John Ellmore

Search engines try to optimize relevance. This is hard. 15 years ago, search engines were much more rudimentary. They only factored in a few signals when determining the relevance of a possible result: PageRank, page keywords, manual curation, and similarly simplistic metrics.

One way to look at relevance is in terms of precision and recall. A search engine should optimize for:

The old simplistic metrics generally gave good recall, because a relevant search result probably had the correct keywords on the page, was linked to by other sites, etc. But precision was terrible. Unscrupulous sites crammed keywords into footers and created link farms to artificially inflate their PageRank. This forced decent sites to adopt at least some form of these cheap hacks in order to compete.

Users got used to sifting through poor-quality results. It might have taken several page visits before you found something halfway decent. Unscrupulous sites got more page views, while the quality results would get second-page treatment.

As time has gone on, search engines have gotten smarter. A veritable soup of signals is now used: bounce rate, user geolocation, page performance, domain/entity reputation, natural language processing, machine learning, etc. From a precision and recall perspective, this is good! Both measures have substantially improved. Most users would say that search results today are much "better" than they were several years ago.

Ultimately, search relevance is subjective to the human performing the query. A human's state of mind and intention when searching cannot be losslessly transferred to the search engine (which would be the holy grail of human-computer interaction). Instead, the search engine has to take the limited bits of information that it gets from the user--the query, geolocation, etc.--and use them to extrapolate the user's actual intent and expectations. With these additional signals, search engines can perform that extrapolation much more accurately.

A surprising trait

This is the same process that happens in basic human interaction. A good communicator can read subtle social signals (the bits of information), create a mental model of the other party, and adjust their communication accordingly. Similarly, a good salesperson can do this with their customers, and match them to the perfect offering for their needs.

This is empathy. Perhaps more precisely, it's cognitive empathy. And just like a good communicator needs to empathize to understand their audience, decent human-facing algorithms must to empathize with the user in order to deliver optimal results. An empathetic algorithm understands the user's state and tailors its output accordingly.

Slowly approximating people

Practically speaking, human-facing algorithms are iterated and optimized over time to improve their quality (often to increase revenue). In order to improve in quality, it has to more precisely model the user and the user's state of mind. As a result, human-facing algorithms will naturally evolve towards empathy.

In the early days of search engines, an SEO expert could look at search engine ranking algorithms and see the writing on the wall: search engines will increasingly adapt their algorithms to mimic the rankings that a real human would give. Keyword stuffing and other hacks might've worked at the time, but they would become increasingly irrelevant. Search engine algorithms are effectively just crappy heuristics for humans, but those heuristics will become more accurate as search engines get better at empathizing with searching end users. So effective SEO should focus on valuable content to the actual human who will consume it, with the understanding that rankings will improve over time as empathetic search engines increasingly recognize the value in the content.

(This is not to say that common SEO strategies are irrelevent or unnecessary. There will always be a need for specific techniques and technical tweaks. But many of those techniques, like optimizing page speed or writing good copy, end up directly benefitting the end user of the content--regardless of whether that was the primary intention or not.)

The need for empathy in algorithms is apparent in many common use cases today:

We can imagine a hypothetical superhuman who knows us incredibly well, and could set our thermostat perfectly for us and recommend products that we'd actually like. Right now, many of the real-world algorithms that do these things can't perform them with the same optimality that a human person could. But we can expect that algorithms will continue to assimilate more of our information, create better models of us, and eventually perform at the same level as (or better than) a real human.

Privacy--where empathy becomes crippled

In human relationships, empathy is generally easier when you know more about the person. In order to become more empathetic, algorithms must acquire more information about the user. As a result, humans must be willing to sacrifice privacy to gain empathy.

We might trust that hypothetical superhuman controlling our thermostat with our personal information, because we're comfortable with trusting people; it's part of human nature to build trust and desire to maintain that trust. But when the other party becomes a cold database in the hands of a corporation, this tradeoff becomes much more uncomfortable.

As a result, privacy neuters the benefits that we might get from empathetic algorithms because it places limits on how "close" we allow an algorithm to get to us. This is unfortunate, but it's an absolute necessity.

Architecting for (and with) empathy

Recommendation algorithms are a dime-a-dozen; you can find simple algorithms which let you plug in various interaction events and user traits, and which spit out a list of objects to recommend to the user. These absolutely have their place.

But a more holistic approach means taking a step back and truly considering how we can be empathetic. After all, humans are more complicated than the articles they read or the videos they click on. Perhaps we can pick up on higher-order signals that would only be readily apparent to another human and factor those in to our algorithms' behaviors.

Any sort of empathetic behavior adjustment runs the risk of seeming creepy, invasive, or manipulative. So correspondingly, we must personally demonstrate empathy when designing thes behaviors.