A Vision for the Future of an AI-Integrated Research Ecosystem

Imagine the world of science as a giant, bustling library where researchers are the authors, editors, and librarians. For a long time, this library ran on a simple rule: humans write the books, humans check the facts, and humans decide what gets published. But recently, a new kind of helper has arrived: Generative AI. Think of it as a super-fast, super-smart robot assistant that can write drafts, find facts, and even summarize entire books in seconds. While this sounds like a dream come true for getting things done faster, it has created a bit of a panic in the library. Everyone is asking: "If a robot helped write the book, who is the real author? Is the story true? And if robots start checking the books too, will we end up with a library full of nonsense?" This paper steps into that chaos to ask a bigger question: instead of just trying to catch people using robots, how do we redesign the whole library so that trust is built-in from the start?

The authors, Ryan E. Dougherty and Natalie Kiesler, are not just guessing; they are looking at the cracks in the current system. They argue that the current focus on "policing" AI—making sure people admit when they used it—is like trying to fix a leaking roof by just putting a bucket under the drip. It's necessary, but it doesn't stop the rain. They suggest that the real problem is that the whole way we share science (papers, reviews, and rewards) is already struggling, and AI is just making the pressure worse.

To help us understand where we might be heading, the authors invite us on a trip to the year 2036 to look at two very different futures.

The Bright Future
Imagine a researcher in 2036 who discovers something amazing about how to teach coding. In this world, they don't spend a year fighting through paperwork. Instead, AI tools act like a super-organized team of assistants. They instantly check the math, verify the data, and connect the new discovery to everything else we know. A human expert then steps in to ask the big questions: "Is this actually important? Does it make sense?" In this future, the "paper" isn't just a static document; it's a living, breathing object connected to the raw data and code. Knowledge flows faster, is easier to find, and is more trustworthy because the AI handles the boring, error-prone work, leaving humans to focus on the big picture.

The Dark Future
Now, imagine a different 2036. Here, the library is drowning. AI is churning out millions of research papers and reviews every day. No human can read them all, so they rely on other AI to summarize the summaries. The volume of "science" is huge, but nobody knows what is real and what is fake. The pressure to publish is insane, and the trust in the system has collapsed. In this scenario, the challenge isn't finding new discoveries; it's figuring out which ones to believe. The humans are exhausted, trying to keep the system from falling apart, but the noise is too loud to hear the truth.

The authors use these two stories to highlight a fundamental tension: the same technology that could make science transparent could also make it a mess of lies if we don't change how we do things. They don't claim one of these futures is guaranteed; instead, they use them to show that we have a choice to make right now.

The Four Big Questions
To navigate this choice, the authors suggest we need to rethink four main things:

1. Are "papers" still the right way to share ideas?
   Currently, a research paper is like a compressed snapshot of a project. It assumes the reader can't see the raw data or the code behind it. But today, we have the tools to share everything. The authors suggest that the "paper" should stop being the main event. Instead, the real research object should be the data and code itself, which should be easy to find and reuse (a concept they call "FAIR"). The paper would just be one way to tell the story of that data. If we do this, it becomes clear where AI helped and where humans did the heavy lifting.

2. What should a "review" look like?
   Right now, peer review (where experts check a paper before it's published) is broken. Studies show that different groups of reviewers often disagree wildly on the same papers. Sometimes, reviews are just random guesses. Plus, reviewers are often anonymous and unpaid, which leads to careless or even AI-generated reviews that slip through. The authors argue that a review shouldn't just be a verdict (pass or fail); it should be a valuable contribution in itself. We need to treat reviews like real scholarship: credit the reviewers, check their work, and make sure they are actually helping, not just checking boxes.

3. Do we still need human reviewers?
   The authors suggest that AI is great at the boring parts of reviewing, like checking if a citation is real or if the math adds up. But humans are needed for the "soul" of the review: judging if the idea is important, fair, and creative. If AI does the verification, maybe we don't need as many humans doing the same tedious checks. However, we absolutely need humans to spot bias and decide what matters. The danger is that if we let AI do everything, we might lose the human judgment that keeps science honest.

4. Whose interests are we serving?
   Everyone in the system wants different things: authors want to publish fast, reviewers want less work, and editors want high-quality papers. AI makes it easy to automate the annoying parts of everyone's job, but if we all just automate our way out of the hard work, the whole system might speed up but learn nothing. The authors argue that we need to change the rules so that doing good, honest science is what gets rewarded, not just doing things quickly.

Three Big Challenges
The authors conclude that we can't just "fix" this with a single tool. They propose three big challenges the community needs to tackle together:

• Calibrated Trust: We need a way to measure how reliable AI tools are. Right now, saying "AI helped" is just a claim. We need standard tests to see how often an AI makes up facts or misses errors, so we know when to trust it.
• Provenance (The Paper Trail): We need systems that automatically record exactly what tools were used and how. Imagine a digital receipt that comes with every paper, showing exactly where AI was used and where humans stepped in. This needs to be built into the software researchers use.
• Integrity in a Synthetic World: When AI can write anything, we can't just rely on "detecting" AI to stop misuse. That's a losing battle. Instead, we need to build a culture where the value of the work is clear, regardless of who or what wrote it.

The Path Forward
The paper doesn't offer a magic wand. Instead, it suggests a shift in mindset. We need to stop trying to police AI and start building the infrastructure that makes trustworthy science the default. This means valuing data as much as papers, treating reviews as real work, and creating new ways to measure trust.

The authors end with a call to action for the whole community: let's have a big conversation about what we actually value as scientists. Do we want speed, or do we want truth? Do we want more papers, or better ideas? The technology will make us faster either way, but whether it makes us better depends on the choices we make today. The future of science isn't just about what the robots can do; it's about what we, as humans, decide to build with them.