2024 Calendar
TechTalk Daily

AI and Context: Why Training AI with So Much Knowledge Leaves it Unable to Understand Context

AI and Context: Why Training AI with So Much Knowledge Leaves it Unable to Understand Context

By Daniel W. Rasmus, Serious Insights



Machine learning training sets focus on precise aspects of a domain. A machine learning algorithm trained on the identification of cancer cells does that job with accuracy and precision. The algorithm has seen so many cases that its ability to identify the next case proves uncanny to those not familiar with the underlying technology. Those who do understand the model are not amazed. The cancer identification algorithm reinforces their own learning that machine learning, or in some camps, artificial intelligence, can outpace a human’s ability to identify patterns among often subtle signals.

But unlike humans, the algorithm does only one thing well. It has no sense of context, no ability to understand causation, nor to identify other diseases related, or not, to cancer cell identification. Most machine learning remains in the space of savants. Highly accurate, almost preternaturally accurate at one thing, and as bereft at almost everything else as they are masters to the savant domain.

Context is important. I even stretch that often to say, context is everything.

Even GPT-3, which was trained on millions of pieces of content, doesn’t get context. It has no understanding of its applications, or why some segments of the population fear it. It simply regurgitates various bits of data from its store to create a complex pattern that meets the goal state input by the user. Despite tools like Chat-GPT that demonstrate the ability to parse text and offer complex analysis, it can only do so because its data set included data related to those questions. Stray away from the core, challenge it on deeper understanding, and it fails.

As I was re-reading some of my work at the Giga Information Group, I thought I would share a cover story from 2002 about the context that should challenge AI researchers to consider what lies beyond the precision of big data focused on singular topics—and why humanity’s biggest advantage over AI is the diversity of context. Knowledge can and will be modeled eventually. But those models will be unable to understand how each human sees answers through their unique lens of experience, how an answer to one may spark innovation, and to another fear. I doubt AI’s ability to ever acquire that type of empathy for context.

Finding Value in Context [from 2002 Giga Information Group Research Digest]

Making Metadata Meaningful Is the First Step

Almost all IT strategic plans at some point have contained the magic catchphrase of information technology: the right information, at the right place, at the right time. Only with recent development around the use of context in computing can IT even start to deliver on such a goal. Previous efforts have primarily used systems analysis and traditional application delivery to anticipate what people will need in a particular context. But that kind of effort is difficult to maintain and usually fails over time as new distractions divert investments away from maintaining shifting contexts to other priorities. Because the world changes so frequently, any manual effort at maintaining context requires a large investment for it to demonstrate sustainable success.

At its most fundamental level, context is information about information. Consider the approach of paleontologists who scour micro-sites to glean evidence about the environment where larger, more easily found animals lived at another time. In North America’s Hell Creek Formation, home to Triceratops and Tyrannosaurus Rex, much of the current investigation involves examining fossils, some so small they are impossible to see without a microscope, in order to determine the context within which these large animals lived.

The context includes small mammals, mollusks, trees, grasses, flowers and their pollens, along with information gathered by other specialists that include geology, stratigraphy, taphonomy and specialists in various animal and plant branches of paleontology. The records, both large and small, that lie beside T-Rex femurs and skulls tell more about what was happening in the Cretaceous Period than a fully articulated dinosaur skeleton.

If we apply this same approach to information, we realize that most of our unstructured content and, to a great degree, our structured content exists outside of anything to inform us of its context. When we retrieve a document, we may know who wrote it, when it was written and where it resides now, but we have little or no insight into its creation, its evolution, why it was created, how it has been used since its creation, or how well it performed on what it was intended to do.


The word intent is of particular note because it not only describes issues around the content but also describes issues about understanding the context of the information requester. One of the largest issues with search engines is that they must glean intent from one or two words entered into a search box on a Web page. And the search engines retrieve only matches against those words or, in more sophisticated systems, words like those words or related by concept to those words. They do not, however, understand the context of the requester, nor do they attempt to leverage the context of the documents to increase the search relevancy.

At the most fundamental level, these are not issues that lie beyond the capability of current technology, but because they are not embedded in the operating system, context becomes a separate technology investment, and like other intangible technologies, one that does not rise to the top of the mission-critical list. However, some organizations have made investments in context and found that they have provided significant returns on their investments through higher-value customer interactions and reduced response time in a variety of situations.

At Ford, for instance, the Autonomy suite (now Micro Focus Autonomy) of products is being used to bring a new level of personalization to the employee e-learning portal. Autonomy builds dynamic profiles of staff member interactions with information sources and maintains them over time, reflecting a historical context, but one that is also weighted toward more recent information.

When an employee engages with the e-learning portal, rather than being presented with a list of courses available or, in better situations, that match the employee’s role (which would be defined manually by manually anticipating context), the Autonomy system suggests courses that match the current profile of the employee. It does not rely on previously defined course organization or any structured approach to analyzing the employee behavior. The statistical model derives a context for the employee and matches it to courses that have a profile with a similar context.

Autonomy has also been used in call center and in customer relationship management (CRM) applications to integrate unstructured content with transaction systems by matching a transaction with unstructured content to provide increased visibility into the customer dialog or into alternative solutions to a problem.

Another technology that derives context is the Lotus Discovery Server (no longer available, like other Lotus products mentioned in this post). One of its unique features is to understand the relationships implicit in content stored within—in the Lotus suite of collaboration tools, the most interesting of which is the relationship to QuickPlace, which not only classifies but uses that classification so documents and people become inherent in the context of the QuickPlace environment.

A document in a QuickPlace environment is related to the ideas being discussed in the environment, other concepts from other documents in the space and to the people in the space. The people in the space are then related to the concept of the document and to the document directly. If one of them leaves the company, for instance, Discovery Server can help provide team-oriented context so someone other than the author can be quickly located to answer a question about a document.

Those examples look at the basic context and how it can be derived from environmental queues that exist in current technology. Other queues help inform the retrieval side of the equation. If a person is looking for information and is a member of a team environment, the system can enhance its understanding of a question by applying information about that team. In a delivery example, the system could use awareness technology to understand where the person is and what kind of device that person is using—which is also a type of context.

When a system understands information holistically—when it can put together that Leslie is working on a Java conversion project in Santa Clara for customer X and is requesting information from her cellular phone at 6 in the evening—and that string of metadata means there is likely an elevated issue with the project, or alternatively, that Leslie working outside of normal business hours is completely normal, then we will be able to at least begin to deliver on the promise of the right information, at the right place, at the right time … in the right context.

Read more from Daniel Rasmus at Serious Insights

Daniel W. Rasmus, the author of Listening to the Future, is a strategist and industry analyst who helps clients put their future in context. Rasmus uses scenarios to analyze trends in society, technology, economics, the environment, and politics in order to discover implications used to develop and refine products, services and experiences. His latest book, Management by Design proposes an innovative new methodology for the design workplace experiences. Rasmus’s thoughts about the future of work have appeared recently in Chief Learning Officer Magazine, Government eLearning!, KMWorld and TabletPC. A wildly popular article on CIO.com titled, 10 Lessons from Angry Birds That Can Make You a Better CIO, went viral on the Internet.