Five years ago, if someone published a detailed AI architecture diagram, I assumed they had spent years designing systems. The diagram was a signal. It stood in for a body of work I could not see directly, and the assumption was usually safe.
I no longer make that assumption.
This is not because people have become less knowledgeable. It is because AI has become good at producing work that looks like expertise. The gap between looking competent and being competent has always existed. What has changed is how cheaply that gap can now be crossed in appearance.
Consider what has happened to the cost of production. A few years ago, a clean reference architecture, a maturity model, a framework, a technical explanation, or an executive summary each represented real effort. Producing one required someone who understood the material well enough to organize it. Today, any of these can be generated in minutes. The output is often coherent, well-structured, and persuasive.
That is a real achievement, and it deserves to be named as one. The barrier to learning complex ideas has fallen. The barrier to communicating them has fallen further. People who would once have struggled to express a technical thought can now produce a clear artifact that carries it. This is good for education, good for access, and good for anyone trying to move an idea from their head into a form others can use.
But there is a distinction worth holding carefully.
Lowering the barrier to creating a knowledge artifact is not the same as lowering the barrier to acquiring knowledge.
The artifact and the understanding were once more tightly coupled than they are today. The coupling was never absolute. People have long produced polished work they did not fully grasp, through ghostwriters, templates, borrowed code, and vendor material. What changed is the strength of the link and the cost of breaking it. Producing a convincing artifact once required at least enough understanding to assemble it, or the budget to hire someone who had it. Generative AI has weakened the link and sharply reduced the cost. The artifact can now exist with little understanding behind it, and from the outside, the two are difficult to tell apart.
This cuts in more than one direction. The same tool that lets someone skip understanding can also help someone build it. A person who works through several iterations of a design, questioning each choice and testing it against what breaks, may come away understanding most of it. The problem is not that AI produces empty work. It is that the finished artifact looks the same whether understanding sits behind it or not.
This is where my own field gives me a specific way to look at the problem. In digital forensics, we draw a hard line between an artifact and a finding. An artifact is a recorded trace or data object that may carry evidentiary significance. A finding is a conclusion supported by evidence that an accountable practitioner with the relevant competence is prepared to defend. A file on a disk is an artifact. The statement that a particular person created it, knew it was there, and controlled access to it is a finding. The artifact is the easy part. The finding is where judgment lives, and judgment is what survives scrutiny.
Applied here, competence becomes visible when the artifact is challenged: in the ability to explain its choices, identify its failure conditions, and respond when reality departs from the design.
Imagine a reference architecture labelled highly available. It shows two regions, duplicated application tiers, and replicated data. Then someone asks what happens when the central identity provider is unavailable. Can the services still authenticate? Can operators obtain emergency access? Does the failover path depend on the service that has failed? If the person presenting the architecture cannot answer, the diagram may still be useful, but it has not established resilience. The challenge did not invalidate the artifact. It revealed the boundary of the understanding behind it.
AI has become good at generating artifacts. It has not removed the need for the judgment those questions test. It has made it possible to skip that judgment while still producing something that looks like its product.
It would be easy to turn this into a complaint about AI. That would be a mistake, and it would also be inaccurate. The tool is not the problem. The problem is older than the tool. It is how we decide who to trust.
The pattern is consistent across eras. There was a time when a polished slide deck signaled that someone had done serious work, because assembling one took serious work. Then the deck became easy to produce, and it stopped carrying that signal. Today a polished AI diagram carries the weight the deck used to carry. Tomorrow it will be a polished AI-generated demonstration, something that looks like a working system, presented by someone who could not rebuild it. The surface changes. The underlying error stays the same. We keep mistaking polished communication for demonstrated competence.
This is the part of the shift that matters most, and it is broader than AI.
Our traditional signals of expertise are becoming unreliable. For most of professional life, we have relied on proxies. A credential, a title, a clean document, a confident explanation. These proxies were never infallible, but institutions, reputation, credentialing, and the cost of production made several of them harder to produce without the competence they implied. Generative AI has weakened some of those constraints, without any intent to deceive on the part of the person using it. Someone learning a subject in good faith can now generate work that outpaces their understanding, and they may not realize the gap exists. The signal degrades whether or not anyone is acting in bad faith.
It helps to be concrete about what these signals do. Trust in professional life is allocated through them. Who gets hired, who gets promoted, who receives funding, who is invited to speak, who is treated as an authority, whose testimony is believed, whose work clears review. In each case a decision-maker reads a proxy and infers the competence behind it. When the proxy can be produced without the competence, the inference breaks, and every decision built on it inherits the error. This moves the issue past appearances. It becomes a governance problem, because the systems that distribute opportunity and authority are calibrated to signals that no longer mean what they once did.
That degradation is the real story. It is not a story about impostors. It is a story about measurement. When the instruments we have always used to read competence stop being trustworthy, the responsible response is not suspicion. It is verification. The answer is not to distrust every polished artifact. It is to stop allowing polish to substitute for evidence of competence.
This is the discipline that becomes central in this environment. Assurance is the practice of establishing justified confidence in a system or a claim through evidence rather than appearance. It does not ask whether an output looks correct. It asks whether the output can be shown to be correct, tested against reality, and defended when the pressure is on and the output is challenged.
In practice, this changes where competence becomes visible. It shows itself less in what someone publishes and more in how they respond when the work is questioned. Can they defend the reasoning behind a choice? Can they hold up when an assumption changes? Can they explain the trade-offs they accepted and the ones they set aside? Can they name what they are uncertain about? Can they concede when the evidence no longer supports the conclusion? These are not tests a generated artifact can pass on its author's behalf.
Much of the discussion focuses on what AI can produce. A question that deserves equal attention is how we verify what it produces, and how we verify the people presenting it. Stated plainly, polished information artifacts are no longer the scarce resource. Anyone can generate them. What remains scarce is justified belief, the ability to say why something is correct and to stand behind it.
So the question I find myself asking has changed.
It is no longer, "Can this person create an impressive AI diagram?" The impressive diagram is now available to almost everyone, and by itself it tells me far less than it once did.
The better question is, "Can they explain why it is correct, identify where it may fail, and revise it when the evidence changes?"
In the age of generative AI, expertise is becoming less about what you can produce, and more about what you can defend.