I saw a talk at a conference last year where the presenter said something along the lines of “if you’re not using artificial intelligence in research, you’re going to be left behind.” This bothered me then and still bothers me today.
We've been watching a slow-motion train wreck in academia. Researchers are using artificial intelligence (AI) tools to write papers, analyze data, and generate ideas. Meanwhile, institutions scramble to create policies, publishers debate disclosure requirements, and everyone pretends we can manage this with the same systems that worked when the biggest threat was someone fudging a PCR gel image (which, of course, still happens).
We can't.
The traditional approaches to research integrity—honor systems and detection tools—are broken in the age of generative AI.
The AI Spectrum
AI use in research isn't a simple case of "cheating" versus "legitimate use." It exists on a spectrum that hasn’t been adequately mapped. At the most problematic end, we have blatant academic dishonesty: submitting entirely AI-written papers without disclosure, fabricating data or hallucinating citations, creating peer reviews. These cases are clear violations that most would universally condemn.
At the acceptable end, we find uses that most would consider legitimate: AI-based citation formatting, coding and debugging analyses, brainstorming, basic editing and proofreading. Even here, though, questions remain about what level of disclosure is appropriate.
But, like many things in academia, the real challenge lies in the massive gray area where most people operate. This includes using AI to restructure your literature review, having AI help with English language editing, letting AI suggest analytical approaches for your data, using AI to generate multiple versions of your abstract, or having AI help respond to reviewer comments. Some of this feels reasonable. Some feels questionable. All of it depends on context, disclosure, and norms that don't exist yet. The gray area is huge, growing, and context-dependent. What's acceptable varies by discipline, institution, journal, and even across researchers.
The Honor System Is Dead
Academic research has always relied on trust. We trusted that researchers accurately reported methods, didn't fabricate data, and properly attributed sources. This worked when misconduct required deliberate, technically complex deception. AI changes everything.
For the first time, research assistance is available to anyone with an internet connection. An undergraduate can generate a literature review draft in minutes that used to take a doctoral student months to write. The barrier to what we might call "research enhancement" (or misconduct, depending on your perspective) has dropped to nearly zero.
More fundamentally, the honor system assumes clear boundaries. But AI use exists on that messy spectrum where boundaries are fuzzy, contextual, and shifting. When a tool can help with everything from writing to analysis to idea generation, asking researchers to self-police becomes impossible. The honor system also relies on peer oversight. But AI-assisted work often looks indistinguishable from human-generated content. Peer reviewers, already overwhelmed and time-limited, can't realistically evaluate every submission for appropriate AI use. And we already know many reviewers are using AI tools themselves.
Although the lines are already blurry, AI companies are developing systems to further obscure them until they disappear. The next generation of AI tools won't just help with writing—they'll integrate seamlessly into research, offering real-time suggestions, automated analysis, and assistance at each step. The goal is to create tools so sophisticated and produce content so human-like that the distinction between human and AI contributions becomes zero. We're not just facing a current problem with fuzzy boundaries—we're heading toward a future where those boundaries won't exist.
AI Detection is Not the Answer
Faced with these limitations, we’re turning to AI detection tools. These promise to identify AI-generated text and restore some objective measure of authenticity.
This approach is doomed.
First, as many of us know, the tools don't work. AI detectors produce false positives, flagging human writing as AI-generated, and false negatives that miss obvious AI content. As AI systems improve, detection becomes even harder as detection will always lag behind generation technology. A major reason detection tools don’t work is they can't handle nuance. They might catch a fully AI-written paragraph, but what about text that's been AI-edited, AI-restructured, or AI-inspired? What about ideas generated through AI brainstorming? The sophisticated ways researchers use AI are invisible to detection algorithms.
Further, detection creates a surveillance culture. When we’re focused on catching violations rather than fostering responsible use, AI use is driven underground. Instead of encouraging thoughtful integration, we get secretive use and defensive researchers. Most importantly, detection misses the point. The goal shouldn't be eliminating AI from research. It should be to ensure AI use enhances rather than undermines research integrity.
What Actually Works
If honor systems and detection tools aren't the answer, what is? We need a fundamental shift from policing AI use to normalizing transparency about it.
We need to normalize disclosure. Researchers should document and disclose AI workflows just like any other methodological choice. This isn't confession—it's reproducibility¹. When someone uses AI to help analyze thousands of documents, that's methodology worth reporting. Discipline-specific guidelines would be a good step toward normalizing disclosure. A historian using AI to process archival materials has different integrity considerations than a microbiologist using AI to help write a methods section.
Focusing on education should be a priority. Researchers need training on AI capabilities, limitations, ethical implications (including environmental²), and best practices. Most people want to do the right thing—they just don't know what that is. Instead of trying to detect AI use after the fact, build submission systems that encourage disclosure. Make it easy to report AI assistance. Create templates for AI methodology sections.
We also need boundaries around human connection. Although AI is useful in many aspects of research, it shouldn't replace collaboration and mentorship. Using AI to draft emails to colleagues, generate responses in Slack or Basecamp, or simulate important face-to-face discussions undermines the relationships that make science meaningful and productive. The conversations, spontaneous remarks, and trust-building that happen through direct human interaction can't be automated away. AI should augment our capacity to do research, not substitute for the human connections that drive the science forward.
Let Me Be Clear: I Use the Tools
Writing is the part of my job I love the most! Yet, I've found AI has enhanced both my writing and my coding and just about anything else I use it for. Yes, there are hallucinations. But I'm the expert and I still maintain control. So, it's up to me (as it should be) to verify everything produced. And I often find errors! Yet, I struggle to see the downside of using these tools to improve the communication and pace of analysis of scientific data. Who wants to read a crappy paper? And don't try to tell me that no one has ever mis-coded something and come to the wrong conclusion. Mistakes happen in coding with humans too. The difference is that with AI assistance, I can iterate faster, explore more approaches, and communicate more clearly. The responsibility for accuracy, interpretation, and scientific validity remains entirely mine. That's how it should be.
Speaking of transparency, I used AI assistance in writing this post. I used AI tools to help brainstorm ideas, consider different perspectives, and improve clarity in several sections. I also used AI to help me decide what was superflous and condense repetition in previous drafts. The core ideas, arguments, experiences, and overall structure are my own, but AI helped me articulate these ideas more effectively. This is exactly the kind of disclosure that should become routine in academic work. (Note bene: I was using the em dash long before AI adopted it.)
The Stakes
The choices we make now will shape scientific knowledge production for decades. If we treat AI as either forbidden, we'll drive its use underground. If we ignore it, we miss opportunities to harness its potential to actually do good. The path forward requires prioritizing transparency and recognizing that research integrity in the AI age will look different from what came before. Researchers are already using AI. The question isn't whether this will happen—it's whether we'll create systems that encourage responsible use or drive it into the shadows.
¹The nature of generative AI means that identical prompts can produce different outputs, creating reproducibility challenges that traditional research methods don't face.
²AI tools come with significant environmental costs through their energy consumption and carbon footprint. Each query to large language models requires substantial computational resources. As we advocate for responsible AI use in research, we should also consider these environmental impacts as part of our ethical framework.