Roughly half of the new articles published on the web are now primarily written by AI (Graphite, 2026). I had to read that number twice. In 25 years of running digital teams, production was always the bottleneck: never enough writers, never enough hours. That constraint is gone, and something quieter broke in its place.
AI-generated content is a quality problem not because machines write poorly, but because they write faster than humans can critically review. Fluent text triggers cognitive ease, automation bias lowers our guard, and verification quietly collapses. The result is a growing gap between how much we publish and how much we actually judge.
I have written before about the shift from predictive to generative AI and what it did to consumers. This article is about the other side of the desk: the editor, the teacher, the thesis advisor, the strategist. The person whose job is to say “prove it.” Here is why that job is failing, what the science says about it, and a framework to get it back.
Production Stopped Being the Bottleneck
The numbers tell a fast story. Twelve months after ChatGPT launched, primarily AI-generated pieces already accounted for roughly 36% of new articles on the web. Since early 2025, the split with human-written work has hovered around fifty-fifty (Graphite, 2026).
The data, honestly stated: Graphite’s research sampled tens of thousands of English-language articles from Common Crawl and classified them using three independent AI detectors, reporting false positive rates below 2% (Graphite, 2026). AI detectors remain imperfect instruments, and paywalled human content is underrepresented in open crawls. Treat the figure as a strong directional signal, not a decimal-point truth.
The strategic point is economic. The marginal cost of producing a competent draft has fallen to nearly zero. The cost of reviewing that draft well, checking its claims, questioning its framing, catching its subtle errors, has not changed at all. It still requires a trained human mind and real time.
When one side of an equation scales exponentially and the other side stays flat, you get a structural gap. I call it the verification gap. And it is not a technology problem. It is a psychology problem.
Why Fluent Text Turns Off Your Brain
Kahneman (2011) described two modes of thought: System 1, fast and intuitive, and System 2, slow and analytical. System 1 uses a shortcut called cognitive ease. When information feels fluent, familiar, and easy to process, we experience it as more true, and System 2 never gets called in to check.
Now consider what a large language model is optimized to produce: text that is maximally fluent. Clean grammar, confident tone, smooth transitions. AI output is engineered, by its very training, to generate the exact signal that tells your brain no further scrutiny is needed. The polish is the problem.
This is the same System 1 machinery that decides what we click and trust in search results. It served us reasonably well when fluency correlated with human effort and expertise. That correlation is now dead.
We saw this coming decades ago. Research on human interaction with automation found that people monitoring reliable automated systems develop complacency and automation bias: they stop checking, miss failures the system does not flag, and accept incorrect outputs it does produce. Critically, these effects appear in both novices and experts, and simple practice does not eliminate them (Parasuraman & Manzey, 2010). Pilots, radiologists, and plant operators all show the pattern. Content reviewers are not exempt.
So the human brain arrives at this moment doubly disarmed. Cognitive ease tells us fluent text is trustworthy, and automation bias tells us a reliable-seeming tool needs no supervision. Both signals are wrong, and both are firing constantly.
The Offloading Spiral: From Editor to Rubber Stamp
The 2025 evidence is uncomfortable. Lee et al. (2025) surveyed 319 knowledge workers across 936 real-world uses of generative AI and found that the more confident people were in the AI, the less critical thinking they actually enacted. Effort shifted from creating content to verifying it, but that verification weakened precisely where trust in the tool was highest.
Gerlich (2025) studied 666 participants and found a significant negative correlation between frequent AI tool use and critical thinking skills, mediated by cognitive offloading: the habit of delegating mental work to the tool. The effect was strongest among younger users, the very people building their professional judgment right now.
A necessary caveat, because honesty is the brand here: both studies are correlational and rely partly on self-report. They cannot prove that AI use causes cognitive decline. But they converge with decades of automation research, and the mechanism they describe matches what I see in real teams.
Picture an illustrative scenario. A content team that once published five researched articles a week now “reviews” forty AI drafts a day. Nobody decided to stop editing. The volume simply made real editing impossible, so review degraded into skimming, and skimming degraded into approving. Each individual shortcut felt reasonable. The aggregate is an organization that publishes things nobody has actually read critically.
GoriUX principle: a reviewer who fully trusts the tool is no longer a reviewer. They are a latency step between the machine and the publish button.
This dynamic does not stop at marketing. School assignments, university theses, peer review, market research, legal briefs: every layer of knowledge work that depends on a human checking the work is exposed to the same spiral. The gap must be addressed at every level where judgment is supposed to live, and the earlier in someone’s education it is addressed, the better, given where the offloading effect concentrates (Gerlich, 2025).
What Happens When Nobody Checks
If you want to know where unverified content leads, the machines have already run the experiment on themselves. Shumailov et al. (2024) showed in Nature that when AI models are trained on data generated by other models, they suffer model collapse: irreversible defects in which the tails of the original distribution disappear, and outputs drift toward repetitive, impoverished sameness.
Read that as a systems warning, not just a technical finding. A model consuming its own unverified output degenerates. A content ecosystem consuming its own unverified output degenerates the same way, and so does a team, a discipline, or a student. Verification is not overhead. It is the mechanism that keeps any information system tethered to reality.
There is a strategic silver lining for those who hold the line. Graphite’s research also indicates that AI-generated articles largely fail to appear in what Google and ChatGPT actually surface and cite (Graphite, 2026). Answer engines, as I argued when examining the behavioral science of AEO and user engagement, reward content that genuinely satisfies intent. Verified, accountable content is becoming scarcer, which makes it more valuable, not less.
The ethical line: publishing content no human has verified is not efficiency. It is externalizing the cost of error onto your reader, your student, or your client, who must now do the checking you skipped, usually without knowing they need to. Persuasion aligns real value with informed decisions; shipping unread machine output at scale does the opposite. If you would not sign it, do not publish it.
The Verification Budget: A Practical Framework
You cannot close the verification gap with good intentions, because the failure is psychological, not moral. Willpower loses to cognitive ease every time. What works is structure: treat verification as an explicit, finite budget that gets allocated deliberately, the same way you treat money or ad spend.
The Verification Budget, three rules you can apply Monday:
- Match effort to stakes, in writing. Before producing anything with AI, classify it: low stakes (internal notes) gets a skim, medium stakes (blog posts, reports) gets full claim-checking, high stakes (medical, legal, financial, academic) gets source-by-source human verification. Deciding the tier before generation prevents System 1 from deciding for you afterward.
- Verify claims, not prose. Fluency is not evidence. Extract every factual assertion, statistic, and citation from the draft and check each against a primary source. A beautifully written paragraph with a fabricated reference is worse than an ugly one with a real one.
- Keep one question the machine cannot answer for you: “What would make this wrong?” If you cannot articulate how the output could fail, you have not reviewed it. You have only read it. This single question forces System 2 back into the loop (Kahneman, 2011).
The same framework scales down to a classroom, where students defend their sources aloud, and up to a thesis committee or an executive team. The unit of quality control is always the same: one human mind, deliberately engaged.
Will AI reviewing AI solve this? Partially, for surface errors. But verification without independent human judgment is exactly the closed loop Shumailov et al. (2024) warned about. The auditor cannot share the biases of the audited.
Judgment Is the Scarce Asset Now
Every technology shift redistributes scarcity. AI made competent text abundant, and in doing so it made verified judgment rare. That is where the strategic advantage moved, and most organizations have not noticed yet.
In the warrior’s terms I keep returning to, impeccability means taking full responsibility for your actions, including the ones you delegate. Delegate the drafting, the formatting, the volume. That is intelligent use of a powerful tool. But the moment you delegate the judgment, you have not gained efficiency. You have vacated your post.
The professionals, teams, and institutions that thrive in the next decade will not be the ones that produce the most. They will be the ones still capable of looking at any piece of content and answering, with evidence, one question: is this true?
Delegate the work. Never delegate the judgment.
FAQ
Research by Graphite (2026), based on Common Crawl samples classified by three AI detectors, indicates that primarily AI-generated articles have made up roughly half of new articles on the web since early 2025. The figure is directional rather than exact, since AI detectors are imperfect and paywalled human content is underrepresented in open web archives.
Automation bias is the documented human tendency to over-trust automated systems: accepting their outputs without checking and missing errors the system does not flag. It affects both novices and experts and is not eliminated by practice (Parasuraman & Manzey, 2010). Applied to content, it means reviewers approve fluent AI drafts without genuine scrutiny.
Current evidence shows a consistent association, though not proven causation. Higher confidence in generative AI predicts less enacted critical thinking among knowledge workers (Lee et al., 2025), and frequent AI use correlates negatively with critical thinking skills through cognitive offloading, especially in younger users (Gerlich, 2025). The prudent response is to build deliberate verification habits.
Model collapse is the degradation that occurs when AI models are trained on data generated by other AI models. Shumailov et al. (2024) showed it causes irreversible defects: the rare, distinctive parts of the original data distribution disappear and outputs converge toward repetitive sameness. It illustrates what happens to any information system that consumes its own unverified output.
Treat verification as an explicit budget. Classify every piece by stakes before generating it, verify factual claims against primary sources rather than judging the prose, and require the reviewer to answer “what would make this wrong?” before approval. This forces analytical thinking back into a process that fluent AI text otherwise switches off (Kahneman, 2011).
Join 3,200+ subscribers who still read critically
If this tension between speed and judgment resonates with what you are seeing in your own team or classroom, let’s continue the conversation on LinkedIn or X.
References
Gerlich, M. (2025). AI tools in society: Impacts on cognitive offloading and the future of critical thinking. Societies, 15(1), Article 6. https://doi.org/10.3390/soc15010006
Graphite. (2026, May). AI now writes as many online articles as humans do. Graphite Five Percent Research. https://graphite.io/five-percent/ai-now-writes-as-many-online-articles-as-humans-do
Kahneman, D. (2011). Thinking, fast and slow. Farrar, Straus and Giroux.
Lee, H.-P., Sarkar, A., Tankelevitch, L., Drosos, I., Rintel, S., Banks, R., & Wilson, N. (2025). The impact of generative AI on critical thinking: Self-reported reductions in cognitive effort and confidence effects from a survey of knowledge workers. In Proceedings of the 2025 CHI Conference on Human Factors in Computing Systems (Article 1121). Association for Computing Machinery. https://doi.org/10.1145/3706598.3713778
Parasuraman, R., & Manzey, D. H. (2010). Complacency and bias in human use of automation: An attentional integration. Human Factors, 52(3), 381-410. https://doi.org/10.1177/0018720810376055
Shumailov, I., Shumaylov, Z., Zhao, Y., Papernot, N., Anderson, R., & Gal, Y. (2024). AI models collapse when trained on recursively generated data. Nature, 631(8022), 755-759. https://doi.org/10.1038/s41586-024-07566-y
Leave a Reply