When I was studying physics many years ago at the Universidad Nacional Autonoma de Mexico, I learned to program in Fortran, a language whose structure still preserves the logic of the punch cards once fed into the first generation of digital computers many, many years earlier. When I was taught statistics, I was shown how to consult the tables at the back of textbooks (yes, we used books then) to identify the significance of our laboratory tests. Fortunately, every time we measured gravity, the number came back (sort of) the same.

These ways of engaging with knowledge, and with the production of claims about the world, are distant from many of our students today. With the click of a button, they can access software that was unimaginable when I was a student, running complex statistical tests with a few clicks, and needed at worst some basic familiarity with Excel. In many substantive ways, their relation to the production of some kinds of knowledge is today mostly automated.

The production of the kind of knowledge scientific we teach our students and hopefully produce ourselves depends on understanding the methods, techniques, and paths through which we claims about the world are crafted. The value of scientific knowledge rests, in large part, on the collective forms of verification, testing, and reliability assessment that communities have build around the production of novel claims. This is, in many ways, the secret sauce of science: the infrastructure of mutual observation and review that makes claims travel throughout the world.

This is, for example, the logic of peer review and, in many ways, the logic of Kuhnian paradigms. It is also the logic of scientific training, where people learn through exemplars and by tinkering with systems to learn about the limits, tolerances, and interconnections of their components, leading to knowledge claims that mostly remain robust to variations in who conducts a study and where it is done.

Why am I bringing this up? Discussions about the use of artificial intelligence in research have become highly contentious and equally widespread. Part of the debate (which many of us are involved in) concerns the intractability of AI models: unlike a statistical regression, for example, we cannot know what went into the production of a specific output as generated by an LLM; there is no standardized, collective infrastructure of trust and verification for the kinds of objects that LLMs produce every time we feed them a prompt. This is added to established concerns about hallucinations; the creation of fake; the citation of materials that differ from those a human researcher would identify as relevant; the scope of an analysis; and the originality of thought (an LLM cannot break a paradigm).

There are many reasons why using artificial intelligence in research may be a dangerous way of diluting knowledge claims, or even of producing false claims dressed up as real knowledge. Yet the fact remains that scholars across the world use AI in a range of applications, all the time. I have seen this as a peer reviewer, when papers that score highly on AI detection systems have arrived on my desk, and as a professor, in research papers submitted by students that contain a clear component of AI-generated material.

The question, then, is what to do. How might we resolve this conundrum without necessarily assuming we need to reach a point of collective disinvestment that make AI disappear entirely from the world of science?

At its core, the practical issue of epistemic concern is verification, not the use of AI in and of itself. When we feed data into a system like Stata, particularly for complex statistical analyses, we do not assess the underlying code line by line, checking every character that was used to conduct a given analysis to guarantee that the software package does what we want and what it should. Instead, we trust that a system of verification, an infrastructure of revision and assessment, has produced code that performs adequately under certain well-known and well-bounded conditions.

With artificial intelligence this is far more complex. Pinning down what LLMs are actually doing is not the same as going through Stata code. We do not know specifically what went into their training, for example, nor the weights applied to produce a given result, nor many other variables relevant to the labor of verification.

What we do know, however, are the claims themselves, however they were produced, and how those claims can be assessed in relation to existing stocks of knowledge by an expert analyst. Ultimately, whether a claim in a paper was produced by or with the help of AI matters less than whether an expert, exercising deliberative judgment and applying clear tests of validity, robustness, and so forth, assessed that claim to be correct. If a claim produced by a machine is correct, the fact that a machine produced it is largely immaterial; what matters is whether the claims and process used to create them have been verified through some collectively accepted process that is transparent, transportable, and intelligible to peers.

What does this entail? It means that the limits and conditions of applicability for AI in research (i.e. what’s acceptable and what’s not) depend on whether systems of verification exist. If research pipelines explicitly incorporate AI in the production of novel findings, and those pipelines include tractable, transparent, and transportable forms of verification for the claims made and the analyses conducted, those claims carry, for all intents and purposes, close to the same epistemic standing as the claims found in a traditional paper. This logic also establishes boundaries for application. Producing textual content through AI, for example, may be less defensible than using it to manipulate data or to write code, given that established forms of verification exist for both (you can validate the data, and check the code against a baseline of expertise); generated prose cannot be tested in the same way..

What emerges from focusing on verification, then, is a mapping of the space of possible applications of AI that might genuinely identify opportunities for innovation in research pipelines (produce savings and efficiencies in the research process) and risky zones where AI is bound to produce less verifiable claims.

This does not settle the debate, of course, but it offers a way of thinking about how we might evaluate the use of AI in research that is grounded on established practices without the need of reinventing the wheel. If a scholar discloses the use of AI, the next step must be to disclose the steps taken to verify the data, the code, the methods. This mirrors, for example, how we traditionally approach such things as topic models in research where, upon a given output, there are established protocols for arguing whether it’s relevant or not.

This shift toward verification also suggests where investments in AI might make sense and where they do not. Investing in the production of content is perhaps not the smartest strategy for contributing to science. For many other tasks, from producing code or conducting some forms of data analysis, AI may indeed introduce efficiencies that the community should discuss. (Translation belongs among these examples, as it could reduce the costs that researchers whose first language is not English face in publishing in English, though at the cost of their own voices as authors).

This is my modest proposal and invitation, to think about how to evaluate and declare the use of AI across different tasks within a research project, while maintaining the integrity of the claims we make when we write.