Surfacing insights through AI (not that one): Augmented intelligence

How technology and human expertise can be combined to help deliver a Digital Transformation that lasts beyond tomorrow

One of the fundamental principles of digital transformations is the need to be able to centralize an organization’s data and make it accessible to those who need it. But, this step is only the first in a journey to achieving a digital transformation that lasts beyond tomorrow. Considering the vastness of information in the modern era, just possessing it is not enough. You must be able to find and draw upon the insights that the information provides.

We sat down with InfoDesk’s Director of Editorial Services, Tiki Archambeau and Taxonomy Manager, Ryan Williams to discuss what organization needs to be doing in the ‘Age of Information’ to ensure they are surfacing the insights that really matter:

We’ll start simple, how can an organization ensure that they’re making the most out of their information?

Tiki: Such an important question! The hottest, most valuable piece of information could be sitting on someone’s hard drive or a long lost shared drive collecting dust. We’ve all been there: “Now where did I file that email…?” From my experience, the best way to make the most of information – that is, get the most value from it – is to: 1) have it categorized and easy to find; & 2) allow it to be distributed or accessed via multiple channels. My email filing system makes a phone book look petite. But I can find even the most mundane detail quickly. For the juicy information, I post it on our internal Wiki for the team to access.

In the news business, we recognize that news (information I call “actionable intelligence”) loses its value with the passage of time. Which is why we work hard to update reports with the very latest information as close to deadline as possible. Yes, there is absolutely value in real-time 24/7 information delivery – but there is more value in targeted and relevant intelligence briefings. It really comes down to what you SHOULD know versus what you COULD know. Letting humans engage in the final stages of refinement is a big boost on that front.

Ryan: To get the most out of their information, organizations should consider making use of a wide range of tools and techniques. Filmmakers with big budgets employ the latest CGI effects; but many of them do good old-fashioned practical effects using ropes and wires, scale models, and actual explosions. If your script calls for a superhero to transform into a lion on-screen, then you probably need some high-end digital effects to pull that off. But, if what you need is a shot of a fiendish villain petting a housecat while explaining their evil plan, then you should probably just find yourself an actor and a cat to put in front of the camera. 

Similarly, for organizations, the question should never be, “What’s latest and greatest? What’s fancy and shiny and new?” but rather, “What tools are the best for the job?” Sometimes an organization’s digital transformation will be fueled by the thoughtful integration of the latest AI-driven tools and techniques. In other cases, human curation of information will provide the most efficient and effective results. And more often, the answer will not be to choose one approach over another, but instead to choose both. 

Tiki, as our resident expert, what are the benefits of human curation?

Tiki: [Cracks knuckles, grins] There are so many benefits that I made a career of it! I’ve given presentations with a continuous scrolling list of the advantages of human curation – mainly to trigger that primal gene that causes us to cheer for the humans while watching The Terminator. But seriously, the human brain is the most powerful supercomputer ever invented. It can discern opinions, plausibility, impact, level of appropriateness, and more.

So for the sake of time, let’s just take the top 3 benefits:

  1. Fake News – Trust me when I say the misuse of this phrase in recent times is on par with walking by a crooked picture every day. The ability to catch what is truly fake news is a perk that Artificial Intelligence (AI) just doesn’t have. It’s a combination of plausibility, language cues, and knowledge of sources that puts humans over the finish line first on this front.
  1. Advocacy journalism – In today’s media landscape, it feels like everybody has an agenda. Truisms are reframed to fit one person or organization’s world view. Opinion is perfectly fine, but when people or organizations rely on factual information to make critical decisions, truth is paramount. Machines are just not as adept as humans at filtering this… yet.
  1. Propaganda (aka state-controlled media) – With the resurgence of authoritarian regimes globally, one of the biggest victims has been the press. Media outlets towing the party line are not reliable when it comes to revealing a country’s vulnerabilities. Imagine investing in a new factory in a state where a rosy picture is painted by the press, only to learn later that ethnic tensions and climate change threaten that region. Humans understand the value of corroboration and trusting our gut better than those gutless computers.

And Ryan, a similar question to you – based on your expertise, what are the benefits of machine learning and semantic enrichment in information management?

Ryan: Tools based on machine learning techniques have the potential to grow with your organization, and to change with your organization’s changing needs. You can train AI-based information retrieval tools to learn and adapt over time, and to respond to the evolving needs and interests of your workforce and customers. With those kinds of tools at your disposal, the humans in your organization concentrate more of their efforts and brainpower not only into finding information, but also into analyzing it, understanding it, and making good, well-informed decisions based upon it.

Semantic enrichment and machine learning are a (very) little bit like jazz. When a jazz artist composes a tune, they often write down only the chord changes and a brief melody to be played at the beginning and the end of the song. The performers on stage use those chord changes and that melody as an underlying structure for improvisation. No one tells a jazz musician what specific notes to play, but the composer’s choices will limit the range of possibilities for what notes might sound good in context. 

The principles underlying artificial intelligence and machine learning are not entirely dissimilar. Developers “compose” machine learning algorithms that determine the range of possibility for what the machine might do. Instead of writing chord changes, developers write code that the machine uses to “improvise” results in new contexts. This is one of the principles underlying InfoDesk’s semantic searching: Instead of explicitly telling the machine exactly what information to retrieve for in one specific context, we are instead giving the machine a framework that it can use to identify the right semantic information in many different contexts, and with only limited instructions. With machine learning, humans write the tune, but we leave it up to the machine to figure out how to play it. (An important caveat: None of the above should in any way be construed as an endorsement of the idea that algorithms can or should be involved in playing jazz. Algorithms just ain’t got that swing.)

Right, so we’ve got the benefits of each in isolation. Now let’s talk about the advantages of using both in conjunction?

Ryan: There is, in fact, no alternative to using both, because even the latest and greatest artificial intelligence techniques are heavily dependent on the work of humans. A common misperception about artificial intelligence is that it does not require the involvement of human minds. But this could not be further from the case. Machines cannot learn if humans do not first carefully create the context in which that learning can take place. Here at InfoDesk, for example, our semantic searching relies not only on algorithms developed by humans, but also on the careful work of the professional taxonomists who create and maintain the specialized vocabularies that power our search tools. Artificial intelligence isn’t magic; a computer cannot simply pull up the right information like a rabbit out of a hat. In fact, it’s not a trick at all, but instead a careful application of computing power in conjunction with human intelligence. 

Tiki: So true, Ryan! As much as I cheer for Team Humans, only AI can mine the reams of data that give us options – and value. Rather, humans CAN, but not many are willing to hang out for the one million years it’d take for a human to turn that data around. As I mentioned above, there is a time-value to information. But just as there are many ways to copy text, there are many ways to tweak AI. The semantic taxonomy you and your team are working on is a game-changer. By finding and collating information in new ways, it opens possibilities we’ve not yet thought about. I’m imagining looking up “Climate Change”, then having the information in front of me to not only report the latest news, but have my team attach custom graphs, chart out social media trends, and provide analyses and predictions in one easily digestible dashboard. Now we have to have a separate conversation about NLP! Next time….

Ryan: You’re spot on Tiki. You will, indeed, be able to use our in-development Subject dictionary to return finely-tuned semantic results for “Climate Change,” and also for related topics like “Environmental Activism” and “Environmental Health.” And you bring up an excellent point here about time, and about all the amazing things your Editorial team can do when they are expending less of their energy on finding documents, and more of it on analyzing that information and adding value. The same idea holds true for any organization: Machine learning isn’t just about making use of the power of technology; it’s also about freeing people up so that they can unleash more of their human potential on the job. 

I’ll note also that InfoDesk’s current and future tools are only the tip of the iceberg for us when it comes to machine learning. (Hmm, maybe I should have used a different metaphor there, with all our earlier references to climate change). We are hard at work on some additional tools and products that are rooted in artificial intelligence and are designed not only to deliver you all the latest information on a topic, but also to bring the stuff that interests you most to the top.

Surely being so contrasting, there are some significant challenges in combining the two approaches into a single solution. What are these?

Tiki: All the disadvantages of being human become magnified in such an effort. For one, we’re all biased. Ryan’s favorite color may be purple, but mine is green. If I am coding on Ryan’s AI team, I may unintentionally give more weight to green data that surfaces positively. And maybe I give purple data demerits because… well, it’s purple. Yuck. Then purple data doesn’t appear as often. This has real world complications. Amazon, for example, had to scrap their AI recruiting tool after it favored men for technical jobs. IBM’s Watson provided unsafe and incorrect suggestions for treating cancer patients (all hypothetical, thankfully). This is not proof of devious intent as much as proof of the fallibility of human history and rigidly linear thinking.

Ryan: For the record, I feel uncomfortable claiming purple as my favorite color, as that hue has been so thoroughly and gloriously claimed by Prince. But yes, you are completely right, Tiki. Amazon is not alone in getting into trouble for (unwittingly or otherwise) building biases into their AI tools. When working on products like our Subject dictionary, we have tried to correct for the limitations of our own perspectives. For one, this means taking great care with the language we choose to use in our taxonomy terms. But it is also important when training NLP models to select documents that represent a broad array of perspectives. If you are trying to train a machine learning algorithm to bring back results on “Advertising,” but the only documents you give it to train on are overtly sexist advertisements and transcripts of Don Draper’s meetings, you will effectively be training your machine to return sexist results. With information processing of any sort, the rule “garbage in, garbage out” applies. And it is important to remember that algorithms are wholly incapable of making ethical judgements. That kind of analysis is better left to humans for sure.

What is it about the combination of the two that provides not only a solution for today’s challenges but one that will last into the future?

Tiki: Any good process has checks and balances. After I pack my son’s bag for a weekend trip, my wife will inspect it and say: “Where did you put his toothbrush?” A mad rush to the bathroom ensues, quickly correcting the equation. The marriage of AI and humans is an ideal match. AI can handle volume and effectively fit results into a rhetorical circle. Humans can look over what lands in that circle and accept or reject it, corroborate to see that it aligns with expectations, set up another circle with overlap, resize the circle, or perform any number of exercises that tests and adapts to a situation. We humans may be biased, commit frequent mistakes, be forgetful, or get too much sun at the beach, but we are great at adapting.

Ryan: Very well put, Tiki! To precariously extend your metaphor: Artificial intelligence can help you sort through the world’s billions of toothbrushes to find the one you want. But it’s still up to you to brush your teeth. Also, as a human, you get to enjoy a vacation every now and then, whereas algorithms do not (and cannot). Poor things.