Designing for AI Image Recognition
During my time at TrademarkVision and Clarivate, I had the opportunity to work on some AI powered products (back in 2016-2020, when AI was not yet cool!) that are now powering thousands of IP (Patents, Designs and Trademarks) searches all around the world!
One of the major challenges at the time, especially when working with Government clients, was trust.
In the background, the proprietary ML model identifies similar shapes and patterns, comparing millions of trademarks and designs, while the simple and familiar UI allowed users to feel comfortable and in control of the search process.
This approach wasn't just about initial comfort – it was also about building trust over time. By keeping the AI's role implicit, we allowed users to experience the benefits firsthand without having to explain the complexities of what was happening behind the scenes. You can still see this in action on the search page of IP Australia's Trademark search system.
As time went on and people became more confident, we included a few more smart features such automatically suggestion of keywords detected in the various images and other information identified by the models.
You can see in the screenshot below, that this iteration of the search results, available on the IP Australia Designs search, it's displaying relevant classification information and explicitly tells the user that their results are ordered by "image match".
Even more confidently, we now display matches of words and other elements found in the images uploaded by the user on the ELISS search platform on Wine Australia's website.
This iterative approach allowed us to:
- Increase confidence amongst the users of these platform,
- Deliver a better and better experience for all users,
- Generate trust in the users group, ensuring that future evolutions of the project would be possible.
As a designer working on AI-powered products in the early days (in 2017!), I experienced the challenges of integrating this technology in our applications, especially for risk-averse users and stakeholders.
As I mentioned earlier, the initial and biggest hurdle for us was trust. How do you convince users to rely on a complex, blackbox systems like AI, especially when dealing with critical IP data?
I believe that our baby steps approach helped us gain trust over time, and that the initial familiar and user-driven interactions helped us include more and more "smart" features in these systems.
As our audience became more familiar with the new search systems speed and accuracy, we added more and more "smart" functionalities, revealing the power of the AI models:
- Keyword suggestions and classifications based on the images uploaded, offering a window into the AI's analysis,
- Sorting based on image matching can now be changed by the users,
- Scoring is visible, to give poeple confidence on the results.
Today, looking back, it's incredibly rewarding to see how widely this solution has been adopted, as TrademarkVision and Clarivate deployed the core technologies to more Governments around the world, following similar design patterns.
As a designer, I find this fascinating, as we are slowly getting used to having AI powered tools help us in lots of mundane and repetitive tasks. But integrating AI into products where it's never been before it's a journey. that starts with cautious steps, and ensures users don't feel replaced but rather empowered by the technology designed and developed for the future.
I am writing this in 2024, a time when AI is now in every news title and we now have LLMs doing all sorts of tasks. This wasn't the case back in 2018 when I first started working at TrademarkVision, and we have come so far in such a short amount of time that is really mind blowing... I can't wait to see what the future holds, it's going to be an exciting journey ahead!