Prediction markets moved an estimated $50.6 billion globally in July 2026, more than legal U.S. sportsbooks handled in bets that month, according to a Pew Research Center analysis cited by CBS Sports. I had my team at 5W test how ChatGPT, Claude, Perplexity, Gemini, and Google AI Overviews describe Kalshi and Polymarket's legal status right now, and the answer depends less on which engine you ask than on whether that engine can search live before it answers.
Why Do ChatGPT and Claude Answer Differently About Kalshi?
Every AI search engine we tested hedges the same question the same way: is Kalshi a financial exchange or a gambling operation. That hedge is correct. The Commodity Futures Trading Commission calls Kalshi a designated contract market in its own press materials. Washington's attorney general calls the same company's contracts an "illegal gambling operation" in an August 13, 2026 court filing. Two regulators have not agreed with each other, so an AI engine should not pretend they have.
What separates the engines is not the hedge. It is whether they know Washington issued that order at all, which is a question of retrieval, not reasoning.
Why Do Financial Press and Betting Sites Describe Kalshi Differently?
My team reviewed 36 sources covering this category, split roughly between financial and markets press on one side and sports and gambling trade press on the other. Financial outlets, including the CFTC's own releases, default to exchange and derivatives language. Sports and gambling trade press default to bet and wager language, describing the same CFTC-regulated products.
In most categories, that kind of split tracks commercial incentive. Here it tracks jurisdiction instead. The CFTC's own materials read one way. State attorneys general's own materials read the other way. The primary sources are litigating the vocabulary, not just the outcome, and AI systems are learning both vocabularies directly from those filings.
What Happens When AI Search Can't Pull Live Data?
I had Claude answer the same 62 questions about prediction markets twice: once with a full year of research in front of it, and once relying only on what it learned in training, with no search. On questions about which states currently allow Kalshi to operate, the no-search answers missed the record entirely, not partially, because Washington's order, New York's lawsuit, and Minnesota's injunction all happened after that training data was collected.
Why it works: a language model's knowledge is fixed at training time unless it retrieves current information. Washington's King County Superior Court signed a final order against Kalshi on August 13, 2026, requiring the company to geofence Washington residents by September 2, 2026, or pay $120,000 a day, according to the Washington Attorney General's press release. No model trained before that date can know it happened without a live search.
5W published the full data behind this test, including the state-by-state legal record and the scoring for all five engines, in The Prediction Markets AI Visibility Index 2026.
How Do the Five AI Engines Compare on Currency?
Google AI Overviews and Perplexity scored highest in our modeling because both are search-grounded by default. ChatGPT scored lowest because it does not always invoke browsing without an explicit prompt cue. Claude, when search is enabled, scored highest of all, and the same model with search turned off dropped to the bottom of the scoreboard. Same model, same reasoning. The only variable was whether search was live, which isolates retrieval as the actual driver of accuracy in this category rather than model quality.
Why This Matters Beyond Prediction Markets
More than a third of consumers now begin product research with AI, not Google. The prediction markets story shows what happens when the facts an AI needs are moving faster than most categories ever move. A cold answer to "is Kalshi under investigation" does not just miss a detail. It can give a clean answer that is wrong, because the CFTC opened a marketing-practices inquiry into Polymarket in June 2026 that a no-search model has no way to know about.
That is not a Kalshi problem or a Polymarket problem. Any regulated company whose legal status is actively being litigated is exposed to the same gap, the same one 5W measured last year in the US Sports Betting and Gaming AI Visibility Index, where operator citation share concentrated among the five largest sportsbooks.
What Should Operators and Their Communications Teams Do About It?
Publish dated, sourced legal updates the moment a filing or ruling lands, because a claim without a date cannot be checked for staleness by a reader or a model. Treat your own press releases as retrieval material, not just legal notices, since AI systems training on a category are already learning vocabulary directly from CFTC and state attorney general filings. Watch how sportsbooks describe the encroachment of prediction markets on sports betting search terms, because the same users are asking the same AI search engines about both.
The through-line is simple. An AI search engine's answer is only as current as its last search. In a category where a court order can change the legal answer in a week, that is the whole game.
I built 5W's AI Search and Generative Engine Optimization practice around exactly this problem: keeping a brand's AI-facing footprint dated, sourced, and current as facts change. 5W runs AI Search (GEO) programs for brands across consumer, B2B, financial services, healthcare, and technology, building the machine-readable footprint that gets brands cited, not just ranked. Learn more at 5W's Generative Engine Optimization practice.