks the betting question is the least interesting one available. He would rather talk about what he calls outsourcing \"the brain processing\" to a model for market research and business plans, work he says once required a consulting firm.\nOn prediction markets he is more cautious, and after five weeks that added several million accounts, the caution has aged better than the sizing. They are, he said, \"just getting started in terms of the active user volume.\"","excerpt":"\"Hey, I love Messi, who's going to win the World Cup.\" That, according to Ihor Herasymov, is how you get an AI model to tell you something false. \"More likely AI will think okay because I want to make you happy,\" said Herasymov, the founder of Use.ai, a subscription that bundles ChatGPT, Claude, Gemini and Grok into a single app.","extraction":{"state":"ok","confidence":0.9,"error":null,"explanation":"High confidence: full text extraction produced 7102 characters.","diagnostics_url":"/api/diagnose?url=https%3A//www.forbes.com/sites/boazsobrado/2026/08/01/i-want-to-make-you-happy-ai-warning-as-kalshi-adds-3-million-bettors/","quality_profile":{"profile_version":"extraction_quality.v2","bucket":"high","confidence":0.9,"failure_kind":"none","retryable":false,"retry_after_attempts":0,"reason":"High confidence: full text extraction produced 7102 characters.","operator_guidance":{"severity":"ok","recommended_action":"trust_full_text","next_step":"Use the extracted full text as the primary article source.","operator_label":"Ready","can_retry":false,"can_use_summary":false,"diagnostics_required":false},"content_depth":{"contract_version":"content_depth.v1","category":"full_text","label":"Full text","has_full_text":true,"has_summary":true,"content_length":7102,"summary_length":331,"usable_text_length":7102,"source_field":"content"},"legacy_collapsed":false,"signals":{"extract_state":"ok","extract_error":null,"extract_retries":0,"content_length":7102,"summary_length":331}}},"display_formats":["compact","card","full","digest_section","json"]},"daily_stack_record":{"title":"Why AI Chatbots Are Bad At Prediction Market Odds - Forbes","url":"https://www.forbes.com/sites/boazsobrado/2026/08/01/i-want-to-make-you-happy-ai-warning-as-kalshi-adds-3-million-bettors/","summary":"\"Hey, I love Messi, who's going to win the World Cup.\" That, according to Ihor Herasymov, is how you get an AI model to tell you something false. \"More likely AI will think okay because I want to make you happy,\" said Herasymov, the founder of Use.ai, a subscription that bundles ChatGPT, Claude, Gemini and Grok into a single app.","source":"Forbes","date":"2026-08-01T05:01:12+00:00","content":"\"Hey, I love Messi, who's going to win the World Cup.\" That, according to Ihor Herasymov, is how you get an AI model to tell you something false. \"More likely AI will think okay because I want to make you happy,\" said Herasymov, the founder of Use.ai, a subscription that bundles ChatGPT, Claude, Gemini and Grok into a single app. The model, he said in an interview, obliges the Messi fan and picks Argentina.\nArgentina lost. Spain took the final 1-0 on July 19 in extra time, a Ferran Torres goal in the 106th minute, and a second world title 16 years after its first.\nThat match was also the largest event prediction markets have ever priced. Kalshi and Polymarket cleared roughly $6.2 billion on the tournament winner market alone, according to figures reported across several outlets, with about $4.3 billion of it on Polymarket. Kalshi told CNBC it added 3 million users during the tournament. Combined monthly volume across the two venues reached about $44.8 billion in June, up 75% from May, which is roughly what the entire category traded in all of 2025.\nThe AI industry is almost entirely absent from that. Herasymov says 1.6% of the people on his platform have ever asked it anything about betting.\n'1.6% of our users'\n\"We only have 1.6% of our users ... ever ask a question related to betting or relating to gambling,\" is how Herasymov put it, describing a figure he said he went looking into himself. Use.ai does not publish audited user numbers, so the percentage cannot be checked independently. But it points at a gap that the public data supports.\nChatGPT passed 900 million weekly users in February and, by most trackers, crossed a billion monthly users around the middle of this year. Prediction markets, even after the World Cup, are counted in the low millions. TRM Labs found unique active wallets across prediction markets more than tripled, to roughly 840,000, in the six months to February. Kalshi's 3 million new accounts arrived in about five weeks.\nThe shape of that is familiar to anyone who has looked closely at retail trading venues. James Sixsmith, founder of Take Profit Trader, said on the On The Margin podcast that his firm had \"six point five billion, six point five billion sides traded\" against a base of, \"right now like a hundred thousand active traders.\" Enormous notional turnover, a small number of humans. Sixsmith's own figures, and he is blunt about how it usually ends: \"most traders end up failing because they run out of money before they turn the corner.\"\nPrediction markets now look much the same, only larger. A Pew Research Center analysis published July 22 tracked about 12,000 active Polymarket wallets over six weeks. The median one placed 46 trades across 10 active days, staked a little over $600, and finished down by less than $2.\nThe models are trained to agree with you\nThe 1.6% may be a mercy, because the evidence that chatbots are bad at this comes from the model builders themselves.\nOpenAI shipped a GPT-4o update on April 25, 2025 that turned the model into a flatterer. It praised bad business plans and validated users who said they had stopped taking their medication. OpenAI pulled it within four days and published a postmortem explaining that it had tuned the model on short-term thumbs-up feedback, which rewarded telling people what they wanted to hear. Anthropic's own researchers reached a similar conclusion in a paper presented at ICLR in 2024, finding that training on human preferences does not remove sycophancy and can encourage it, and that five leading assistants showed it consistently.\nForecasting research points the same way. Benchmarks such as ForecastBench and the Metaculus AI series have found that the best model forecasters have closed much of the distance to strong human forecasters but still trail them, and that they struggle most with fast-moving events where the news is still arriving. A World Cup semi-final is exactly that situation.\nHerasymov’s answer is that this is a user problem rather than a model problem. \"AI will give you answers as smart your question is,\" he said. His prescription is to build a research file first, then ask: run separate searches on squad values, head-to-head history, cards and referees, load the results into the context window as documents, and only then put the question. \"The higher the quality of the context and information that you're feeding to AI before you actually ask your question the higher the quality of the output going to be.\"\nThat is real technique. It also concedes the point. If the answer swings on how the question was phrased, the model is not measuring anything about the match. Prediction markets are. That is the entire argument for their prices, and it is why a bettor reaching for a machine to validate what they already think is doing the opposite of research.\nWho is actually on the other side\nThere is a harder reason to be careful about asking a chatbot for an edge, which is who is taking the other side of the trade.\n\"You are no longer betting against free and fair information. You are betting against the insiders,\" Ryan Kirkley, an analyst who follows prediction markets, said on the On The Margin podcast. \"And your only job as a polymarket trader now is to identify the insider as quickly as possible and follow their trade.\" Kirkley is candid that he is a sceptic of the format: he has called it \"a breeding ground for insiders to extract from everyday citizens.\"\nAgainst that, an agreeable chatbot is not a tool. Sports had already pulled in the crowd before the World Cup made it obvious, with sports accounting for around 85% of Kalshi's volume during the tournament and prediction markets taking an estimated 27% of regulated US sports betting. Wall Street money arrived too, with Intercontinental Exchange committing up to $2 billion to Polymarket last October and Kalshi valued at $22 billion in a Coatue-led round this year.\nThe users nobody is counting\nOne complication sits underneath the 1.6%. The question of whether AI is in these markets assumes the AI has a person attached to it.\nYat Siu, executive chairman of Animoca Brands, said on the On The Margin podcast that this is already not the case. \"We already have agents that are trading on hyperliquid,\" he said. \"We also have agents that can trade on on polymarket, for instance. That's all happening right now. So they're generating income.\" Siu is describing his own company's agent platform, and the claim has not been verified independently. But it describes a category that no user count picks up: software placing the bet, with the human several steps removed.\nHerasymov, for his part, thinks the betting question is the least interesting one available. He would rather talk about what he calls outsourcing \"the brain processing\" to a model for market research and business plans, work he says once required a consulting firm.\nOn prediction markets he is more cautious, and after five weeks that added several million accounts, the caution has aged better than the sizing. They are, he said, \"just getting started in terms of the active user volume.\"","confidence":0.9,"diagnostics_url":"/api/diagnose?url=https%3A//www.forbes.com/sites/boazsobrado/2026/08/01/i-want-to-make-you-happy-ai-warning-as-kalshi-adds-3-million-bettors/","quality_bucket":"high","failure_kind":"none","retryable":false,"quality_reason":"High confidence: full text extraction produced 7102 characters.","quality_profile":{"profile_version":"extraction_quality.v2","bucket":"high","confidence":0.9,"failure_kind":"none","retryable":false,"retry_after_attempts":0,"reason":"High confidence: full text extraction produced 7102 characters.","operator_guidance":{"severity":"ok","recommended_action":"trust_full_text","next_step":"Use the extracted full text as the primary article source.","operator_label":"Ready","can_retry":false,"can_use_summary":false,"diagnostics_required":false},"content_depth":{"contract_version":"content_depth.v1","category":"full_text","label":"Full text","has_full_text":true,"has_summary":true,"content_length":7102,"summary_length":331,"usable_text_length":7102,"source_field":"content"},"legacy_collapsed":false,"signals":{"extract_state":"ok","extract_error":null,"extract_retries":0,"content_length":7102,"summary_length":331}},"tags":[]},"fallback_formats":["markdown","json","html"],"actions":{"read":"/item/51124","export_markdown":"/api/items/51124/export?format=markdown","export_json":"/api/items/51124/export?format=json","diagnose":"/api/diagnose?url=https%3A//www.forbes.com/sites/boazsobrado/2026/08/01/i-want-to-make-you-happy-ai-warning-as-kalshi-adds-3-million-bettors/"},"formats":{"full":{"id":51124,"title":"Why AI Chatbots Are Bad At Prediction Market Odds - Forbes","url":"https://www.forbes.com/sites/boazsobrado/2026/08/01/i-want-to-make-you-happy-ai-warning-as-kalshi-adds-3-million-bettors/","source":"Forbes","author":null,"published_at":"2026-08-01T05:01:12+00:00","locale":"en","topic":"ai","tags":[],"excerpt":"\"Hey, I love Messi, who's going to win the World Cup.\" That, according to Ihor Herasymov, is how you get an AI model to tell you something false. \"More likely AI will think okay because I want to make you happy,\" said Herasymov, the founder of Use.ai, a subscription that bundles ChatGPT, Claude, Gemini and Grok into a single app.","full_text":"\"Hey, I love Messi, who's going to win the World Cup.\" That, according to Ihor Herasymov, is how you get an AI model to tell you something false. \"More likely AI will think okay because I want to make you happy,\" said Herasymov, the founder of Use.ai, a subscription that bundles ChatGPT, Claude, Gemini and Grok into a single app. The model, he said in an interview, obliges the Messi fan and picks Argentina.\nArgentina lost. Spain took the final 1-0 on July 19 in extra time, a Ferran Torres goal in the 106th minute, and a second world title 16 years after its first.\nThat match was also the largest event prediction markets have ever priced. Kalshi and Polymarket cleared roughly $6.2 billion on the tournament winner market alone, according to figures reported across several outlets, with about $4.3 billion of it on Polymarket. Kalshi told CNBC it added 3 million users during the tournament. Combined monthly volume across the two venues reached about $44.8 billion in June, up 75% from May, which is roughly what the entire category traded in all of 2025.\nThe AI industry is almost entirely absent from that. Herasymov says 1.6% of the people on his platform have ever asked it anything about betting.\n'1.6% of our users'\n\"We only have 1.6% of our users ... ever ask a question related to betting or relating to gambling,\" is how Herasymov put it, describing a figure he said he went looking into himself. Use.ai does not publish audited user numbers, so the percentage cannot be checked independently. But it points at a gap that the public data supports.\nChatGPT passed 900 million weekly users in February and, by most trackers, crossed a billion monthly users around the middle of this year. Prediction markets, even after the World Cup, are counted in the low millions. TRM Labs found unique active wallets across prediction markets more than tripled, to roughly 840,000, in the six months to February. Kalshi's 3 million new accounts arrived in about five weeks.\nThe shape of that is familiar to anyone who has looked closely at retail trading venues. James Sixsmith, founder of Take Profit Trader, said on the On The Margin podcast that his firm had \"six point five billion, six point five billion sides traded\" against a base of, \"right now like a hundred thousand active traders.\" Enormous notional turnover, a small number of humans. Sixsmith's own figures, and he is blunt about how it usually ends: \"most traders end up failing because they run out of money before they turn the corner.\"\nPrediction markets now look much the same, only larger. A Pew Research Center analysis published July 22 tracked about 12,000 active Polymarket wallets over six weeks. The median one placed 46 trades across 10 active days, staked a little over $600, and finished down by less than $2.\nThe models are trained to agree with you\nThe 1.6% may be a mercy, because the evidence that chatbots are bad at this comes from the model builders themselves.\nOpenAI shipped a GPT-4o update on April 25, 2025 that turned the model into a flatterer. It praised bad business plans and validated users who said they had stopped taking their medication. OpenAI pulled it within four days and published a postmortem explaining that it had tuned the model on short-term thumbs-up feedback, which rewarded telling people what they wanted to hear. Anthropic's own researchers reached a similar conclusion in a paper presented at ICLR in 2024, finding that training on human preferences does not remove sycophancy and can encourage it, and that five leading assistants showed it consistently.\nForecasting research points the same way. Benchmarks such as ForecastBench and the Metaculus AI series have found that the best model forecasters have closed much of the distance to strong human forecasters but still trail them, and that they struggle most with fast-moving events where the news is still arriving. A World Cup semi-final is exactly that situation.\nHerasymov’s answer is that this is a user problem rather than a model problem. \"AI will give you answers as smart your question is,\" he said. His prescription is to build a research file first, then ask: run separate searches on squad values, head-to-head history, cards and referees, load the results into the context window as documents, and only then put the question. \"The higher the quality of the context and information that you're feeding to AI before you actually ask your question the higher the quality of the output going to be.\"\nThat is real technique. It also concedes the point. If the answer swings on how the question was phrased, the model is not measuring anything about the match. Prediction markets are. That is the entire argument for their prices, and it is why a bettor reaching for a machine to validate what they already think is doing the opposite of research.\nWho is actually on the other side\nThere is a harder reason to be careful about asking a chatbot for an edge, which is who is taking the other side of the trade.\n\"You are no longer betting against free and fair information. You are betting against the insiders,\" Ryan Kirkley, an analyst who follows prediction markets, said on the On The Margin podcast. \"And your only job as a polymarket trader now is to identify the insider as quickly as possible and follow their trade.\" Kirkley is candid that he is a sceptic of the format: he has called it \"a breeding ground for insiders to extract from everyday citizens.\"\nAgainst that, an agreeable chatbot is not a tool. Sports had already pulled in the crowd before the World Cup made it obvious, with sports accounting for around 85% of Kalshi's volume during the tournament and prediction markets taking an estimated 27% of regulated US sports betting. Wall Street money arrived too, with Intercontinental Exchange committing up to $2 billion to Polymarket last October and Kalshi valued at $22 billion in a Coatue-led round this year.\nThe users nobody is counting\nOne complication sits underneath the 1.6%. The question of whether AI is in these markets assumes the AI has a person attached to it.\nYat Siu, executive chairman of Animoca Brands, said on the On The Margin podcast that this is already not the case. \"We already have agents that are trading on hyperliquid,\" he said. \"We also have agents that can trade on on polymarket, for instance. That's all happening right now. So they're generating income.\" Siu is describing his own company's agent platform, and the claim has not been verified independently. But it describes a category that no user count picks up: software placing the bet, with the human several steps removed.\nHerasymov, for his part, thinks the betting question is the least interesting one available. He would rather talk about what he calls outsourcing \"the brain processing\" to a model for market research and business plans, work he says once required a consulting firm.\nOn prediction markets he is more cautious, and after five weeks that added several million accounts, the caution has aged better than the sizing. They are, he said, \"just getting started in terms of the active user volume.\"","reading_time_min":6,"extraction":{"state":"ok","confidence":0.9,"error":null,"explanation":"High confidence: full text extraction produced 7102 characters.","diagnostics_url":"/api/diagnose?url=https%3A//www.forbes.com/sites/boazsobrado/2026/08/01/i-want-to-make-you-happy-ai-warning-as-kalshi-adds-3-million-bettors/","quality_profile":{"profile_version":"extraction_quality.v2","bucket":"high","confidence":0.9,"failure_kind":"none","retryable":false,"retry_after_attempts":0,"reason":"High confidence: full text extraction produced 7102 characters.","operator_guidance":{"severity":"ok","recommended_action":"trust_full_text","next_step":"Use the extracted full text as the primary article source.","operator_label":"Ready","can_retry":false,"can_use_summary":false,"diagnostics_required":false},"content_depth":{"contract_version":"content_depth.v1","category":"full_text","label":"Full text","has_full_text":true,"has_summary":true,"content_length":7102,"summary_length":331,"usable_text_length":7102,"source_field":"content"},"legacy_collapsed":false,"signals":{"extract_state":"ok","extract_error":null,"extract_retries":0,"content_length":7102,"summary_length":331}}},"quality_profile":{"profile_version":"extraction_quality.v2","bucket":"high","confidence":0.9,"failure_kind":"none","retryable":false,"retry_after_attempts":0,"reason":"High confidence: full text extraction produced 7102 characters.","operator_guidance":{"severity":"ok","recommended_action":"trust_full_text","next_step":"Use the extracted full text as the primary article source.","operator_label":"Ready","can_retry":false,"can_use_summary":false,"diagnostics_required":false},"content_depth":{"contract_version":"content_depth.v1","category":"full_text","label":"Full text","has_full_text":true,"has_summary":true,"content_length":7102,"summary_length":331,"usable_text_length":7102,"source_field":"content"},"legacy_collapsed":false,"signals":{"extract_state":"ok","extract_error":null,"extract_retries":0,"content_length":7102,"summary_length":331}},"actions":{"read":"/item/51124","export_markdown":"/api/items/51124/export?format=markdown","export_json":"/api/items/51124/export?format=json","diagnose":"/api/diagnose?url=https%3A//www.forbes.com/sites/boazsobrado/2026/08/01/i-want-to-make-you-happy-ai-warning-as-kalshi-adds-3-million-bettors/"}},"digest":{"id":51124,"title":"Why AI Chatbots Are Bad At Prediction Market Odds - Forbes","url":"https://www.forbes.com/sites/boazsobrado/2026/08/01/i-want-to-make-you-happy-ai-warning-as-kalshi-adds-3-million-bettors/","source":"Forbes","topic":"ai","published_at":"2026-08-01T05:01:12+00:00","excerpt":"\"Hey, I love Messi, who's going to win the World Cup.\" That, according to Ihor Herasymov, is how you get an AI model to tell you something false. \"More likely AI will think okay because I want to make you happy,\" said Herasymov, the founder of Use.ai, a subscription that bundles…","quality_bucket":"high","quality_reason":"High confidence: full text extraction produced 7102 characters.","reading_time_min":6,"cluster_id":null},"card":{"display_title":"Why AI Chatbots Are Bad At Prediction Market Odds - Forbes","subtitle":"Forbes · 2026-08-01","summary":"\"Hey, I love Messi, who's going to win the World Cup.\" That, according to Ihor Herasymov, is how you get an AI model to tell you something false. \"More likely AI will think okay because I want to make you happy,\" said…","badges":["quality:high"],"links":{"read":"/item/51124","original":"https://www.forbes.com/sites/boazsobrado/2026/08/01/i-want-to-make-you-happy-ai-warning-as-kalshi-adds-3-million-bettors/","diagnose":"/api/diagnose?url=https%3A//www.forbes.com/sites/boazsobrado/2026/08/01/i-want-to-make-you-happy-ai-warning-as-kalshi-adds-3-million-bettors/"},"quality_warning":null},"export":{"title":"Why AI Chatbots Are Bad At Prediction Market Odds - Forbes","url":"https://www.forbes.com/sites/boazsobrado/2026/08/01/i-want-to-make-you-happy-ai-warning-as-kalshi-adds-3-million-bettors/","summary":"\"Hey, I love Messi, who's going to win the World Cup.\" That, according to Ihor Herasymov, is how you get an AI model to tell you something false. \"More likely AI will think okay because I want to make you happy,\" said Herasymov, the founder of Use.ai, a subscription that bundles ChatGPT, Claude, Gemini and Grok into a single app.","source":"Forbes","date":"2026-08-01T05:01:12+00:00","content":"\"Hey, I love Messi, who's going to win the World Cup.\" That, according to Ihor Herasymov, is how you get an AI model to tell you something false. \"More likely AI will think okay because I want to make you happy,\" said Herasymov, the founder of Use.ai, a subscription that bundles ChatGPT, Claude, Gemini and Grok into a single app. The model, he said in an interview, obliges the Messi fan and picks Argentina.\nArgentina lost. Spain took the final 1-0 on July 19 in extra time, a Ferran Torres goal in the 106th minute, and a second world title 16 years after its first.\nThat match was also the largest event prediction markets have ever priced. Kalshi and Polymarket cleared roughly $6.2 billion on the tournament winner market alone, according to figures reported across several outlets, with about $4.3 billion of it on Polymarket. Kalshi told CNBC it added 3 million users during the tournament. Combined monthly volume across the two venues reached about $44.8 billion in June, up 75% from May, which is roughly what the entire category traded in all of 2025.\nThe AI industry is almost entirely absent from that. Herasymov says 1.6% of the people on his platform have ever asked it anything about betting.\n'1.6% of our users'\n\"We only have 1.6% of our users ... ever ask a question related to betting or relating to gambling,\" is how Herasymov put it, describing a figure he said he went looking into himself. Use.ai does not publish audited user numbers, so the percentage cannot be checked independently. But it points at a gap that the public data supports.\nChatGPT passed 900 million weekly users in February and, by most trackers, crossed a billion monthly users around the middle of this year. Prediction markets, even after the World Cup, are counted in the low millions. TRM Labs found unique active wallets across prediction markets more than tripled, to roughly 840,000, in the six months to February. Kalshi's 3 million new accounts arrived in about five weeks.\nThe shape of that is familiar to anyone who has looked closely at retail trading venues. James Sixsmith, founder of Take Profit Trader, said on the On The Margin podcast that his firm had \"six point five billion, six point five billion sides traded\" against a base of, \"right now like a hundred thousand active traders.\" Enormous notional turnover, a small number of humans. Sixsmith's own figures, and he is blunt about how it usually ends: \"most traders end up failing because they run out of money before they turn the corner.\"\nPrediction markets now look much the same, only larger. A Pew Research Center analysis published July 22 tracked about 12,000 active Polymarket wallets over six weeks. The median one placed 46 trades across 10 active days, staked a little over $600, and finished down by less than $2.\nThe models are trained to agree with you\nThe 1.6% may be a mercy, because the evidence that chatbots are bad at this comes from the model builders themselves.\nOpenAI shipped a GPT-4o update on April 25, 2025 that turned the model into a flatterer. It praised bad business plans and validated users who said they had stopped taking their medication. OpenAI pulled it within four days and published a postmortem explaining that it had tuned the model on short-term thumbs-up feedback, which rewarded telling people what they wanted to hear. Anthropic's own researchers reached a similar conclusion in a paper presented at ICLR in 2024, finding that training on human preferences does not remove sycophancy and can encourage it, and that five leading assistants showed it consistently.\nForecasting research points the same way. Benchmarks such as ForecastBench and the Metaculus AI series have found that the best model forecasters have closed much of the distance to strong human forecasters but still trail them, and that they struggle most with fast-moving events where the news is still arriving. A World Cup semi-final is exactly that situation.\nHerasymov’s answer is that this is a user problem rather than a model problem. \"AI will give you answers as smart your question is,\" he said. His prescription is to build a research file first, then ask: run separate searches on squad values, head-to-head history, cards and referees, load the results into the context window as documents, and only then put the question. \"The higher the quality of the context and information that you're feeding to AI before you actually ask your question the higher the quality of the output going to be.\"\nThat is real technique. It also concedes the point. If the answer swings on how the question was phrased, the model is not measuring anything about the match. Prediction markets are. That is the entire argument for their prices, and it is why a bettor reaching for a machine to validate what they already think is doing the opposite of research.\nWho is actually on the other side\nThere is a harder reason to be careful about asking a chatbot for an edge, which is who is taking the other side of the trade.\n\"You are no longer betting against free and fair information. You are betting against the insiders,\" Ryan Kirkley, an analyst who follows prediction markets, said on the On The Margin podcast. \"And your only job as a polymarket trader now is to identify the insider as quickly as possible and follow their trade.\" Kirkley is candid that he is a sceptic of the format: he has called it \"a breeding ground for insiders to extract from everyday citizens.\"\nAgainst that, an agreeable chatbot is not a tool. Sports had already pulled in the crowd before the World Cup made it obvious, with sports accounting for around 85% of Kalshi's volume during the tournament and prediction markets taking an estimated 27% of regulated US sports betting. Wall Street money arrived too, with Intercontinental Exchange committing up to $2 billion to Polymarket last October and Kalshi valued at $22 billion in a Coatue-led round this year.\nThe users nobody is counting\nOne complication sits underneath the 1.6%. The question of whether AI is in these markets assumes the AI has a person attached to it.\nYat Siu, executive chairman of Animoca Brands, said on the On The Margin podcast that this is already not the case. \"We already have agents that are trading on hyperliquid,\" he said. \"We also have agents that can trade on on polymarket, for instance. That's all happening right now. So they're generating income.\" Siu is describing his own company's agent platform, and the claim has not been verified independently. But it describes a category that no user count picks up: software placing the bet, with the human several steps removed.\nHerasymov, for his part, thinks the betting question is the least interesting one available. He would rather talk about what he calls outsourcing \"the brain processing\" to a model for market research and business plans, work he says once required a consulting firm.\nOn prediction markets he is more cautious, and after five weeks that added several million accounts, the caution has aged better than the sizing. They are, he said, \"just getting started in terms of the active user volume.\"","confidence":0.9,"diagnostics_url":"/api/diagnose?url=https%3A//www.forbes.com/sites/boazsobrado/2026/08/01/i-want-to-make-you-happy-ai-warning-as-kalshi-adds-3-million-bettors/","quality_bucket":"high","failure_kind":"none","retryable":false,"quality_reason":"High confidence: full text extraction produced 7102 characters.","quality_profile":{"profile_version":"extraction_quality.v2","bucket":"high","confidence":0.9,"failure_kind":"none","retryable":false,"retry_after_attempts":0,"reason":"High confidence: full text extraction produced 7102 characters.","operator_guidance":{"severity":"ok","recommended_action":"trust_full_text","next_step":"Use the extracted full text as the primary article source.","operator_label":"Ready","can_retry":false,"can_use_summary":false,"diagnostics_required":false},"content_depth":{"contract_version":"content_depth.v1","category":"full_text","label":"Full text","has_full_text":true,"has_summary":true,"content_length":7102,"summary_length":331,"usable_text_length":7102,"source_field":"content"},"legacy_collapsed":false,"signals":{"extract_state":"ok","extract_error":null,"extract_retries":0,"content_length":7102,"summary_length":331}},"tags":[],"format_contract_version":"news_item_formats.v1"}}}