Editor’s Note: Today’s post is by Mohamed Mannaa. Mohamed is a Lecturer in Plant Pathology at Cairo University. Dr. Mannaa also serves as a reviewer and editorial board member for several international scientific journals. Reviewer credit goes to Chef Haseeb Irfanullah.

I write this as a scientist trained before artificial intelligence became part of the everyday scientific workflow, and as an editor and reviewer who now sees its influence from both sides of the publication process. I’m not opposed to AI. I use computational tools myself and accept that technological progress in science is inevitable. What concerns me is not the presence of AI, but what it is increasingly replacing.

In the current research ecosystem, AI is no longer limited to technical assistance. It is now routinely involved in drafting manuscripts, restructuring arguments, refining responses to reviewers, and — more quietly — contributing to peer-review reports themselves. Reviewers, asked to carry an expanding workload with little recognition, are understandably tempted to rely on such tools. Authors then respond, again with AI assistance, to critiques that may themselves be partially machine-generated. The result is a strange loop: scientific exchange still looks human-led, yet the cognitive labor is progressively outsourced. Humans remain formally “in the loop,” while the struggle of thinking, the part that once shaped judgment and craft, is gradually removed.

I remember what it meant to write before this era. Writing a single paragraph often required reading dozens of papers, abandoning weak ideas, and slowly learning how to synthesize evidence into something coherent. It was difficult, frustrating, and time-consuming. But that difficulty was not incidental. It was how scientific thinking was formed. As many writing scholars have argued for decades, writing is not just a way to report what we think; it is one of the main ways we arrive at what we think (see, for example, Emig’s classic discussion of writing as a mode of learning).

When I look back at my own work produced before widespread AI tools, I feel pride, not only in the results, but in the intellectual path that led to them. At that time, a manuscript was generally a reliable reflection of the author’s understanding. Today, I find it harder to assume that polished writing necessarily corresponds to deep comprehension.

This concern extends well beyond scientific publishing. It has become normal for people to ask AI systems to draft routine emails, which are then read and answered by other AI systems. At first, this looks like efficiency. Over time, however, it can erode confidence in one’s own ability to communicate. When people repeatedly outsource wording, judgment, and tone, they begin to distrust their own voice. Eventually, the hesitation is no longer about expression, but about thinking itself.

I know similar anxieties accompanied earlier technological shifts. As a PhD student, I used software to generate graphs and run statistics. My supervisor — trained decades earlier — described the hours he spent drawing figures by hand and computing standard deviations manually. From his perspective, my generation already had it easy. He wasn’t wrong. But there is a critical distinction: those tools replaced manual effort, not cognitive effort. What we are facing now is different in kind. AI does not merely accelerate tasks; it performs synthesis, interpretation, and articulation — activities that were once inseparable from intellectual development. This is precisely why scientific communities are now debating where transparency, authorship, and accountability should sit when AI tools are used (for example, WAME’s recommendations that chatbots cannot be authors and that use should be disclosed).

As a microbiologist, I find it difficult not to view this through an evolutionary lens. In biology, long-term dependence on a host or partner can lead to reductive evolution: organisms that outsource essential functions lose genes, autonomy, and complexity over time. They may survive efficiently within a system, but they no longer function independently. A striking example is Mycetohabitans rhizoxinica (formerly Burkholderia rhizoxinica), an endosymbiotic bacterium living inside Rhizopus fungi. Descended from free-living relatives, it underwent substantial gene loss after adapting to an intracellular niche and became specialized around producing rhizoxin, a metabolite crucial for its host’s pathogenic lifestyle (a detailed genomic account is here). It persists, but largely by serving a narrow function within a larger system. It is difficult not to see a parallel risk in AI-driven knowledge systems. Survival inside an efficient structure does not require autonomy, only functional contribution. When thinking is systematically externalized, participation can remain while agency quietly declines.

The danger is not an immediate collapse of scientific quality. It is more subtle: a gradual shift from understanding to fluency, from judgment to plausibility, and from wisdom to surface coherence. Over time, we may produce researchers who can generate convincing text across disciplines without possessing the depth required to truly advance them. And because the output looks “good,” the loss is easy to miss, until it accumulates.

So, what would “self-control” look like in practice, and why would anyone choose it, especially under pressure to publish, teach, and compete? The honest answer is that incentives matter. If hiring, promotion, and funding reward speed and volume above clarity and originality, dependence will grow. But boundaries are still possible, even within imperfect systems.

One simple line to draw is between using AI as a tool and using AI as a substitute for cognition. Tools can help with grammar, formatting, translation, or code troubleshooting. The red zone begins when AI is doing the argument itself: generating the logic, drafting the claims, producing the “voice,” or writing the peer review you are supposed to think through. In that zone, the person’s name remains on the work, but the mind has stepped back.

A practical rule I’ve found useful (and increasingly recommend to trainees) is: write the first version with your own brain, then use tools only for polishing. If you cannot explain the central argument without looking at the AI output, you don’t own the work. Another is to treat peer review as one of the last protected spaces of human judgment: if we automate reviews, we send the message that evaluation is just another box to tick.

None of this requires rejecting AI or pretending we can roll back time. It requires something more modest and more difficult: acknowledging that thinking is a finite human capacity that weakens when unused. If the struggle to think is systematically removed from scientific practice, the ability to think deeply will erode with it. No amount of computational efficiency can compensate for that loss.