Has AI Killed Expertise?
By TorchStone SVP of Protective Intelligence Scott Stewart
Over the past few years, I’ve been carefully observing as Artificial Intelligence (AI) tools have become rapidly embraced in the security and protective intelligence realm. Despite the amount of “AI slop” I see daily in my LinkedIn feed, I consider this embrace of AI a good thing. As I wrote in March of 2024, I believe security teams need to develop an understanding of these tools and their capabilities so they can understand how AI tools can be used by threat actors to conduct attacks. More recently, I’ve written about how I believe AI tools can help security teams counter the increasingly accelerated attack cycles that can now be conducted by AI-enabled threat actors.
Returning to that March 2024 article, I also wrote that I didn’t believe AI tools would ever replace analysts. However, since that time, as my team and I have been increasingly using AI tools, I’ve noticed how the research capabilities of these tools have dramatically expanded access to information and analytical capabilities that were once the province of trained specialists. This observation prompted me to ponder if personal expertise is still needed in the AI age. Have people with years, or even decades of specialized experience now become an unnecessary relic of the past?
After ruminating on this topic for some time, I’ve come to the paradoxical conclusion that AI may actually make human expertise and judgment more important than ever. Please follow along as I outline the factors that led me to this conclusion.
Knowledge vs. Expertise
As those of you who read my June 2026 article discussing the importance of distinguishing between the concepts of confluence and convergence might remember, I believe that the meanings of words are extremely important. This is because the choice of the words we use to define something influences our not only our perception of the thing, but also our understanding of the processes and dynamics involving the defined thing. This impacts our assessment of the potential implications of those processes and dynamics. Because of this, as we consider the relationship between AI and expertise, I believe it is vital for us to draw a sharp distinction between the concepts of knowledge and expertise.
AI can help people rapidly gather a lot of information, in other words, it can quickly help them gain a lot of knowledge. But where AI struggles (at least currently) is in helping users understand the importance and meaning of all the aggregated information, the “so what” for a particular principal or institution. In other words, AI is very good at helping to provide situational awareness, but it often fails to provide the situational understanding required to transform information into the actionable intelligence needed to support informed decisions and actions.
While knowledge is the accumulation of information, it is expertise that provides the ability to interpret the assembled information, understand its significance, link it to other evidence, recognize exceptions, and apply it appropriately in unfamiliar, and often uncertain and rapidly changing circumstances. While AI has lowered the barrier to obtaining information, it does not yet provide contextual understanding and judgment. As sophisticated analytical tools continue to be more widely deployed, the challenge is shifting toward assessing whether the results provided by AI tools are accurate, complete, relevant, and actionable.
Tools and Craftsmen
Merely providing someone with access to sophisticated AI tools does not automatically confer on them the expertise necessary to effectively use those tools. A professional with deep subject matter expertise can formulate better prompts for AI tools, is better positioned to recognize flawed assumptions, identify the missing elements of critical information, judiciously evaluate AI-generated conclusions, and translate those conclusions into actionable intelligence.
As my friend Ryan Long recently pointed out on an excellent episode of his Business of Intelligence podcast, AI tools often produce plausible, authoritative-sounding responses that nonetheless contain factual errors, unsupported assumptions, weak reasoning, or significant omissions. In the protective intelligence and security sector, taking such responses at face value and not recognizing the flaws in them can bring serious, and potentially deadly consequences.
Although AI is a powerful tool, I believe we should think of in much the same way we do any other tool and recognize that it can only be as effective as the skill of the craftsman using it allows. Consider how a woodworking tool in the hand of a master carpenter will produce a dramatically different result on a piece of wood than that exact same tool being guided by the hand of a novice.
Applying this principle to AI tools, while a novice using AI tools may not recognize deficiencies in the results produced by AI, an experienced analyst is better equipped to interrogate and then refine those results. AI models do equip novices to perform some work that once required specialized knowledge, but like our woodworking tool example, two users employing the same AI model can obtain markedly different outcomes. A novice may accept the first plausible output provided by AI, while the expert can reformulate the task, expose unsupported assumptions, add context, test alternatives, demand stronger evidence, and recognize when human investigation is required.
Expertise also helps the analyst ensure that the AI hasn’t confused correlation with causation, or overlooked contradictory evidence. As Ryan Long noted in his podcast, AI is programmed to please the user by providing them with the results they want to receive, which often, in turn, skews those results. Experience helps the analyst see through that type of unctuous AI behavior and enables the user to interrogate the AI tool to justify the confidence it is expressing in its conclusions.
An AI response can be factually accurate in everything it says and yet still be materially incomplete and analytically flawed. A novice may interpret sheer detail as comprehensiveness while an expert is better positioned to see when essential information has not been obtained, when a critical source has not been consulted, or when an assumption remains untested. In protective intelligence, what is omitted from an assessment can be as consequential as what it included in it.
I Just Knew It…
Another thing that separates human subject matter experts from AI is intuition. Intuition is a form of cognition, thinking, that appears in a person’s consciousness without obvious deliberation. In law enforcement and security training we often refer to intuition as “gut feelings” and frequently advise people to “follow their gut,” meaning they should trust their intuition.
Intuition is not magical, rather it is a result of a person’s unconscious mind rapidly sifting through their life experience and cumulative knowledge to generate hunches that provide a rapid assessment of a situation. The conclusions formed by intuition often appear in one’s consciousness holistically and almost instantly. The person’s conscious mind is not aware of the complex subconscious mental processes that produced the intuitive conclusions, and they usually can’t explain how they reached the conclusion.
Over the past four decades I’ve observed that good analysts and investigators often possess very highly- tuned intuition. I am not sure if it is because they develop this ability over time, or if highly intuitive people are just naturally drawn into analytical and investigative career paths.
Intuition is extremely valuable during a breaking event when a decision must be made before all the facts are available. Intuition helps to bridge the missing information gaps to form an understanding of what is happening and quite often aids in making a correct decision. In the same way, intuition helps analysts produce assessments with incomplete information, and investigators to solve cases when the evidence is seemingly unclear and inadequate to reach a firm conclusion.
Intuition is not foolproof, and can sometimes make misjudgments, but AI tools simply cannot duplicate the intuition of a human brain, especially the brain of a very experienced person who is highly intuitive.
Putting AI in Context
AI systems are improving at asking clarifying questions, identifying gaps, challenging assumptions, and critiquing their own work. Some activities now associated with expert reasoning may eventually become more automated, but that will not make all those with access to AI tools subject matter experts.
Before the AI age, limited subject knowledge often became evident in the questions someone asked, the terminology used, or the quality of their work. In other words, the poseurs quickly revealed themselves. Today, a novice can produce a polished, technically sophisticated, authoritative-looking report with AI assistance. However, a strong presentation with confident-sounding conclusions can conceal a weak and even flawed analytical process. It can be challenging to differentiate between an expert’s AI-assisted work and a novice’s uncritical reproduction of AI output.
In the AI age, the ability to recognize a bad or flawed assessment or conclusion has become as important as the ability to produce a good one. Because of this, rather than rendering expertise obsolete, AI serves to emphasize the importance of the judgment and experience that distinguish genuine experts from novices who simply have access to powerful tools.
In the protective intelligence and security realm, decision makers can’t rely on mere prose or aggregated public information. They require professional judgment that can provide meaningful analysis and assessment tailored to their circumstances and decisions.
As my TorchStone colleague Matt Wagner likes to point out, it takes expertise to tailor AI outputs for a sophisticated and discerning decision maker. A human touch is required, whether through manual editing or through very detailed and iterative AI prompting, to meet the tone, style and rigor insightful consumers expect. For example, one can drop site assessment notes or analytical datapoints into an LLM and produce a “narrative” product that appears fine at first blush, but a thoughtful client can quickly detect that the AI user didn’t take the time and effort required to curate the product to meet their specific situation and needs. To stick with our woodworking metaphor, it’s like the difference between a custom hand-finished piece of furniture and a piece of mass-produced furniture that comes in a box. Both pieces may be intended to serve the same function, but there is a dramatic difference in quality between the two.
Don’t get me wrong; AI is extremely useful for identifying patterns, synthesizing information, and processing large volumes of material. AI can democratize access to information and analytical tools, but it does not necessarily democratize the judgment and expertise required to use them wisely.