Improving the AI Experience
Vendors don't need to wait on regulators or scientists to make these 5 simple fixes.
Part Two of Two
Yes, it’s finally happened. The word “Luddite” is fashionable again. Signs of AI backlash are imminent, but it’s clear Wall Street won’t be abandoning this cluster of technologies any time soon.
Federal regulation such as the Hawley-Blumenthal framework is being contemplated, but the tech industry has reason to be skeptical of legislative solutions. In the meantime, here are five steps that vendors can take to improve outcomes.
AI Wishlist
This list is written from the perspective of UX (my actual day job) rather than a trader’s, academic’s, or analyst’s perspective. Ticker symbols are not included because detailed product reviews and comparisons simply don’t exist yet for many AI solution providers. Very few are publicly traded.
Disclose. Most of us like to know when we are talking to a real person or a simulation on the phone. Same with web or video chat. It helps us manage expectations. Just as food products tell customers what is artificial, software can also label the “artificial ingredients” in a given interaction. A small button or icon (no larger than the lock 🔒 icon commonly used to show the presence of encryption) can indicate the presence of predictive or generative AI.
Flag apparent bias before release. The major problem with predictive AI is that it’s a black box. You know approximately what types of results you’re looking for, and you’re sorting through a vast field with multiple variables (often interdependent) influencing your search. But there isn’t a straightforward formula in the same way that, for instance, banks calculate acceptable costs for housing when determining if they will make a loan. You don’t want your brand to get a reputation for racism or other forms of bias.
So be proactive. Run audits. Make sure everyone in management understands the difference between test, training, and validation data. Write this phase of auditing into your sales contracts and project timelines. If the end results are out of whack with existing demographics, throw the model out.
If you are called upon to explain your system’s discriminatory practices in a court of law, “just because” is never a good enough answer.
Provide citation. The extra step of identifying publicly available sources for citation cannot be optional. Maybe it’s time to train a few models on Wikipedia. Or just add links search engine results on related topics and keywords as a starting point. That’s what human users are stuck doing right now.
Bring back the sandbox. Reliance on the cloud is profitable for Google and Amazon, but it squelches innovation. With the exception of Large Language Models (LLMs) most machine learning models don’t actually require supercomputers to run. Yet client-side tutorials and libraries, even those from well-known names such as PyTorch and Tensorflow, are sporadically maintained, if at all. In the past, Apple offered tools to build machine learning directly into client-side apps for cell phones, laptops, and tablets. By and large these still function but opportunities for customizing the models themselves are limited. Important gaps still exist that AI can fill — particularly around assistive technologies, reducing carbon footprint, and content moderation on community-run forums. Coming up with the right tools doesn’t necessarily require a large budget or a Ph.D. It does require working code from which to refine and iterate.
Invite feedback. There is no question that AI does some things well. But on other tasks, it underperforms. Surveying users and gathering the types of performance ratings that are standard for other customer service interactions should be the norm, not the exception.
Just because your employee is not human, there’s no excuse to skip their performance review!
Bonus Content
Me vs the Machine
I finally tried out ChatGPT from OpenAI. Scroll down if you’d like to read the alternate version of this article (authored exclusively from the prompt).
While I couldn’t find a “feedback” icon in the UI, my overall experience was positive. Simply put, it’s a great alternative to StackOverflow. It excelled at answering programming questions with specific language constraints and even provided code samples. While it didn’t help me directly with any of the specific tasks on my list, I haven’t yet closed the browser window. That should count for something.
AI on AI
Before writing this article, I asked the prompt for its perspective. The result was grammatical and logically coherent. Enough so that I’m prepared to go head-to-head.