I have been publishing occasionally on the EA Forum for some time, often using it to explore ideas that I thought might be scientifically useful, especially around animal welfare, sentience, suffering, interspecific comparisons, and more recently the impact of AI on animals.

Recently, the Forum introduced automatic labels indicating how much of a post appears to have been written by AI: “Partly AI-written”, “Mostly AI-written”, or “~Entirely AI Written”. Still, I disagree with this new policy because I think it directs our attention toward the wrong variable. While the motivation is understandable—there is a lot of bad AI-generated material on the internet, and nobody wants the same for the Forum—perhaps my personal experience of several decades in science, including what happens backstage, can provide an unusual angle on this.

My interest is in the honest pursuit of truth and, whenever possible, in how knowledge can help us do good. At that level, I care very little, except perhaps as a matter of admiration, whether an argument was labored over for months by a dedicated and talented human, produced largely by an AI from an initial human idea, or developed through many rounds of interaction between the two. I care whether the argument is any good. More specifically, I care whether it helps us understand something that we did not understand before, whether its evidence is sound, whether its reasoning survives criticism, and, in the areas I work on, whether it can eventually help us reduce suffering. If AI allows us to do more of that, faster, I regard that as progress.

What exactly are we rewarding?

There are many things one might value in writing. One is craftsmanship. Some people enjoy knowing that a particular person selected every word. I understand that very well. In fact, this is one of the things I most admire in the writing of my favorite author, Jorge Luis Borges. Each word seems to me like a finely sculpted piece of a mosaic, fitted precisely into short pieces dense with beauty and layers of philosophical meaning. I also think authorship of this kind can matter greatly when we read a poem, a memoir, or a love letter. Another is interpersonal communication. In the discussion of the Forum's policy, one commenter put this position unusually clearly: what mattered to him was “genuine unfiltered human interaction”; the quality of the ideas was only part of what he wanted from the Forum. That seems to me a perfectly legitimate preference. It is simply not my main reason for reading scientific or analytical work.

If someone proposes a new way of estimating animal suffering, I want to know whether it works. If somebody finds a flaw in one of my hypotheses, I want to understand the flaw. If a new argument changes which animals we should prioritize, I care enormously about whether that argument is correct. Whether the sentences conveying that argument originated in neurons or GPUs is, to me, a secondary question. This seems particularly relevant from an effective-altruist perspective. Human intellectual labor is scarce. If we have developed machines capable of multiplying it, the interesting question is what useful intellectual work we can now accomplish, not how much human labor remains embodied in each paragraph.

Human intellectual production, seen from backstage

Human intellectual production is one of the great treasures of our existence. But having spent much of my life in science, I am not naive about it. I have encountered extraordinary scientists. I have also encountered an incredible amount of mediocrity. More troublingly, I have seen dishonesty, questionable attribution, selective use of evidence, and other behavior far removed from the idealized picture of the scientist patiently pursuing truth. I have more than once been a victim of plagiarism myself. This has left me with very little inclination to treat “human-written” as a certificate of quality.

Science is made by humans, and it reflects both the virtues and the vices of the rest of humanity. AI systems have their own failure modes, of course. But the conclusion I draw is not that one kind of intelligence should be trusted and the other distrusted. What matters is whether the claims are supported, the references are real, uncertainty is represented honestly, and the reasoning survives scrutiny.

The same applies to quality control. If AI makes it possible to generate thousands of mediocre essays at negligible cost, a forum has every reason to protect readers from them. But the problem is that they are mediocre, repetitive, or uninformative. Those are properties of the work itself. The filter should target those properties, not the provenance of the sentences. If AI-generated work is generally worse, good quality controls will filter much of it out anyway. And if it becomes better, penalizing it simply for being AI-generated would become increasingly difficult to justify.

AI has changed how I work

AI has been extraordinarily valuable to me. English is not my first language. I am also extremely busy with several ongoing projects. Before these tools existed, there were ideas that I simply did not have time to turn into publishable essays, or that took much longer to express adequately. That alone would have been enough to make the technology valuable to me. But my use of AI goes much further than grammar or editing.

I use it to challenge ideas, look for weaknesses, suggest connections I had not considered, restructure arguments, inspect assumptions, explore counterarguments, search through unfamiliar areas, and sometimes develop an initial thought much further than I could conveniently have done alone. I see no virtue in deliberately refusing that help.

On another level, this question is far from abstract for me. In my own work on animal welfare, I am deeply interested in whether AI models and agents can help us map affective states and welfare impacts at a scale and resolution that would previously have been impractical. When we began developing methods for this work, analyses at the scale we imagined appeared likely to require many years of human effort and resources we could not assume would be available. AI has changed that calculation. Some of what once looked prohibitively slow may now be achievable much sooner. To me, that is an opportunity worth exploring, provided the resulting claims remain open to the same scrutiny we should demand of any scientific work.

Conclusion

Science and intellectual life have always relied heavily on proxies for quality: credentials, institutional affiliation, journal prestige, fluency, confidence, citation counts, the reputation of the author. None is completely useless - but also none tells us whether a particular argument is true. If the origin of sentences becomes increasingly ambiguous, perhaps we will have to look harder at what we should have been examining all along: the argument, the evidence, and what follows from them. That seems to me a rather healthy development.

We are entering a period in which intellectual production is becoming dramatically cheaper and more abundant. Much of what gets produced will undoubtedly be rubbish. So was much of what humans produced before. But some of it will be excellent. Some may contain ideas that would otherwise never have been developed at all. It would be a pity to waste that opportunity because we mistook the provenance of prose for a measure of intellectual merit.

And, in case the Forum’s detector leaves any possible doubt: this essay was produced fiercely, shamelessly, and very productively with the assistance of AI. If knowing that makes the argument less persuasive, then perhaps we have identified the disagreement.