Ride the Speed Demon

Dirk Hovy
9 min read

Academia often mistakes effort for value. Everyone learns that the hard way when 90% of the work you put into your first project barely gets a mention in the final paper. But it also holds more broadly. Up until recently, writing a paper took a long time. After the initial idea, it included getting physical books from the library (possibly having them sent there from somewhere else) after locating them in a byzatine paper catalogue, writing things on a typewriter (including endless typos) or using a dictaphone (I vividly remember my father clearly enunciating one sentence at a time, saying “Full stop, next line,” pausing, rewinding, and checking, only to repeat the cycle). Maybe you sent the tape to someone else who typed it up. And that was just the writing part. Then you sent it out. Someone else retyped it, set it, and printed it. Reviews came by regular mail. Every step took time, and even small delays added up. Much of academic life often meant waiting for the system to catch up.

That world is gone. Good riddance, for the most part. Home computers sped writing up, so did the internet, printers, online databases, and better analysis software. Now you can do a first literature review in minutes. An agent can write code to explore and show the data. You can wrap up and document everything in a few hours. AI shows us how much of the effort in research was just administrative drag, not real value. It also shows that not every delay in academia meant rigor. That was often a sunk-cost fallacy. But clunky tools and slow processes could make mediocre ideas seem valuable. If it took a long time to make, it must be worth it, right? Not always. Sometimes a delay was just a delay.

Every academic generation faces a speed demon, some new tool that looks like it will make a previously essential human skill useless. But that never happens. Calculators did not end the need to know arithmetic. Search engines did not end the need for memory. PowerPoint (sadly) did not end bad talks. None of these tools did what people feared. Large language models have not solved language. They certainly have not solved science. But they are good at many parts of research that used to slow things down. Our speed demon just got a lot faster.

And yet, we largely still measure value by effort. Alexander Kustov argues that our academic reality has not yet caught up to the presence of the latest speed demon, the extent to which agentic AI changes research workflows, especially in quantitative and conceptual work (Kustov, 2026a). You were always told that science is a marathon, not a sprint. But if you are a junior academic today, you are seeing a speed demon. And speed has always been seductive in academia, especially when you are junior, and everyone in your network seems to be publishing, reviewing, applying, and somehow also finding time to have #blessed and #humbled takes on social media, and you need to make your mark. That can make it feel like you are racing that marathon at sprint speed against a robot that never gets tired. You might worry you will be obsolete before you even start. If the machine can already do so many parts of research faster than you, what are you here for? It is tempting to panic.

Don’t panic.

A few years ago, I went to a Google event for PhDs. Someone asked why Google wanted to hire PhD students to write code. The executive said they already had enough people for that. They wanted PhDs because they could find new problems. Maybe then they would write code to solve them. That was true then, and it is even more true now. Once you state a problem clearly, the rest is engineering. It can be hard and sometimes elegant, but once you have a problem and a way to solve it, it is still engineering. The real scientific challenge and value come before that.

The speed demon makes this difference even clearer. As it gets faster and easier to find answers, the human advantage is not in knowing more answers. It is in asking better questions. It is in seeing a gap and framing a question worth answering. As Paul Goldsmith-Pinkham writes (Goldsmith-Pinkham, 2026), AI speeds up the middle steps of research, not the start or the end. It can help with grunt work and answers. It cannot make a good question. It cannot decide if a result is worth believing.

In the short term, though, things may get worse before they get better. People will chase easy wins. Many will use AI to turn half-formed ideas into quick papers, old data into new submissions, and vague thoughts into polished nonsense. If you already thought some academic papers were just hot air swathed in confidence, the next few years will not bring much to change your mind. There will be more output, more noise, and more mess. Reviewers and journals will struggle with the onslaught of slop.

But that phase will not last. Once everyone has the same tools, speed will no longer set you apart. If everyone can produce decent code, clear writing, good figures, and basic analyses, those are just the basics. Nobody will get credit for doing things the hard way just to do them. A junior scholar with good ideas and strong AI workflows can now do work that used to require a small lab, or at least some very patient RAs. (That will, however, change what people expect and create a different type of competition.) If AI makes it easy to produce academic-looking text, institutions will have to look elsewhere for signs of quality than perceived effort. There needs to be a new differentiator. That will put more focus on ideas and originality. You will not stand out for running regressions. You will stand out for your idea. Bad news if your identity depended on showing effort the old way. Good news if you focus on ideas. That is why I think more AI will make good ideas matter more.

Dave Karpf, in his response to Kustov, disagrees that AI will replace social scientists, but he does think it will change the value of the traditional journal article (Karpf, 2026). Getting published in the usual venues will no longer necessarily show effort or the worth of an idea. What may be collapsing is not scholarship itself, but the journal article as the main unit of academic output. That sounds dramatic, but he has a point.

All this should be good news for junior academics. You are not in a PhD program to show you can handle manual coding, bad data, or strange publisher formatting rules. You are there to learn how to think clearly, see which problems matter, and then ask interesting questions about them. The speed demon can free you up to do just that.

So what can you do as a junior academic to avoid getting run over by the speed demon? First, learn the tools. Standing out from AI does not mean ignoring it. Quite the contrary! AI is a tool, and you need to know how to use it well. Use AI for the boring parts. Let it speed up the repetitive work. Let it help you test, compare, document, and debug. But you cannot just outsource all your work, sit back, and relax. You do risk losing your skills. You still need to learn the basics before you let AI do them for you. If you use a model to code before you know what code does, or to summarize papers before you can judge an argument, you are not saving time. You are skipping the training you need to supervise the tool later. You need the knowledge to catch mistakes, spot nonsense, and know when the AI is wrong. If the model writes your regression code, you should still know what the numbers mean. If it drafts your literature review, you should still know which papers it missed and why that matters. If it gives you a nice paragraph, you should still ask if it says anything relevant. Checking the work matters even more in a faster system. You can save a lot of time producing, but you will still have to invest some time on checks. Overall, you will still come out ahead on time, but it is not as much as it might first look.

More importantly, learn what not to outsource. Kustov, in a follow-up post (Kustov, 2026b), also points out something more hopeful: as AI gets better at standard synthesis and analysis, the relative value of things AI still cannot do rises. Fieldwork. Trust-building. Archival discovery. Original data collection. Physically being in a place at a certain time. Convincing a shy or nervous person to talk to you about their experience. A model cannot spend two years gaining access to a secretive community. It cannot sit in a dusty archive and realize that the key letter was filed under the wrong name. Basically, everything that involves physical presence and social competence. These are no longer just quaint leftovers from a pre-digital age but rather forms of academic scarcity (Togelius & Yannakakis, 2024).

As Hadley Wickham puts it, AI is like an amplifier, or a magnifying mirror: it makes you more of what you already are. AI can give you more time to ask better questions, or it can help you make mistakes faster. If you have a good idea, it helps you test it faster and makes you more productive. If you have a bad idea, AI can help you make a cleaner, more convincing version of it. If you are sloppy, it makes you sloppier, just faster. There is an old carpenter saying: you cannot build things as fast with hand tools as you can mess them up with a power tool. AI is an academic power tool, though it might also help you spot problems sooner and avoid wasting time. You decide what the mirror shows you.

There is one more reason to be hopeful. The speed demon will force academia to decide what it really values. We will reward what cannot be mass-produced: The skill to find a real problem and dig into it. Only now, you can do it much faster.

Sources