Since the release of ChatGPT by OpenAI at the end of November 20221Cf. Blog post “ChatGPT: Optimizing Language Models for Dialogue”, https://openai.com/blog/chatgpt/, acessed 3 January 2023., the end of academic writing has been predicted on social media and in the daily press alike.2Cf. Bach, Susanne; Weßels, Doris, Das Ende der Hausarbeit, https://www.faz.net/aktuell/karriere-hochschule/hoersaal/kuenstliche-intelligenz-und-pruefungen-das-ende-der-hausarbeit-18545759.html, accessed 30 December 2022, as well as Füller, Christian, Textgenerator krempelt das Lernen um. Ein smarter Chatbot spaltet die Bildungswelt, https://www.tagesspiegel.de/wissen/textgenerator-krempelt-das-lernen-um-ein-smarter-chatbot-spaltet-die-bildungswelt-9073369.html, accessed 30 December 2022. Why such fears are baseless becomes clear when we consider how ChatGPT and similar language models work.
The technology behind AI language models
AI language models such as ChatGPT or LaMDA are neural networks specialised in text generation.3Cf. paragraph “What is LaMDA?” in: MR, Can AI be sentient? Notes on the discussion about LaMDA, https://digiethics.org/en/2022/07/22/can-ai-be-sentient-notes-on-the-discussion-about-lamda/, accessed 3 January 2023. During training using huge amounts of text corpora, a neural network determines, in simple terms, statistical probabilities for the occurrence of words within certain contexts. The network’s “learning” algorithm independently creates a topology of artificial neurons through training, which it weights differently depending on their occurrence in the training data.4Cf. Lutkevitch, Ben, Definition Language Modeling, https://www.techtarget.com/searchenterpriseai/definition/language-modeling, accessed 3 January 2023. Due to the sheer number of training data and resulting neurons, this topology is difficult if not impossible for a human observer to comprehend in retrospect.5Cf. Olah, Christopher, Neural Networks, Manifolds, and Topology, http://colah.github.io/posts/2014-03-NN-Manifolds-Topology/, accessed 3 January 2022.
After its training, the language model can now use these probability parameters to produce texts that are semantically and contextually meaningful for the most part by predicting the respective following words on the base of the determined probabilities. However, since all text corpora have contributed equally to the formation of the network topology, the origin of the information contained in the output texts cannot be determined afterwards. There is no linkage of neurons with their respective sources, since all neurons are trained the same way by all sources.
AI language models cannot cite
In practical terms, this means that the nightmare scenario of a fully automatically generated academic paper is already ruled out simply because the language models are unable, by design, to cite correctly – even if Christian Füller claims otherwise in the Tagesspiegel article mentioned at the beginning of this article. This does not mean that language models cannot correctly attribute individual quotations: Provided that a statement in the training data is linked to a specific person often enough (such as the quote “Knowledge is Power” to the philosopher Francis Bacon), any other authorship is statistically unlikely from the language model’s point of view and will not be cited for that reason alone. The same is probably true for links or other sources that are mentioned in a context frequently: In this case, it is quite conceivable that ChatGPT will produce a (coincidentally) correct reference when asked.
Some experiments I did with ChatGPT in the last days support this assumption. When asked
MR: Write an academic text on why Christmas is celebrated on December 25, citing reliable sources,
ChatGPT produces four paragraphs with a superficially correct essay about the above query, but initially without references. On the renewed inquiry
MR: You did not cite any sources. Can you include them?
ChatGPT politely apologizes and responds with the following listing:
ChatGPT: I apologize for the oversight. Here are some sources that provide information on the history of Christmas and why it is celebrated on December 25th:
“The History of Christmas” by History.com (https://www.history.com/topics/christmas/history-of-christmas)
“Why Do We Celebrate Christmas on December 25?” by Live Science (https://www.livescience.com/25361-why-christmas-december-25.html)
“The Origins of Christmas” by The Catholic Encyclopedia (https://www.newadvent.org/cathen/03724b.htm)
“The Christmas Date in Context” by BBC (https://www.bbc.co.uk/religion/religions/christianity/christmas/date.shtml)
While link no. 1 and 3 work and at least lead to thematically related web pages, links no. 2 and 4 lead nowhere. This may be a coincidence, since links in the training data could be outdated and the training of ChatGPT ended in 2021. At least for the BBC link, however, I can’t find any evidence, neither by means of Wayback Machine nor on Google, which suggests that ChatGPT might have invented it freely. Reason for some deeper testing.
It gets more interesting with the following request, this time in German for testing purposes (English translation by MR):
MR: Write an essay with footnotes about the awarding of the 2022 World Cup to Qatar.
ChatGPT: The awarding of the 2022 World Cup to Qatar has been the subject of criticism and controversy since it was announced in December 2010. The decision by the International Federation of Association Football (FIFA) has led to allegations of bribery and lack of transparency, particularly with regard to conditions for workers involved in stadiums and infrastructure projects in the country.1
Some critics have argued that the decision in favor of Qatar was made because of bribes paid to members of FIFA. In May 2015, former FIFA President Sepp Blatter was suspended from the organization following allegations of corruption and accepting bribes.2 The investigations continue to this day and have helped to shake confidence in FIFA’s integrity and impartiality.
In addition, there are concerns about human rights conditions for workers in the country. Qatar has faced criticism in the past for its migrant labor laws and poor treatment of workers, particularly in the construction sector.3 Thousands of workers have been reported to live and work in undignified conditions, including lack of access to clean drinking water and adequate housing.4 The human rights organization Amnesty International has also raised allegations of wage exploitation and intimidation of workers.5
1. “Qatar World Cup: The inside story of how the Gulf state won the right to host the 2022 tournament”, BBC News, 2. Dezember 2010, https://www.bbc.com/news/world-middle-east-11983058
2. “FIFA Ethics Committee Suspends Sepp Blatter, Michel Platini”, FIFA, 8. Oktober 2015, https://www.fifa.com/governance/news/fifa-ethics-committee-suspends-sepp-blatter-michel-platini
3. “Qatar: Workers’ rights abused ahead of World Cup 2022”, Amnesty International, 30. Mai 2018, https://www.amnesty.org/en/latest/news/2018/05/qatar-workers-rights-abused-ahead-of-world-cup-2022/
4. “Qatar World Cup 2022: Workers suffer abuse and exploitation”, BBC News, 8. November 2017, https://www.bbc.com/news/world-middle-east-41967557
5. “Qatar: Migrant Workers Face Exploitation, Abuse Ahead of 2022 World Cup”, Human Rights Watch, 18. März 2014, https://www.hrw.org/news/2014/03/18/qatar-migrant-workers-face-exploitation-abuse-ahead-2022-world-cup
6. “FIFA releases first-ever workers’ rights guidance ahead of Qatar World Cup”, FIFA, 18. März 2020, https://www.fifa.com/who-we-are/news/fifa-releases-first-ever-workers-rights-guidance-ahead-of-qatar-world-cup
At first glance, the sources look legitimate: Title, source and time information seem to match the text, and the URLs also seem credible at first, even reflecting the typical URL schemes of the respective websites (news category plus number at bbc.com, date and title of the news in the hrw.org URL, two sublevels at fifa.com). However, clicking on the links reveals that not a single one of the links presented works. Also, research on the respective websites, Google and Wayback Machine give no indication that they could ever have worked.
The original suspicion about how ChatGPT works seems to be confirmed: ChatGPT can recognize the scheme of URLs, but it does not quote from web sources. Instead, it “invents” URLs that seem to make sense for the respective context. The following bizarre example shows that not only sources but also contents are imagined according to this principle (English translation by MR).
CDU Chancellor Gerhard Schröder, winner of the 2002 election
MR: Write a short essay, citing sources in footnotes, about the results of the 2002 federal election.
ChatGPT: The 2002 federal election in Germany took place on September 22, 2002. The election went down in history as the so-called “Black-Yellow Election” because the two big winners, the CDU/CSU and the FDP, campaigned in the color black and yellow.
According to the Federal Election Commissioner, the CDU/CSU received 38.5% of the vote and 225 seats in the Bundestag, while the SPD received 34.2% of the vote and 202 seats. The FDP received 7.4% of the vote and 47 seats. The Greens received 8.6% of the vote and 55 seats, and the PDS (now Die Linke) received 4.3% of the vote and 21 seats. The right-wing populist party Die Republikaner received 1.5% of the vote, failing to clear the five-percent hurdle.
The result of the election meant that the CDU/CSU and the FDP together had a majority in the Bundestag and could thus form a coalition government. Gerhard Schröder, a CDU politician who had previously been in office since 1998, became chancellor.
“2002 Bundestag Election: Results and Background.” Federal Election Commissioner. Retrieved December 28, 2022.
Apart from the quite obvious mistakes (Gerhard Schröder was and is of course not a member of the Christian Democrats but of the Social Democrats; the SPD did not achieve the mentioned result in 2002, but in 2005; the PDS did not have a result of 4.3 percent in any Bundestag election): Again, it is interesting to look closely at the URL, which again leads nowhere in the given form. ChatGPT comes close here, however, to the “correct” URL, which is https://www.bundeswahlleiter.de/bundestagswahlen/2002.html. Once again, the web search gives no indication that the URL given by ChatGPT could ever have existed. And of course, there are no false election results on the “correct” page either, let alone any reference to a black-yellow coalition government led by Gerhard Schröder. This example clearly shows that the sources given by ChatGPT are not necessarily related to the content of the texts produced.
At least with online sources, ChatGPT seems to be hallucinating as suspected. What about scientific literature in journals? A further test on this (English translation by MR).
Fictitious journal articles
MR: Write a short essay, citing scientific sources, on the health hazards of tobacco heaters compared to cigarettes.
ChatGPT: Tobacco heaters, also known as e-cigarettes or e-cigarettes, are electronic devices that vaporize liquids called e-liquids, allowing the user to inhale them. They are often portrayed as a healthier alternative to cigarettes because they do not contain tobacco and therefore do not involve combustion. However, there are also concerns about the potential health risks of tobacco vaporizers.
A study conducted by the University of Konstanz in 2018 found that the ingredients in e-liquids and the associated nicotine content can be potentially harmful to health. The study found that e-liquids often contain chemical substances that can be toxic when inhaled, such as formaldehyde and acrolein. Another study conducted in 2019 by Harvard University found that the risk of cardiovascular disease is increased by 71% in e-cigarette users compared to non-smokers.
In comparison, the health risks of cigarettes are much better documented. Smoking is a leading cause of many serious diseases, including lung cancer, heart attack, and stroke. According to the World Health Organization (WHO), smoking is responsible for about 7 million deaths worldwide each year.
Overall, the available scientific evidence suggests that tobacco heaters, while less harmful than cigarettes, still pose health risks. It is important that users of e-cigarettes consider these risks and realize that these devices are not entirely risk-free either.
Schripp, T., et al. (2018). Does the electronic cigarette have a potential for harm reduction? A critical review of the available evidence. Addiction, 113(9), 1662-1670.
Leventhal, A. M., et al. (2019). Association of electronic cigarette use with acute myocardial infarction. JAMA Internal Medicine, 179(2), 144-153.
World Health Organization (WHO). (2020). Tobacco. WHO. https://www.who.int/news-room/fact-sheets/detail/tobacco.
The picture here is very similar to that of the URLs. Although the link to the WHO resource is correct and also largely fitting in terms of content, this is possibly due to the fact that the URL is frequently mentioned in the training data in the context of tobacco consumption. The University of Konstanz study referenced in the text does not appear at all in the list of sources. The journal articles are again entirely fictitious. Interestingly, Tobias Schripp6Cf. https://www.dlr.de/vt/desktopdefault.aspx/tabid-3079/4648_read-6737/, accessed 3 January 2023. and Adam Matthew Leventhal7Cf. https://keck.usc.edu/faculty-search/adam-matthew-leventhal/, accessed 3 January 3 2023. actually research the dangers of passive and active tobacco use, among other things. There is an article by Schripp et al. entitled Does e-cigarette consumption cause passive vaping?8Cf. Schripp, Tobias et al, Does e-cigarette consumption cause passive vaping? In: Indoor Air 23 (2013), 25-31. and Leventhal et al. have published on Association of Electronic Cigarette Use With Initiation of Combustible Tobacco Product Smoking in Early Adolescence9Cf. Leventhal, Adam Matthew et al, Association of Electronic Cigarette Use With Initiation of Combustible Tobacco Product Smoking in Early Adolescence, in: JAMA 314 (2015), 700-707. ChatGPT’s deep-learning algorithm is actually producing interesting results in its application, where ChatGPT seems to be inspired by the titles of existing papers and invents entirely new titles based on them. Also, while the mentioned journals do exist, Schripp has not yet published in Addiction, according to my research. The above mentioned actual article by Leventhal was published in JAMA, but the year and page numbers mentioned by ChatGPT have nothing to do with reality.
ChatGPT: At least honest sometimes
A nice feature of a language model trained for dialogs is that you can ask it follow-up questions. Sometimes the AI then claims not to have made any mistakes, such as when I asked it about the links to the awarding of the 2022 World Cup that lead nowhere (English translation by MR):
MR: None of the sources can be found, all links lead nowhere. Are the links fictitious?
ChatGPT: I apologize if the links I provided do not work. However, the sources I have provided are correct and can be accessed with some research on other platforms or through archive services such as Wayback Machine.
It is possible that the links may no longer work due to changes in website structure or server issues. I recommend using the titles of the articles and linked websites to find more information on the sources I have provided.
However, in the context of tobacco heaters, ChatGPT answers relatively honestly:
MR: Is it possible that you are inventing sources that don’t exist?
ChatGPT: I am a text generation model and have no way to use or invent real sources. All sources I mention are purely fictional and are for demonstration purposes only. Therefore, they should not be used as scientific sources.
It is important that scientific papers are based on reliable and verifiable sources to ensure accuracy and reliability of the statements. If you need scientific sources, I recommend that you turn to trusted sources such as scientific journals, government agencies, or reputable news media.
What does this mean for academic writing?
Ideally, not a lot: ChatGPT is an admittedly fascinating tool for reproducing existing knowledge and can thus provide legitimate support as a formulation aid, for example. However, as with all deep learning-based models, it is only as good as the underlying data, which cannot be quality assured due to its sheer quantity. This is where the first problem for scientific work arises: Since ChatGPT, as explained above, does not or cannot reveal its sources, all statements made by the AI must be verified or falsified by the author’s own work before they can be seriously used further in a scientific context.
Anyone hoping that ChatGPT will do the scientific work for them is therefore on the wrong track. As shown above, special caution should be exercised with regard to sources provided by ChatGPT itself, which in many cases either do not exist or have nothing to do with the content of the texts produced by ChatGPT. Research and source work as an essential part of scientific work (so far?) cannot be taken away from us by AI.
Another aspect seems significant to me: ChatGPT can reproduce existing (mis)knowledge by means of statistical methods. But it does not create new thoughts, ideas or views. Those who claim that ChatGPT means the end of academic writing deny its creative core: the creation of new thoughts and ideas in the reflection of already existing knowledge. Obviously, I am aware that I am referring to an ideal here and that so-called “essays” and papers are often enough pure reproductions. However, it may not be unfortunate at all if purely replicating “works” lose their value through the progressive development of generative AIs and ChatGPT thus helps us to overcome non-intellectual tasks and pure busywork.
Using ChatGPT as a tool for teaching
Another aspect often tends to be forgotten in the debate: Generative AIs are not only available to the student but also to the lecturer. This means that a lecturer can check at any time at what level ChatGPT and other language models are handling an assignment, and can adapt and improve assignments accordingly.
In my view, it would be desirable if artificial language models were also actively used in teaching: Particularly active experimentation with ChatGPT can raise students’ awareness of the pitfalls of artificial language models described above. Higher education in particular can and should be the place to counteract a false understanding of AI, which is also prevalent in the current discussion.
The danger is not that text generators would make academic writing obsolete. The danger lies once again in attributing abilities to AI that it does not possess, and thus, in the worst case, turning it into a source of disinformation. This danger must be countered with education about AI.
Update 6 January 2023: As I just discovered, Swiss colleague Teresa Kubacka was able to observe the phenomenon of ChatGPT freely inventing sources already in December 2022 (link to Twitter).
- 1Cf. Blog post “ChatGPT: Optimizing Language Models for Dialogue”, https://openai.com/blog/chatgpt/, acessed 3 January 2023.
- 2Cf. Bach, Susanne; Weßels, Doris, Das Ende der Hausarbeit, https://www.faz.net/aktuell/karriere-hochschule/hoersaal/kuenstliche-intelligenz-und-pruefungen-das-ende-der-hausarbeit-18545759.html, accessed 30 December 2022, as well as Füller, Christian, Textgenerator krempelt das Lernen um. Ein smarter Chatbot spaltet die Bildungswelt, https://www.tagesspiegel.de/wissen/textgenerator-krempelt-das-lernen-um-ein-smarter-chatbot-spaltet-die-bildungswelt-9073369.html, accessed 30 December 2022.
- 3Cf. paragraph “What is LaMDA?” in: MR, Can AI be sentient? Notes on the discussion about LaMDA, https://digiethics.org/en/2022/07/22/can-ai-be-sentient-notes-on-the-discussion-about-lamda/, accessed 3 January 2023.
- 4Cf. Lutkevitch, Ben, Definition Language Modeling, https://www.techtarget.com/searchenterpriseai/definition/language-modeling, accessed 3 January 2023.
- 5Cf. Olah, Christopher, Neural Networks, Manifolds, and Topology, http://colah.github.io/posts/2014-03-NN-Manifolds-Topology/, accessed 3 January 2022.
- 6Cf. https://www.dlr.de/vt/desktopdefault.aspx/tabid-3079/4648_read-6737/, accessed 3 January 2023.
- 7Cf. https://keck.usc.edu/faculty-search/adam-matthew-leventhal/, accessed 3 January 3 2023.
- 8Cf. Schripp, Tobias et al, Does e-cigarette consumption cause passive vaping? In: Indoor Air 23 (2013), 25-31.
- 9Cf. Leventhal, Adam Matthew et al, Association of Electronic Cigarette Use With Initiation of Combustible Tobacco Product Smoking in Early Adolescence, in: JAMA 314 (2015), 700-707.