At work, we tested ChatGPT Deep Research to produce a research report. For the prompt, we used the same prompt that I recently followed to produce my own research report. The prompt involved a detailed review of many specific questions relating to European aquaculture and fish markets. My own report is still unpublished, so ChatGPT does not have access to it. I didn’t refine the prompt at all after seeing ChatGPT’s report.

(I haven’t linked to either report here. Mine is still not publicly available, so it’d be hard to compare. In this post, I just want to write about my conclusions from this exercise.)

TL;DR:

  • This technology is freaking insane. The report took 11 minutes and it seems like it’s at the level of a mid-quality professional research analyst. Compare this to 4 weeks for my own report.
  • There are many reasons to believe it will continue to improve significantly, and many of the limitations seem readily fixable.
  • I think that research will look very, very different in a decade’s time.
  • I’m going to explore alternative careers. I still think research is worth doing, and research skills are still worth having; if anything, the ability to distinguish truth from not-truth will become more important. But this new form of research could involve much heavier use of generative AI, and my heart isn’t in it. I’d rather spend my time on other activities.

Detailed thoughts:

I read through the report generated by ChatGPT Deep Research itself and in terms of accuracy, it seems comparable to a typical report by a colleague I might be asked to review - it seems mostly correct, with a few details that I think are wrong but are really arguable either way, plus it has some insights I didn’t include in mine.

It seems like generative AI is getting quite close to doing what we do - secondary research involving flexible judgment and aggregating a variety of sources. The bottleneck is human judgment stuff (e.g. some guy on the ground in Chile just told me XYZ, and I need to integrate this into my decision) and access to data (e.g. the technical and legal hurdles which are currently preventing it from scraping and using all scientific papers but could be overcome very easily). Of course, interacting with the physical world will remain a challenge for AI. But it’s rare that we do original research at my current organisation - in almost all cases, we’re doing secondary research using publicly available information (it’s different for primary researchers/experimental scientists). I think it’s already better than the lower-quality researchers who are currently working in research globally (not animal advocacy necessarily). Also need to take into account the fact that this took 11 minutes, compared to my systematic review report which took 4 weeks (and I’m pretty sure even that is more efficient than the average researcher).

It seems to me that the next few iterations of this Deep Research functionality will overtake human researchers very quickly. I remember playing around with GPT back when nerds knew about it but before it was easily publicly accessible (that is, pre-ChatGPT release). I think I was playing around with it when I was in the airport to EAGx Australia 2022. So around September 2022, or 2.5 years ago. The gap in the quality of output between that iteration and ChatGPT Deep Research is enormous. If you had shown me ChatGPT Deep Research then and said “Conditional on this being possible within the next 3 years, how likely is it that GPT will be better than any human at research within a further 5 years or whatever”, I probably would have assigned that outcome a high likelihood. And that’s the world we’re now in.

So there are still a few bottlenecks and shortcomings. But I’m not aware of any specific reason why those bottlenecks and shortcomings won’t be addressed - beyond the next iterations of GPT itself and then overcoming the technical/legal hurdles, it seems more likely to me than not that LLMs will outpace human researchers within like 3 years.

Which raises the question of what the future of research will look like. Even a hyperintelligent GPT wouldn’t be something we could simply defer to in all decisions at my organisation (Animal Ask), for instance. We’d need at least a single human decision maker. But maybe the model of a typical organisation is more like “1 nerd + GPT” rather than “4 nerds”.

Update 2025-03-21

A colleague asked me a couple of further questions, which are relevant here.

I noticed that many of the sources in the Deep Research response […] were not from peer-reviewed papers. How much of a limitation do you think that poses for producing high-quality research?

Yes, I think this is currently a major limitation. However, I think that it is very easily solvable in principle - it seems like there might be some very minor technical and legal barriers preventing this, but I think it’ll be solved quite soon (next few years). e.g. technologically it could be as simple as scraping all academic databases and then setting up some pipeline such that Deep Research picks the best sources from those academic databases, reads the full texts (from sci-hub or whatever), then bases its answers primarily on that information.

There are small technical and legal hurdles in scraping all academic databases and obtaining/parsing full-text documents that are currently not solved, but could easily be solved by either a) an AI company signing a deal with academic publishers to scrape data (as AI companies have already done with news publishers) or b) a clever nerd who doesn’t care about the law publishing an open-source plug-in to give some LLM access to all this database and full-text information (there are already such plug-ins for Zotero, for example, so it’d only be a matter of time before a nerd on the internet makes a similar plug-in for LLMs).

So I think this is more something that will be solved in the next year or two, basically whenever people get around to it - there is no conceptual advance required in order to do this.

In what ways do you expect research to be different in the coming years? As in, what will a human’s job look like? Asking because I’m also exploring which skills are likely to become more relevant as AI improves.

The million-dollar question! hahaha

These are just my guesses. Not sure if you should weight my views on this very highly.

My suspicion is that:

  • Secondary research will involve fewer researchers, using a mix of existing research skills and prompt engineering/AI tools, having a higher per capita research output. The mix of existing skills vs AI skills depends on how autonomous and “good at making conceptual leaps” AI becomes. Two years ago, it was a 95/5 mix in favour of classic research skills. Right now, with the current iteration of Deep Research, it’s more like a 70/30 mix. With the next iteration of Deep Research (e.g. what I’ve described above), we might see 50/50.
  • Primary research is similar, but I expect will maintain a consistently higher share of “classic research skills” than secondary research. e.g. we need somebody to pipette liquid into test tubes, physically go to supermarkets to survey consumers, design experiments, think of hypotheses, and be absolutely fastidious about working through logic from the hypothesis to the experimental design. I think most of these skills can also be automated, but I think there’ll be a lag compared to secondary research.

An important thing here is that research will become more efficient - i.e. we get more units of research per unit of human labour. This might mean society hires fewer researchers (as we can achieve the same with fewer people) or it might mean that society hires more researchers (as there are greater financial and social returns to a given unit of investment in research).

I’m thinking about the next five years or so, and not much further. Beyond that, it gets hard to predict. e.g. we could end up with a fully automated pipeline where all elements of the scientific process are performed by AI (“P.A.S.T.A.” as 80,000 Hours has called it). Or, AI advances could plateau before then, and we’d end up with primary and secondary researchers automating large components of the scientific process but remaining present for some aspects. Part of this is technological (what the actual limits on AI possibilities are, which nobody knows) but part is also social (how much society is comfortable handing over aspects of the scientific process and/or policy-making to machines, which nobody knows).

oh and this article is relevant, I agree with these views, though this is more longer time-frame than I normally think about: A Taxonomy of Jobs Deeply Resistant to TAI Automation - EA Forum