Bench Talk

In 2018, Google showed off impressive demonstrations of its Duplex technology, a voice AI trained to make short calls and book appointments. Trained to speak like a human with pause fillers like “umm” and “hmm,” Duplex successfully navigated real conversations to make dinner and haircut reservations. Six years later, why aren’t we at the point where our robots call other robots, and we never need to pick up the phone?

Part of it is simply because it’s impossible to differentiate helpful robocalls from spammy robocalls,[1] or businesses can even opt out of receiving Duplex calls entirely. Thus, Duplex remains an occasionally useful but niche feature, like Google’s “Hold for Me” feature that will wait on hold for you and spare you from the eternal elevator music.

With the meteoric rise of AI and all that it can do, a world that hates phone calls desperately wants an AI-powered personal assistant that can do it all. No longer just a luxury, the AI personal assistant should be able to manage our inboxes, calendars, and calls while relaying the executive summary to us. While certain AI use cases do sift through the noise for us, a fully functional personal assistant is unlikely to unseat the current low-tech workaround.

The Eyeball Economy

Your attention is the most valuable resource in the online economy. Everything you see online is the result of rigorous A/B testing to be as enticing as possible, down to the specific colors of links and notifications. Google made headlines in 2009 when it came out that designers had tested 41 shades of blue to see which was the most clickable.

This is just one example of the significant investments in learning how to keep your attention. Between our inboxes, news feeds, social media feeds, televisions, and good old-fashioned paper news, we have an incredible breadth of information at our fingertips, all competing for your attention. But what we gain in breadth, we lose in depth.

The more information sources we have, the harder it is for us to pay attention to any single source of information. We can easily see how the internet has accelerated this phenomenon—now we can check social media during a news broadcast or check our email during a meeting. “Undivided attention” has been left in the past.

Filtering Information with AI

The pieces seem to be in place for AI-powered assistants to take over—language models can digest information in the blink of an eye. There is already a great use case in online searches, as Google has been working diligently on “zero-click” searches, where the algorithm determines the answer you’re looking for and displays a relevant snippet without needing to click into a website (Figure 1). This seems to be catching on, too. The SEO agency SEMRush measured that just over half of all searches are zero-click.[2]

Figure 1: “Zero-click” searches allow for instant information without needing to visit another web page (Source: Author)

Zero-click searches make searching on your computer or phone faster, and they also power voice search results. While these snippets help with quick searches such as “what’s the weather today?”, long-form results like recipes still require us to click through to the results. This is where AI wins. Figure 2 shows a comparison between giving both Google and ChatGPT the same prompt: “potato soup recipe.”

Figure 2: Comparing recipe results on Google and ChatGPT (Source: Author)

Google gave me four nearly identical results, putting the burden on me to select one then navigate through the cookie pop-ups, email notification pop-ups, and banner ads to find the crucial “jump to recipe” button. The search results are still a few steps away from an actual recipe.

ChatGPT, on the other hand, immediately gave me a recipe. What’s more, I can ask ChatGPT to double the recipe, make it spicier, or substitute ingredients. I can even ask it to give me the recipe in the style of my favorite TV chef (Figure 3).

Figure 3: An example of the type of innovation large language models (LLMs) can generate. (Source: Author)

The Challenges in Expanding Beyond Recipes

Generative AI presents an opportunity to fundamentally change the way we get information from the internet. Instead of cycling through half a dozen news and social media apps to stay up to date every day, we could have one meta-app that curates an executive summary of our personal bubbles. It would be like an RSS feed that you can talk to. “Good morning. Your local competitive bowling team advanced to the regional championship, and this week, your city council is meeting about a bill that’s relevant to you. Grab a jacket, it's chilly today.”

Unfortunately, we are still far away from a true AI personal assistant. Dozens of vendors offer AI “personal assistants,” but most only function within a specific application as a chatbot. Applications and social media sites prefer you browse their services personally instead of sending your AI proxy. They make money from their human users, not from bots scraping their API. Social media giants and independent creators built virtual walls around their content when it came out that large language models (LLMs) like ChatGPT freely trained themselves on publicly available content.[3]

On the user side, data privacy concerns also hinder an AI personal assistant singularity. In the wake of data privacy regulations like GDPR, sharing data between applications is increasingly difficult. This is good for users, but also means it’s unlikely we’ll see a program that can interface with our inboxes, calendars, news feeds, and social media feeds. Our left hand does not know what our right hand is scrolling, and there’s a degree of safety in keeping our digital footprint fragmented. Siloing our usage across multiple applications protects us from a single breach stealing everything.

The Surprisingly Low-Tech Solution

Inboxes identify important emails and add dates to our calendars automatically, and news feeds tailor themselves to our interests, but right now, nothing ties it all together. Many people solve the problem of fragmented digital information with a simple analog solution: paper planners (Figure 4). Paper planners enable busy people to combine obligations from multiple sources into one single source of truth. And they come with stickers!

Figure 4: It's much more satisfying to physically cross things off your to-do list (Source: Andrey Popov / stock.adobe.com)

While generative AI offers exciting opportunities to change the way we search, personal and practical limitations restrain the potential of an actual AI personal assistant. Until the tech world comes up with a solution more comprehensive than the humble paper planner, the AI personal assistant will remain a fantasy. We’ll have to keep scheduling our own appointments for now.