I have been working with AI on a regular basis. I have trained models, built automations, used agents, and connected tools through MCP servers. I value what these systems make possible. But using them closely has also changed the question I ask about productivity: when a machine produces an answer in seconds, how much work remains before I can trust it?
Let’s consider a technical research question. I give ChatGPT, Claude, Gemini, Grok, and DeepSeek the same carefully prepared prompt: the domain context, the task, the required evidence, and the expected format. They can return different conclusions. This is not automatically evidence that the models are unintelligent. Some questions permit several defensible approaches. But where factual claims conflict, I still need to reconcile them.
In this case I now have five answers to read, claims to trace, sources to inspect, and assumptions to test. The cost of generating a response has fallen dramatically; the cost of establishing which response is dependable has not necessarily fallen at the same pace. I think of the additional effort of comparing these responses as a cross-model reconciliation cost.
The same issue appears with agents. Imagine an agent generates 100 research reports. If each takes 15 minutes to inspect, that is 25 hours of review. That does not prove AI made the task slower: writing the reports manually could have taken much longer. It does show why measuring generation time alone is misleading and we need to know how to address it.
I use two working labels for the resulting burden. Verification debt is the backlog of claims and outputs that have been produced or reused without sufficient checking. Comprehension debt is the gap between information an organization possesses and what the responsible people understand well enough to explain, challenge, and act on. These burdens are related but distinct. A report can be factually accurate yet poorly understood; another can be well understood but wrong.
The risks grow when unverified material enters a database, repository, or knowledge base and is later retrieved as though it were established fact. An error can be repeated, summarized, and amplified. On the other hand, simply demanding that humans read everything is not a scalable solution.
The response should be workflow design, not a retreat from AI. Use authoritative source links, claim-level evidence, reproducible checks, automated tests, uncertainty labels, and human review proportional to the consequences of error. When several models disagree, don't vote on the answer. Look for independent evidence or a test that can resolve the disagreement.
My concern is not that AI creates too much information. It is that organizations may confuse possessing information with understanding it, and generating a deliverable with completing trustworthy work.
The question I keep returning to is simple: after the generation, review, correction, and decision-making are counted, how much useful work did we actually finish?
Reducing comprehension debt
Verification debt has a workflow fix. Comprehension debt needs something else, because understanding does not live in documents. It lives in people. You can store artifacts in a wiki, but you cannot store understanding there. So improving the artifacts alone never closes the gap.
The first move is to make understanding part of the deliverable. Every consequential output should arrive with its claim, its evidence, what would change the conclusion, and what is uncertain. And someone specific should own understanding it. The test is simple: can that person explain the conclusion, the key evidence, and the main risk in their own words? If not, the work is not finished.
The second move is to generate less. AI made generation free, but attention did not scale with it. Every report nobody is assigned to understand becomes debt the moment it is created. The cheapest way to reduce the debt is to stop issuing it.
The third move is to refresh understanding when the decision is live, not when the document is written. A short walkthrough at decision time beats a perfect document from six months ago that nobody re-read.
Comprehension debt behaves like any debt. Stop borrowing so much, pay down what you owe by getting it into someone's head, and do not mistake the IOU for the money.