Recently Notion, a service for creating and managing shared documents, announced Notion AI. This new wonder of technology promises to create documents for users. For example blog posts:

Write the page title. Notion AI will take care of the rest.

My first reaction was what nonsense!. But thinking it over I think Notion seems to have a good understanding of the types of documents stored on its servers: outdated project documentations that nobody ever looks at again (or even finds again), pointless quarterly plans that are just created to be created without helping to coordinate or make management decisions, meeting notes that everybody insists on recording, but forgets about their existance or content the moment the document is written down. I am not saying that project documentations, quarterly plans or meeting notes are generally useless. But I have seen many a team forget that such texts are tools rather than ends in themselves and that the time investment to create them should be balanced with their benefits.

In the domain of science we also find lots of texts that nobody cares about. Scientific publications (also refered to as papers) are one of them. Originally papers were a means of communicating research results, but recently they have become a kind of decoration on the CVs. Papers are not written any more, they are produced. So why not just have machines do it? This idea is not even new: SCIgen has generated papers in the area of Computer Science. As of 2021, 243 such papers have been published [1]. This is not a large number compared to the number of papers published per year, but it is a start.

When we generate the papers, it would of course be absurd to have them reviewed by humans, so let's also generate the reviews. In fact, even if we stay with human-produced papers for a while, we should still consider to use computers for generating the reviews (as we all know computers are so much more objective than humans, so no more feelings in the review process). The review system is in the process of collapsing anyway. When I was still serving as an editor for scientific journals and conferences, it was getting harder and harder to find people to write reviews. At some point you stop bothering the handful of friends of whom you know they will provide decent reviews, because you want them to stay your friends. Other editors seem to have the same problem. Having submitted a paper to a good journal (this year), we got an apologetic e-mail that our paper had not yet been reviewed after three months because the editor could not find reviewers. So to overcome this reviewing bottleneck, we should just let computers do the job, so the paper counting on CVs is not being stifled any more by the review process.

Analogous to publications, we encounter the same reviewing bottleneck for grant applications. And, of course, writing the proposals is another time-consuming exercise. Computers cannot only create the text, they could split the work into arbitrary work packages and create fancy gantt charts. Honestly, in this respect, computers can hardly invent more fantastic promises and pie-in-the-sky time plans than you currently find in human-made grant proposals.

When grant proposals and their reviews are computer-generated for everyone, the grant assignment will result in a random assignment process. Not even this idea is new. Funding agencies are already experimenting with random grant acceptance [2]. The only difference would be to pacify traditionalists who feel uncomfortable assigning money without the proper process.

Using the time that is currently spent on producing papers that nobody reads and grant proposals that have no connection to the research that will be finally done, and turning it into real research time (i.e. thinking, observing, experimenting, trying out procedures even though one does not know whether they will work) could be the crucial factor to save the scientific system from collapse by ridiculousness. The decisions of who will make a career will be rather random, but then, most of the time in history, becoming a scientist was based on the random attribution of wealth and social status. But I would rather give a few randomly selected individuals the freedom to advance science that restraining everyone in the system as we have it now.

As much as I like my dream about getting real science done when computers take over document production, I doubt that it will become a reality. Administrators will get wind of the computational support and start to withdraw any funding that is left to scientific institutions so that the few remaining scientists will be busy teaching (which we could also think about replacing by computers) or fulfilling ever more administrative exercises. At least one thought is gratifying: the AI prophets distributing futuristic scenarios of AI taking over any kind of job will have succeeded in killing off their own positions.