Chainlink used to curb LLM hallucinations
Chainlink says major financial players are using its oracle tech to reduce hallucinations in large-language-model workflows that process corporate actions, tackling what it describes as a roughly $58 billion annual problem. The announcement frames oracles as a guardrail—providing verified external data to models so outputs match real-world records rather than inventing facts. (x.com)
Banks still spend about $58 billion a year handling things as basic as stock splits, dividend notices, and merger paperwork, because the data often arrives as messy documents that humans have to re-check by hand. Chainlink says some of the biggest market plumbing firms are now using its system to stop large language models from inventing details while reading that paperwork. (dtcc.com) (prnewswire.com) A corporate action is a company event that forces investors and custodians to update records, like when a firm pays a dividend, splits its shares, or gets acquired. One bad field in that chain can send the wrong cash amount, wrong election deadline, or wrong share count into multiple firms at once. (prnewswire.com 1) (prnewswire.com 2) Large language models are good at turning long, ugly documents into neat tables, but they also hallucinate, which means they can state a date, ratio, or payment option that was never in the source. In finance, that is less like a chatbot typo and more like a clerk typing the wrong number on a wire form. (prnewswire.com 1) (prnewswire.com 2) An oracle is a data messenger that pulls information from outside a blockchain and checks how that information should be formatted before other systems use it. Chainlink’s pitch is that the oracle should act like a fact-checking rail between the model and the bank’s back office, so the model does not become the final authority on what happened. (chain.link) (prnewswire.com) The first phase of this project, announced in October 2024, had Chainlink, Swift, Euroclear, and six financial institutions test whether models from OpenAI, Google, and Anthropic could pull structured fields out of unstructured corporate action announcements. The output was turned into what the group called a “golden record,” which is one agreed version of the event that everyone can read instead of each firm rebuilding the same record separately. (prnewswire.com) (finextra.com) The second phase, announced on September 29, 2025, expanded the group to 24 organizations, including The Depository Trust & Clearing Corporation, Swift, Euroclear, UBS, DBS Bank, BNP Paribas Securities Services, ANZ, Wellington Management, and Schroders. Chainlink said its Runtime Environment checked multiple model outputs, confirmed the results, converted them into International Organization for Standardization 20022 messages, and sent them over the Swift network in minutes instead of days. (prnewswire.com) That International Organization for Standardization 20022 format matters because banks already use it for machine-readable financial messages, so the project is trying to fit new artificial intelligence tools into old production pipes instead of asking the industry to replace everything at once. The selling point is not “put finance on a blockchain tomorrow,” but “clean the data before it hits the systems banks already trust.” (prnewswire.com) (finextra.com) The scale of the mess explains why firms are trying this. Chainlink cited Citi’s 2025 Asset Servicing report saying the average corporate action event touches more than 110,000 firm interactions, costs about $34 million to process, and still leaves 75% of market participants manually revalidating data, while automation rates remain below 40%. (prnewswire.com) There is also an awkward irony here: Chainlink is selling oracles as a cure for artificial intelligence errors at the same time its own data feeds have had failures in crypto markets. In 2025, reports tied Chainlink oracle problems to a Moonwell exploit of about $1 million and an Euler liquidation event of more than $500,000, which is a reminder that “verified data” systems still depend on how well the verification is designed. (ambcrypto.com) (crowdfundinsider.com) What Chainlink is really arguing is that large language models should do the reading, while a separate control layer does the deciding. If that split works in production, the winner will not be the chatbot that sounds smartest, but the pipeline that can prove a dividend date, share ratio, or merger election matches the source record before money moves. (prnewswire.com) (dtcc.com)
