Welcoming Our New Robot Overserfs

Daniel Tunkelang
4 min readMay 12, 2023

These are early days for generative AI, but it already feels like we are at the dawn of a revolution in human-computer communication. Practically overnight, we can communicate with machines conversationally, with both parties using natural language rather than keywords, facets, or SQL.

So, what’s next?

Search as Communication

In my two decades working on search engines, I have characterized search as a communication problem. Even when our technology for content understanding and query understanding was primitive — at least by today’s standards — I advocated for an approach to information seeking that deemphasizes ranking and instead focuses on what Gary Marchionini called human-computer information retrieval (HCIR).

The focus of HCIR to empower information seekers by offering them more control, but at the same time demanding that they take responsibility for this control. We should not expect machines to read our minds. Instead, it is our responsibility to tell machines what we want, and theirs to help us express ourselves and explore the available information.

The emergence of ChatGPT and other generative AI tools has not changed my thesis that search is fundamentally a communication problem. I am not convinced that the current instantiations of generative AI solve the search problem — in part, as Amelia Wattenberger eloquently explains, because text inputs have no affordances. But I do believe that generative AI has the potential to transform the way that we seek and find information.

Simple Interfaces for Simple Needs

The “ten blue links” search paradigm, which heavily emphasizes ranking and encourages people to enter short keyword queries, has persisted for years because it is so simple.

This paradigm has seen incremental improvements, the most notable being “universal search” that blends diverse result types (e.g., web pages and images) in a single page. Search engines have also leaned into question answering, providing answers to some queries above the links to results.

A simple search interface is a great fit for a large portion — perhaps the majority — of people’s information-seeking needs. If searching requires typing and reading, then most searchers prefer less of both. For simple information-seeking needs, search engines should help searchers satisfice.

Taming the Complexity

But our current search interfaces do not do so well when people’s needs become even a little more complex. Search engines struggle to understand anything beyond than a simple noun phrase. To the extent that search engines take context into account, they tend to give it too much or too little weight. Perhaps most importantly, today’s search engines do not provide robust ways for searchers to provide feedback on the results. Indeed, we have not progressed much beyond faceted search in the past two decades.

Compared to our current search interfaces, we humans are pretty good at communicating complex needs to one other. We use feedback to direct conversations, ask clarifying questions when we don’t understand each other, and are pretty good at incorporating contextual and non-verbal cues.

Generative AI has already demonstrated some of these capabilities. ChatGPT does an impressive job of accepting and adapting to feedback. It also maintains a enough conversational memory to incorporate previous interactions in the conversation as context for the current one.

So, are we there already?

Making AI Feel Natural

I suspect that we will communicate with AI the way we communicate with one another when the experience feels similar. While we have adapted to texting in the past few decades, most of us expect to see and hear each other when we communicate. Indeed, the past few years have taught us that there’s no substitute for in-person communication, but that the closest alternative is video chat. Mark Zuckerberg keeps telling us that we will all plug into the metaverse, but clearly that is not happening anytime soon.

And yet Generative AI is already close to offering us this sort of natural communication. We’ve had decent voice recognition for a while, but now generative AI can respond to us conversationally with a natural-sounding voice. It can produce video that will become increasingly photorealistic and real-time as technology progresses. We may have to get through an uncanny valley, but at this rate it seems reasonable to predict that we are only a few years away from being able to communicate with generative AI in the same ways with communicate with one another.

I believe that, when — and only when — the technology reaches that point, we’ll embrace conversational interfaces beyond current niche applications.

The Future is Here

William Gibson, who popularized the term “cyberspace”, famously said that “The future is already here — it’s just not very evenly distributed.” Generative AI is here, and we are only starting to explore its applications.

Having spent much of my career working on search engines, I am excited about the potential of generative AI to transform information seeking.

I, for one, welcome our new robot overserfs.