SAN FRANCISCO, June 27, 2025 (GLOBE NEWSWIRE) — CTGT, which enables enterprises to deploy AI for high-risk use cases, announced today an upgrade to its platform that removes bias, hallucinations and other unwanted model features from DeepSeek and other open source AI models. In testing, CTGT was proven to enable DeepSeek to answer 96 percent of sensitive questions, compared to 32 percent originally.
CTGT demonstrated its new platform upgrades for the first time at VentureBeat Transform 2025, which took place June 24-25 in San Francisco. CTGT was one of just seven finalists for the VentureBeat Innovation Showcase, and was named the top winner of the Showcase by judges at the conclusion of the event.
In recent months, CTGT has built infrastructure for a leading insurance provider to ensure that their AI correctly coded complex claims properly, leading to the insurer avoiding over $1M+ in financial exposure in just one week. CTGT is also architecting the ship of GenAI email compliance with a major bank by automatically updating the company’s compliance models in production, without having to take them offline.
According to McKinsey, $67.4 billion in global losses were linked to AI hallucinations across industries in 2024. Deloitte data has shown that 47 percent of enterprise AI users admitted they made at least one major business decision based on hallucinated output. The problem appears to be worsening, with newer versions of some models hallucinating more than the older ones. The recently launched ChatGPT 4.5 model has a 30 percent hallucination rate – its worst ever. DeepSeek R1’s hallucination rate is 14.9 percent, and much of its misinformation is intentional, coded into the model by the Chinese government.
In a high-stakes environment, these models can’t be trusted.
CTGT launched in February 2025 with $7M in funding from Gradient (Google’s early-stage AI fund), General Catalyst, Y Combinator, Liquid 2, Deepwater and notable angels including François Chollet (Google, creator of Keras), Michael Seibel (Y Combinator, co-founder Twitch) and Paul Graham (Y Combinator). Initially CTGT was focused on reducing the compute needed to train and deploy AI models – something it did quite well, reducing compute by an order of magnitude by eliminating the need for fine-tuning, prompt engineering and backpropagation.
“DeepSeek’s introduction of R1 drove many enterprises to reach out to us looking for help – they were interested in taking advantage of DeepSeek’s low cost and speed but concerned about bias,” said Cyril Gorlla, co-founder and CEO of CTGT. “In testing, CTGT was able to identify the exact model features causing bias and remove them, so that DeepSeek R1 could function without bias. CTGT has since performed the same tests on other open source models including Llama, proving that it can identify and remove bias, hallucinations and other unwanted features in minutes, without any time-consuming and costly fine-tuning or retraining. This capability is now available to any enterprise using the CTGT platform.”
“AI model trustworthiness is a major obstacle to enterprises achieving return on their AI investments,” said Gorlla. “Each week I’m fielding calls from Fortune 500 executives wanting advice on how to solve this problem. After doing custom work in this area for the past few months, we’ve now productized our approach into the CTGT platform and made it available to anyone.”
About CTGT
CTGT enables enterprises to deploy AI for high-risk use cases, with a breakthrough mathematical method that removes hallucinations, censorship and bias at the model level automatically in real time, with no training or fine-tuning needed. Companies are using CTGT’s platform to improve model accuracy, reduce compute and get models deployed up to 500x faster. Based on technology developed at the University of California at San Diego, CTGT is backed by General Catalyst, Y Combinator and notable angels including Paul Graham and François Chollet. Learn more at CTGT.ai.
Media contact:
Michelle Faulkner
Big Swing
617-510-6998
michelle@big-swing.com
A photo accompanying this announcement is available at https://www.globenewswire.com/NewsRoom/AttachmentNg/decd52d7-2fc3-4ee2-8a31-0bf45462e989
