At the beginning of February, OpenAi, the world's most famous artificial intelligence company, published deep research, which was designed to conduct in-depth, multi-step research “. With a few strokes of a keyboard, the tool can produce a paper on any subject in minutes Many academics are eager to do it. Level, you can publish papers that you 'wrote in one day', said Kevin Bryan of the University of Toronto. “I am thinking of the quality that is comparable to a good study level assistant, and send that person away for a week or two With a task, “said Tyler Cowen of George Mason University, an economist with cult-like status in Silicon Valley.
Do you have to disable $ 200 a month for deep research? Mr Cowen has hyped rages in the past, as he did with Web3 and Clubhouse, a once popular social-media network. On the other hand, if deep research approaches a form of artificial super intelligence, as many believe, then $ 2,400 a year is the biggest bargain in the history of the world. To help you decide, your columnist has made the tires of the new model. How good is a research assistant deep research, for economists and others?
The obvious conclusions first. Deep research is unable to conduct primary research, from organizing polls in Peru to getting a sense of the body language of a chief executive whose company you could fail. Neither can it brew a coffee, making it a bad replacement for a human assistant. Another complaint is that the output of deep research is almost always prose, even if you ask to be more lively. On the other hand, most people were never good writers anyway, so it will hardly be careful whether their AI assistant is a bit boring.
However, use deep research as an assistant for a while, and three more important issues come to the fore: “data creativity”, the “tyranny of the majority” and “intellectual shortcuts”. Start with data creativity. The OpenAi model can ask for clear questions – “What was the unemployment rate of France in 2023?” – without breaking the step. It can make marginal more complex questions – “Tell me the average unemployment rate in 2023 for France, Germany and Italy, weighed by population” – with ease.
However, when it comes to data questions that require more creativity, the model is struggling. It is wrongly estimated the average amount that an American household under the leadership of a 25 to 34-year-old in 2021 spent whiskey, although anyone who is familiar with the Bureau or Labor Statistics data in the exact answer ($ 20) in a A few seconds. It cannot accurately tell you which part of the British companies currently uses AI, although the Statistics Office produces a regular estimate. The model has even more difficulty with more complex questions, including those of the analysis of source data produced by statistical agencies. Human assistants retain a lead for such questions.
The second problem is the tyranny of the majority. Deep research is trained on a huge range of public data. For many tasks this is a plus. It is amazingly good at producing detailed, produced summaries. Mr. Cowen asked the POT to produce a paper of ten pages in which the rental theory of David Ricardo was explained. The export would be a respectable addition to each textbook.
Nevertheless, the enormous amount of content that is used to train the model creates an intellectual problem. Deep research is usually based on ideas that are often discussed or published, instead of the best things. Information volume Tyrannises Information quality. It happens with statistics: deep research is susceptible to consulting sources that are easily available (such as newspapers), instead of better data that may be behind a payment wall or can be found more difficult.
Something similar happens with ideas. Consider the question – much discussed by economists – about whether American income inequality is increasing. Unless asked to do differently, the model faints that inequality has risen since the 1960s (as the conventional wisdom is) instead of staying flat or only increasing a little (the image of many experts). Or consider the true meaning of the “invisible hand” of Adam Smith, the fundamental idea in the economy. In a paper published in 1994, Emma Rothschild from Harvard University demolished the idea that Smith used the term to refer to the benefits of free markets. Is aware of Mrs Rothschild's investigation, but nevertheless repeats the popular misconception. In other words, those who use deep research as an assistant risk about the consensus display, not that of the cognoscenti. Creativity and thought, from public intellectuals to investors.
The idiot fall
A third problem with the use of deep research as an assistant is the most serious. It is no problem with the model itself, but how it is used. Despite, you notice that you take intellectual shortcuts. Paul Graham, an investor from Silicon Valley, has noted that AI models, by offering to do people to do people for them, to make the risk of making them stupid. “Writing is thinking,” he said. “There is even some kind of thinking that can only be done by writing.” The same applies to research. For many jobs, research is to think: noticing opponents and gaps in conventional wisdom. The risk of outsourcing all your research to a Supergenius assistant is that you reduce the number of options for getting your best ideas.
Over time, OpenAI can iron its technical problems. At a certain point, deep research can also come up with great ideas, making it an assistant to the main investigator. Use deep research up to that time, even for $ 200 a month. Just don't expect the research assistants to replace soon. And make sure it doesn't get stupid.
© 2025, The Economist Newspaper Limited. All rights reserved. From the economist, published under license. The original content can be found on www.economist.com
