Professional Documents
Culture Documents
VITOR MIRANDA/ALAMY
produce detailed and informative answers to
complex user questions.
But using conversational AI for specialized
research is likely to introduce inaccuracies,
bias and plagiarism. We presented ChatGPT
A chatbot called ChatGPT can help to write text for essays, scientific abstracts and more. with a series of questions and assignments
that required an in-depth understanding of
for research
experience relapse after treatment?’, it gen-
erated an overly general text arguing that
treatment effects are typically long-lasting.
However, numerous high-quality studies show
that treatment effects wane and that the risk
Eva A. M. van Dis, Johan Bollen, Robert van Rooij, of relapse ranges from 29% to 51% in the first
Willem Zuidema & Claudi L. Bockting year after treatment completion2–4. Repeat-
ing the same query generated a more detailed
and accurate answer (see Supplementary
Conversational AI is a caused excitement and controversy because it information, Figs S1 and S2).
is one of the first models that can convincingly Next, we asked ChatGPT to summarize a
game-changer for science. converse with its users in English and other systematic review that two of us authored
Here’s how to respond. languages on a wide range of topics. It is free, in JAMA Psychiatry5 on the effectiveness of
easy to use and continues to learn. cognitive behavioural therapy (CBT) for
This technology has far-reaching conse- anxiety-related disorders. ChatGPT fabricated
quences for science and society. Researchers a convincing response that contained several
and others have already used ChatGPT and factual errors, misrepresentations and wrong
S
other large language models to write essays data (see Supplementary information, Fig. S3).
and talks, summarize literature, draft and For example, it said the review was based on
ince a chatbot called ChatGPT was improve papers, as well as identify research 46 studies (it was actually based on 69) and,
released late last year, it has become gaps and write computer code, including sta- more worryingly, it exaggerated the effective-
apparent that this type of artificial tistical analyses. Soon this technology will ness of CBT.
intelligence (AI) technology will have evolve to the point that it can design exper- Such errors could be due to an absence of
huge implications on the way in which iments, write and complete manuscripts, the relevant articles in ChatGPT’s training set,
researchers work. conduct peer review and support editorial a failure to distil the relevant information or
ChatGPT is a large language model (LLM), a decisions to accept or reject manuscripts. being unable to distinguish between credible
machine-learning system that autonomously Conversational AI is likely to revolutionize and less-credible sources. It seems that the
learns from data and can produce sophisti- research practices and publishing, creating same biases that often lead humans astray,
cated and seemingly intelligent writing after both opportunities and concerns. It might such as availability, selection and confirma-
training on a massive data set of text. It is the accelerate the innovation process, shorten tion biases, are reproduced and often even
latest in a series of such models released by time-to-publication and, by helping people amplified in conversational AI6.
OpenAI, an AI company in San Francisco, to write fluently, make science more equi- Researchers who use ChatGPT risk being
California, and by other firms. ChatGPT has table and increase the diversity of scientific misled by false or biased information, and