It’s frighteningly easy to poison a large language model (LLM). What’s even more unnerving is you don’t need access to do it.
A new study from NYU scientists dives deep into how much medical misinformation an AI can handle before it starts giving wrong answers. They found that even a tiny amount of fake info – just 0.001% of the data it learns from — can mess up the AI.
While the study focuses on someone deliberately trying to trick the AI during training, it also highlights a bigger problem: All the misinformation already floating around online that these AIs are learning from. Plus, there’s the issue of outdated medical info still hanging around in some databases. Sci-Hub, a free access/piracy site for scientific research, found only 15% of papers retracted by their publisher were removed from the site.
The good news is that any poisoned data must compete for attention with what might be accurate information. So, the ability to poison an LLM might depend on the topic. The bad news: LLMs get trained on data taken from the internet, and a lot of that is garbage.
Now here’s this week’s roundup of AI-powered martech releases:
- Amazon Ads announced a new capability within Amazon Marketing Cloud (AMC). This new feature allows advertisers to generate SQL queries for their desired audience using natural language. This eliminates the need to write code, significantly reducing the time required to develop audience queries. Advertisers can then run these queries within AMC to create new audiences and activate them across Amazon DSP and ads console.
- Cordial’s Cordial Edge uses multimodal AI to generate personalized marketing models for individual retail brands. These models consider various data types, including brand creative, illustrations, photography, and text, to optimize marketing messages for increased purchases.
- PearlMountain Limited released an update to its video creation platform, FlexClip. This update includes two new tools: an AI Video Generator and a Motion Tracking feature. The AI Video Generator enhances creative possibilities, while the Motion Tracking feature simplifies complex editing tasks.
- Jellypod launched its AI podcast studio. This platform offers features such as AI-powered voice cloning, content control, and robust content integration. Jellypod aims to simplify podcasting by assisting creators with all aspects of production, from ideation to final edits and publishing.
- Dappier announced the availability of AskAI banner ad deployment. This feature allows digital publishers to integrate a conversational AI answer engine within banner ads. This in-ad unit provides contextual knowledge of user content and offers advertisers a high-attention environment to connect with consumers.
- HUMAN Security partnered with TollBit to protect against unauthorized AI scraping bots, enforce robots.txt directives, and enable publishers to monetize their content.
- Jivox updated Jivox IQ DaVinci platform. This update includes deep API integrations with major retailer DSPs, such as Amazon DSP and Walmart Connect DSP. The platform combines first-party audiences with creative and media in a single workflow.
- Integral Ad Science launched its AI-driven Total Media Performance (TMP) solution. TMP integrates media quality signals with in-flight optimization to help advertisers achieve better campaign outcomes.
- Creative Realities launched AdLogic CPM+TM, a campaign planning and management platform. This platform offers programmatic functionality and is designed for in-store retail media networks. AdLogic CPM+TM enables advertisers and publishers to plan, manage, and optimize campaigns within a unified AdTech stack.
- BrandRep launched an AI-powered keyword generator. This tool uses artificial intelligence to identify relevant keywords for small and mid-sized businesses. The goal is to improve SEO performance and drive growth in competitive digital spaces.
About the author
Constantine von Hoffman is managing editor of MarTech. A veteran journalist, Con has covered business, finance, marketing and tech for CBSNews.com, Brandweek, CMO, and Inc. He has been city editor of the Boston Herald, news producer at NPR, and has written for Harvard Business Review, Boston Magazine, Sierra, and many other publications. He has also been a professional stand-up comedian, given talks at anime and gaming conventions on everything from My Neighbor Totoro to the history of dice and boardgames, and is author of the magical realist novel John Henry the Revelator. He lives in Boston with his wife, Jennifer, and either too many or too few dogs.
Read More
Constantine von Hoffman