Unsplash+

The AI Productivity Trap Is Coming for Your Lab

Why faster won’t mean freer in scientific research, and what early-career scientists need to do about it now.

The grant deadline is three weeks out. You open ChatGPT, paste in last year’s R01, and ask it to revise the specific aims for your new direction. Twenty minutes later, you have a polished draft. You should feel relief.

Instead, you find yourself thinking about how many more grants you could realistically submit this cycle. How many manuscripts you’ve been “meaning to write.” How the postdoc down the hall already has three preprints up since January. The time you just saved doesn’t feel like surplus. It feels like an obligation to produce more.

That quiet shift, the one from “I got my afternoon back” to “I should be doing more with this afternoon,” is the AI productivity trap. It’s coming for every scientific lab in the country, and most of us are walking into it without seeing the shape of what’s about to happen.

The Promise You Were Sold

The pitch is clean. AI summarizes papers, drafts manuscripts, codes your analysis, formats your figures, writes your responses to reviewers, and cleans up your grant prose. The pitch says you’ll spend less time on the administrative drag of being a scientist and more time on the part you actually love. The thinking. The discovery. The science itself.

The reality is turning out to be different. Yes, AI does compress those tasks. The studies are real. An MIT experiment published in Science found that professional writers using ChatGPT completed tasks 40 percent faster, with output quality rising 18 percent (Noy & Zhang, 2023). A study of 5,000 customer support agents in the Quarterly Journal of Economics found a 15 percent productivity gain on average, with the largest gains going to less experienced workers (Brynjolfsson, Li & Raymond, 2025). Code generation, literature summarization, formatting work — these compress in measurable ways. The lab-level pitch isn’t fictional.

But here’s what the studies also show, and what almost nobody is reading carefully. A large NBER working paper that linked AI use to actual Danish payroll and hours data across 25,000 workers and 7,000 workplaces found that the average AI user saved just 2.8 percent of their working hours, and that AI chatbots had “no significant impact on earnings or recorded hours in any occupation” (Humlum & Vestergaard, 2025). Most of the saved time didn’t become leisure or deep-thinking time. It got reabsorbed into new tasks, many of them generated by managing the AI itself: prompt engineering, output verification, integration work.

Even more striking, a randomized controlled trial by METR on experienced open-source developers found that those using AI tools were 19 percent slower to complete real tasks from their own projects, even though they believed they were 20 percent faster (Becker et al., 2025). Subsequent analysis has surfaced selection-effect critiques of the study, and the question is genuinely contested, but the broader picture across the literature is clear: the headline productivity numbers from controlled writing experiments do not survive contact with real work environments intact.

The pattern is consistent. The unit cost of producing scientific work is falling. The volume of scientific work expected of you is rising to absorb the savings. You’re not faster than you were three years ago. You’re producing more, in less time, with a thinner margin of judgment behind it.

Why Science Gets Hit Harder Than Most Fields

The trap operates everywhere knowledge work happens, but science has four features that make the absorption mechanism especially brutal.

The grant ratchet. Faster grant drafting raises the expected baseline of how many proposals you submit. Study sections see more applications. Funding rates fall further. The faster everyone writes, the more everyone has to write to maintain the same odds. Nobody designed this. It emerged from the collective use of the tool.

The verification tax in citations. AI hallucinates references with breathtaking confidence. A draft that looks polished can contain a citation that doesn’t exist, a finding misattributed to the wrong paper, a statistic that was never published. A 2023 study in Nature Scientific Reports tested ChatGPT on multidisciplinary literature reviews and found that 55 percent of GPT-3.5’s citations and 18 percent of GPT-4’s citations were entirely fabricated, with another 43 percent and 24 percent (respectively) containing substantive errors (Walters & Wilder, 2023). A 2025 Deakin University study of mental health literature reviews found that 56 percent of GPT-4o’s citations were either fake or contained errors, with 64 percent of fabricated citations linking to real-but-unrelated papers, making them especially difficult to catch (Cleveland Clinic JMIR analysis, 2025). In casual writing, this is embarrassing. In scientific writing, it ends careers. The work of catching these errors falls back on you, and the stopwatch that measures “AI saved me time” never accounts for the audit you now have to perform on every sentence.

The coordination bottleneck doesn’t move. Most of what slows down a lab isn’t writing speed. It’s IRB review, IACUC approval, core facility queues, collaborator response times, manuscript revisions, and the four-month gap between submission and decision. AI accelerates the parts you control individually. It does nothing for the parts you don’t. The result is that you produce more drafts that sit in more queues, waiting for the same human bottlenecks to release them.

Judgment atrophy in a field that runs on judgment. This is the deepest one. Science is the discipline of making careful inferences from incomplete data under genuine uncertainty. That’s not what AI does. AI does pattern completion across what already exists. The more you let it summarize the literature, draft your hypotheses, write your interpretations, and generate your figures, the less you exercise the muscle that actually does the science. And that muscle, like any other, weakens when it isn’t used.

The scientists I worry about most are the ones who don’t notice this happening. They feel productive. The outputs are increasing. Their CVs look stronger than ever. And quietly, the thinking is happening less often, less rigorously, and less originally.

What Vision-First Actually Means

I’ve spent 30 years working with scientists at Harvard, Stanford, and the Broad on the part of their careers that no one trains them for. The leadership, the strategy, the decisions about what to spend their finite scientific lives on. Across all of it, one pattern shows up reliably. The scientists who navigate the chaos well are the ones who’ve done the work of clarifying their vision before they reach for any tool, AI or otherwise.

At EDC we call this Vision-Alignment-Strategy, or VAS. The architecture is simple. Vision comes first, and it’s about who experiences the impact of your science. Not what you’ll publish. Not what you’ll be known for. Who, specifically, lives differently because of what you discovered. Patients. Students. Trainees. A whole field. Name them, by category if you have to, but name them.

Alignment is second. It’s the question of whether your daily work, your collaborations, your grant strategy, your tools, and your time allocation actually serve the vision. Most scientists I work with have never asked this question explicitly. They’ve inherited a default workflow from their training and never inspected whether the workflow points where they’re trying to go.

Strategy is third, and only third. This is where tools enter. AI included. The discipline is keeping AI confined to this layer, the execution layer, never letting it migrate up into the vision layer or the alignment layer. AI can help you articulate your specific aims more crisply once you know what they are. It cannot help you decide what they are. The moment you let it try, you’re outsourcing the thing that’s actually yours to do.

The companion frame is what we call EQ:IQ. Intellectual rigor and emotional integration as a single operating system, not two separate skills. AI is heavily IQ-flavored. It can pattern-match across vast corpuses, structure arguments, generate plausible interpretations. It has no EQ. None. It cannot tell you whether the direction you’re taking will erode the relationships in your lab, whether you’re chasing the work because it serves your vision or because you’re afraid of being seen as falling behind, whether the prestige of a particular collaboration is worth what it’ll cost your sleep and your marriage.

The trap closes most reliably on scientists who let AI’s IQ-heavy outputs substitute for their own integrated EQ:IQ judgment. The outputs look right. They sound right. They publish well. And five years in, you look up and realize you’ve built a career that satisfies every external metric and serves no one you actually wanted to serve.

Practices That Keep AI in Its Lane

A few specific habits I recommend to the early-career scientists I work with.

Write your vision by hand before any AI sees it. Pen, paper, no laptop. Write a narrative of one day in the life of your science five years from now. Who are you helping. What does the lab look like. What does the science actually do for the world. Do this slowly, in two or three sittings, with a cup of tea or a glass of wine. The slowness is the point. The instrument you’re using is your own attention, not your typing speed.

Use AI as adversary, not author. Once you have a draft of your specific aims, your hypothesis, your interpretation, hand it to AI and tell it to attack. “Where is this argument weak? What competing hypothesis am I dismissing too quickly? What would a skeptical reviewer say?” AI is genuinely good at this. It’s much worse at generating the original thinking. Use the tool where it actually adds value.

Verify every reference, every statistic, every claim. No exceptions. If AI surfaces a citation, open the paper. Read the relevant section. Confirm the claim. The half-hour you spend on this is the audit you owe yourself. The day you stop doing it is the day a fabricated reference makes it into a published manuscript with your name on it.

Establish AI-free zones in your week. A morning. A whole day. Whatever you can defend. Use it for the deep work that AI can’t do for you. Reading a paper carefully and writing notes by hand. Thinking through an experimental design without prompting. Talking with a collaborator without checking your phone. The muscle of original scientific thought needs regular use or it weakens.

Fix coordination before adopting more tools. If your lab meetings are unfocused, your decision rights are unclear, or your communication channels are scattered, AI will accelerate the chaos, not the work. Audit the human system first. Tool acquisition without system clarity is how you turn a busy lab into an overwhelmed one.

Run a vision audit every six months. Sit with your written vision from earlier. Ask honestly whether the last six months of work served it, drifted from it, or quietly replaced it with something else. Drift is normal. Unnoticed drift is the problem.

The Question Worth Holding

The most useful question I’ve ever asked scientists I work with is this. If AI disappeared from your workflow tomorrow, would you still know what you’re trying to do?

If the answer is yes, AI is in its proper place. It’s accelerating a journey you’ve already chosen. If the answer is no, something has quietly inverted. The tool is shaping the destination, not just the route. And that’s a problem no productivity gain is worth.

Science is one of the few human activities that genuinely runs on judgment, originality, and the long patience of letting a question stay open until you actually understand it. AI is going to be part of that work going forward. I’m not anti-tool, and the scientists I work with aren’t either. But AI belongs in the execution layer of your scientific life, not the navigation layer. The navigation is yours. It always has been. The cost of forgetting that is higher than the productivity gains will ever offset.

The labs that will matter in 20 years are not the ones that adopted AI fastest. They’re the ones whose principal investigators kept their vision human, their judgment intact, and their tools in their proper place. Vision first. Alignment second. Strategy last. EQ and IQ integrated, always.

The trap is real. But the way out isn’t to refuse AI. It’s to build the architecture that keeps AI honest about what it is and what it isn’t.

That architecture is the work. The tools are just the tools.

Sources

• Becker, J., Rush, N., Barnes, E., & Rein, D. (2025). Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity. METR. arxiv.org/abs/2507.09089
• Brynjolfsson, E., Li, D., & Raymond, L. (2025). Generative AI at Work. The Quarterly Journal of Economics, 140(2), 889–942. academic.oup.com/qje/article/140/2/889/7990658
• Humlum, A., & Vestergaard, E. (2025). Large Language Models, Small Labor Market Effects. NBER Working Paper №33777. nber.org/papers/w33777
• Noy, S., & Zhang, W. (2023). Experimental evidence on the productivity effects of generative artificial intelligence. Science, 381(6654), 187–192. science.org/doi/10.1126/science.adh2586
• Walters, W. H., & Wilder, E. I. (2023). Fabrication and errors in the bibliographic citations generated by ChatGPT. Scientific Reports, 13, 14045. nature.com/articles/s41598–023–41032–5
• MIT Technology Review (2025, December). AI coding is now everywhere. But not everyone is convinced. technologyreview.com/2025/12/15/1128352

This article was refined with AI tools to enhance readability, but the ideas and perspectives are entirely my own.