Filters: a Critical Component in the Safe use of AI
How to use generative AI safely
Generative AI is opening up a whole world of possibilities for creativity in Marketing. From automated content creation to the incredible videos featured on Sora: first impressions, it’s easy to see how AI will transform marketing campaigns. We’ll see faster development cycles, more personalized content and even new ways to engage with your consumer. It is probably the most exciting time to be in Marketing since the advent of the commercial internet and AI will be just as game changing or a technology as http.
One of the weird things about generative AI though, is that it does not act like you’d expect technology to act. We’ve come to expect technology to be predictable and repeatable – you use a calculator to multiply two numbers, you get the same answer each time. A set of rules were followed to arrive at a definite conclusion. But in the world of generative AI, models work probabilistically, making decisions based on probabilities or likelihoods. This is necessary to handle more complex tasks where the outcome is uncertain, such as for image generation. This is part of the fun of using generative AI to create, you don’t know exactly what you’re going to get when you enter a prompt. Some are better at prompt engineering than others, but regardless how good you are, you cannot dictate the specific outcome of any generation engine.
This makes filters essential in the world of generative AI. The large, well-known generation engines apply mechanical filters to stop the really bad stuff from making it to the end user and we see a lot of organizations who are using these generative AI engines add an additional layer, usually in the form of people, to double check the output not only for brand safety but to ensure it is on brand for the organization. As use of generative AI becomes more prevalent, this solution will not scale or there will be cases where human intervention won’t be possible, such as the AI-powered Test Drive A Win website Washington’s Lottery recently launched and took down. The concept was simple, users were asked to upload a headshot and throw a digital dart at a board to see which dream vacation they could pay for with their lottery winnings. Their face would be superimposed on a scene of that dream vacation as a bit of fun, instant gratification. A woman from Tumwater gave it a go and landed on Swim with the Sharks. She received an image of herself sitting on a bed in an aquarium with fish swimming around her where she was wearing a bathing suit bottom but no top, her bare breasts exposed. Washington’s Lottery logo was in the bottom right corner of the image. This shocked the woman and was reported to the Lottery who pulled the site down “out of an abundance of caution”.
Our brand governance platform BrandGuard acts as an automated filter system. With over 20 checks in five categories including Brand Safety, we screen for nudity, violence and profanity by identifying whether or not these elements exist in the image or text. Beyond basic Brand Safety, BrandGuard can be trained on Brand Principles such as look and feel, content guidelines and tone of voice as well as Precision Elements such as logo, fonts and colors to act as a brand filter, ensuring nothing off brand or inconsistent makes it through. Built using an ensemble of different models and methods, BrandGuard examines every piece of content to ensure brand suitability, supporting our company’s vision of enabling the safe use of AI for marketers.
If you're interested in a demo, click here or contact us at sales@brandguard.ai.