ChatGPT: Powerful and versatile tool, but with limitations and controversy
Originally published in Linkedin at: 28 Feb, 2023.
ChatGPT has become a highly popular tool due to its remarkable range of features that function seamlessly. This NLP (Text-to-Text) model is based on a multi-layer transformative neural network and is trained by reinforcement learning with human feedback [1]. It should be noted that ChatGPT is a variant of the Generative Pretrained Transformer (GPT) model developed by OpenAI.
Among its multiple natural language processing (NLP) features, ChatGPT stands out for its ability to translate texts [15], summarize them, and analyze sentiments [2]. In addition, restricted only to language, it allows users to write poems [11], full articles and research papers [18] easily and comfortably [3].
ChatGPT is capable of understanding and responding to natural languages thanks to its use of deep learning, which allows it to generate text that is similar to that produced by humans, solve mathematical problems but not so complex [16], tell stories, explain complex topics from any discipline such as quantum computing and statistics [5], telling jokes [14][35], marketing [31], developing code [32][33][34], among other things.
Coding examples
Let’s start with an example for Salesforce, we’ll ask ChatGPT to create a trigger for cases before inserting so that it doesn’t allow the subject to contain bad words:
It looks like it can work, but the code is still improvable so I asked to use clean code:
Wow! I see no problems in this code other than that it separates the logic to check for profanity and to get the error message while still having descriptive names on the functions and variables. Now it’s time to do the unit tests:
I’m surprised, ChatGPT was able to generate the corresponding test class for the new trigger without any problems. However, there is room for refactoring! The name of the tests could be more descriptive, the case creation logic could be put into a method and reused as well as the SOQL query for the insertedCases. Let’s see how ChatGPT reacts when asked to refactor two aspects on purpose:
Refactors the build, but not the Query. Apparently ChatGPT did not understand this instruction, so we proceed to give negative feedback and give the query refactor as expected result and improve the names of the tests:
So, after the feedback, I ask again to refactor:
ChatGPT managed to refactor the query! But it did not take into account the name of the tests given in the feedback, although with this it is already demonstrated that it is capable of developing a functional trigger with 100% coverage while still having clean code.
Let’s see an example of a form with Vue:
And with React:
But what if we asked for something more complex? Let’s try a calculator in Flutter:
It goes on at length later but by the time it was finishing a Widget, it finished generating code. Apparently you have to be more specific or impose certain limits for the response to be fully generated. So the question arises: are there limitations for ChatGPT?
Limitations
There are limitations pointed out by OpenAI itself [7]:
- ChatGPT sometimes writes plausible-sounding answers that are incorrect or nonsensical. Solving this problem is challenging, due to the inherent characteristics of the model and how it is built.
- ChatGPT is sensitive to small changes in the wording of the input or trying the same request multiple times. For example, given a wording of a question, the model may claim that it does not know the answer, but given a slight rephrasing, it may answer correctly.
- The model is often excessively verbose and overuses certain phrases, such as repeating that it is a language model trained by OpenAI. These issues arise from biases in the training data (trainers prefer longer answers that appear more complete) and from known over-optimization issues.
- Ideally, the model would ask clarifying questions when the user provides an ambiguous query. Instead, current models often guess what the user intended.
- Sometimes you may respond to harmful instructions or exhibit biased behaviors. In addition, it can have some false negatives and positives, so it is dependent to a certain degree on human feedback.
As disadvantages, it can present biases due to its human feedback. An example is the one that can have bias and tries to be politically correct [8] (in fact, its use can become more political [24]); does not distinguish truth from fiction and is prone to “hallucinations” [12][17][21][22]; is maybe a bit stupid [14].
In addition, universities are concerned about plagiarism [9][18][25][26][27][29][30], but it also seems to be an educational tool [13] and it is necessary to train teachers as well as students. for its correct use [20][28].
And Microsoft?
Bing, Microsoft’s search engine, is renewed with artificial intelligence. At the company’s Redmond, Washington, US facility, managers recently unveiled a new version of the tool that incorporates startup OpenAI’s ChatGPT viral chatbot technology [12].
And Google?
Google has not been left behind in the trend of searching for information through chatbots. As a response to the integration of ChatGPT in Bing, the company has developed Bard, although it is only available to authorized Google personnel. However, errors in the operation of the chatbot have been reported [10][13].
It is worth mentioning that there are various alternatives to ChatGPT, such as ChatSonic, Chinchilla, Bloom, Replika, Jasper, LaMDA and Meta AI Speech from Brain.
Conclusions
Apparently not everything is hunky-dory, despite the fact that it is a powerful and versatile tool capable of generating text from a text input, there are several limitations that must be taken into account. These limitations are due to the capacity of the model and partly to human feedback, as well as possible biases and hallucinations in text generation.
Also, it is important to note that ChatGPT is not a source of truth and its use to generate plagiarism is not recommended. It has also been noted that he has difficulty translating into distant languages. Despite these limitations, it is important to note that as the use of ChatGPT increases, the feedback it receives will also increase, which will help improve the quality of the model (this was demonstrated in the example of unit tests in Apex). Microsoft has taken advantage of this opportunity and has incorporated ChatGPT into Bing, while Google is working on developing its own chatbot.
Bibliographic references
[1] Gozalo-Brizuela, R., & Garrido-Merchan, E. C. (2023). ChatGPT is not all you need. A State of the Art Review of large Generative AI models. arXiv preprint arXiv:2301.04655. https://arxiv.org/pdf/2301.04655.pdf
[2] German Majumder (2022). ChatGPT-what is it and how does it work exactly?. Medium. https://medium.com/geekculture/chatgpt-what-is-it-and-how-does-it-work-exactly-62e7010524d3
[3] Ankit Patel (2023). What is ChatGPT and How Does it Work? [Examples & Alternative]. XongoLab. https://www.xongolab.com/blog/what-is-chatgpt-and-how-does-it-work/
[4] Sobania, D., Briesch, M., Hanna, C., & Petke, J. (2023). An Analysis of the Automatic Bug Fixing Performance of ChatGPT. arXiv preprint arXiv:2301.08653. https://arxiv.org/abs/2301.08653
[5] Rudolph, J., Tan, S., & Tan, S. (2023). ChatGPT: Bullshit spewer or the end of traditional assessments in higher education?. Journal of Applied Learning and Teaching, 6(1). https://journals.sfu.ca/jalt/index.php/jalt/article/view/689
[6] Omar, R., Mangukiya, O., Kalnis, P., & Mansour, E. (2023). ChatGPT versus Traditional Question Answering for Knowledge Graphs: Current Status and Future Directions Towards Knowledge Graph Chatbots. arXiv preprint arXiv:2302.06466. https://arxiv.org/pdf/2302.06466.pdf
[7] OpenAI (2023). ChatGPT: Optimizing Language Models for Dialogue. https://openai.com/blog/chatgpt/
[8] Anna Lagos (16 Feb, 2023). ChatGPT tiene postura política: puede dedicar poemas alabando a Obama y López Obrador; pero no a Trump o Bolsonaro. Wired, https://es.wired.com/articulos/chatgpt-tiene-postura-politica-puede-dedicar-poemas-alabando-a-politicos-como-obama-lula-o-amlo-pero-no-a-trump-bolsonaro-o-pena-nieto.
[9] Sofia Barnett (1 Feb, 2023). ChatGPT obliga a las universidades a replantearse el plagio. Wired, https://es.wired.com/articulos/chatgpt-obliga-a-universidades-a-replantearse-plagio-inteligencia-artificial-generativa
[10] Jorge Garay (13 Feb, 2023). Estudiante accede a información confidencial de Microsoft engañando a ChatGPT. Wired, https://es.wired.com/articulos/estudiante-accede-a-informacion-confidencial-de-microsoft-enganando-a-chatgpt-sydney
[11] Fernanda Gonzalez (14 Feb, 2023). ¿Un chat de IA escribió la carta de amor que recibirás este 14 de febrero?. Wired, https://es.wired.com/articulos/chats-de-inteligencia-artificial-escriben-cartas-de-amor
[12] Aarian Marshall (8 Feb, 2023). Microsoft presenta su nueva implementación de ChatGPT para impulsar Bing y vencer a Google. Wired, https://es.wired.com/articulos/microsoft-recurre-a-chatgpt-para-impulsar-bing-y-vencer-a-google
[13] Fernanda Gonzalez (9 Feb, 2023). Bard responde mal una pregunta y Google pierde 100,000 millones de dólares. Wired, https://es.wired.com/articulos/bard-responde-mal-una-pregunta-y-google-pierde-100-mil-millones-de-dolares
[14] Amit Katwala (11 Dic, 2022). Los textos que crea la inteligencia ChatGPT son estúpidos porque se entrenó con humanos. Wired, https://es.wired.com/articulos/por-que-chatgpt-es-una-inteligencia-artificial-tan-convincente
[15] Jiao, W., Wang, W., Huang, J. T., Wang, X., & Tu, Z. (2023). Is ChatGPT a good translator? A preliminary study. arXiv preprint arXiv:2301.08745. https://arxiv.org/pdf/2301.08745.pdf
[16] Frieder, S., Pinchetti, L., Griffiths, R. R., Salvatori, T., Lukasiewicz, T., Petersen, P. C., … & Berner, J. (2023). Mathematical capabilities of ChatGPT. arXiv preprint arXiv:2301.13867. https://arxiv.org/pdf/2301.13867.pdf
[17] Bang, Y., Cahyawijaya, S., Lee, N., Dai, W., Su, D., Wilie, B., … & Fung, P. (2023). A Multitask, Multilingual, Multimodal Evaluation of ChatGPT on Reasoning, Hallucination, and Interactivity. arXiv preprint arXiv:2302.04023. https://arxiv.org/pdf/2302.04023.pdf
[18] Liebrenz, M., Schleifer, R., Buadze, A., Bhugra, D., & Smith, A. (2023). Generating scholarly content with ChatGPT: ethical challenges for medical publishing. The Lancet Digital Health. https://www.thelancet.com/pdfs/journals/landig/PIIS2589-7500(23)00019-5.pdf (accessed 17 Feb, 2023).
[19] Karpf D. Money will kill ChatGPT’s magic. The Atlantic, Dec 21, 2022. https://www.theatlantic.com/technology/archive/2022/12/chatgpt-aichatbots-openai-cost-regulations/672539/ (accessed 17 Feb, 2023).
[20] García-Peñalvo, F. J. (2023). La percepción de la Inteligencia Artificial en contextos educativos tras el lanzamiento de ChatGPT: disrupción o pánico. Education in the Knowledge Society (EKS), 24, e31279-e31279. https://revistas.usal.es/tres/index.php/eks/article/download/31279/29185
[21] Alier-Forment, M., & Llorens-Largo, F. (2023). EP-31 Las Alucinaciones de ChatGPT con Faraón Llorens In Cabalga el Cometa. https://bit.ly/3ZCNBVT
[22] Bowman, E. (2022, 19 Dic). A new AI chatbot might do your homework for you. But it’s still not an A+ student. NPR. http://bit.ly/3QL6z8A
[23] Codina, L. (2022, 12 Dic). Cómo utilizar ChatGPT en el aula con perspectiva ética y pensamiento crítico: Una proposición para docentes y educadores. Lluís Codina. http://bit.ly/3iKBFAE
[24] Cowen, T. (2022, 6 Dic). ChatGPT Could Make Democracy Even More Messy. Bloomberg. http://bit.ly/3XwqjPs
[25] Stokel-Walker, C. (2022, 9 Dic). AI bot ChatGPT writes smart essays — should professors worry? Nature. https://doi.org/10.1038/d41586-022-04397-7
[26] Ropek, L. (2023, January 4th). New York City Schools Ban ChatGPT to Head Off a Cheating Epidemic. Gizmodo. http://bit.ly/3kp8Ha9
[27] Cotton, D. R., Cotton, P. A., & Shipway, J. R. (2023). Chatting and Cheating. Ensuring academic integrity in the era of ChatGPT. https://edarxiv.org/mrz8h/download?format=pdf
[28] Zhai, X. (2022). ChatGPT user experience: Implications for education. Available at SSRN 4312418.https://www.researchgate.net/profile/Xiaoming-Zhai/publication/366463233_ChatGPT_User_Experience_Implications_for_Education/links/63a2817aca6a9d254f8de010/ChatGPT-User-Experience-Implications-for-Education.pdf
[29] Khalil, M., & Er, E. (2023). Will ChatGPT get you caught? Rethinking of Plagiarism Detection. arXiv preprint arXiv:2302.04335. https://arxiv.org/pdf/2302.04335
[30] King, M.R., chatGPT. A Conversation on Artificial Intelligence, Chatbots, and Plagiarism in Higher Education. Cel. Mol. Bioeng. 16, 1–2 (2023). https://doi.org/10.1007/s12195-022-00754-8
[31] Más Que Marketing | David García-Amaya. 10 Usos Prácticos de ChatGPT 🥇 | Funcionalidades para Alucinar con la Inteligencia Artificial 🧠. Youtube, 29 Dic, 2022. https://www.youtube.com/watch?v=Op3Q1CXF0u8
[32] Más Que Marketing | David García-Amaya. 13 Usos Prácticos de ChatGPT 🥇 (Parte 2) ¡Seguimos ALUCINANDO 🧠! | Guiones, resúmenes, layouts…. . Youtube, 5 Ene, 2023. https://www.youtube.com/watch?v=Kt0iLVrY2pE
[33] Kiko Palomares. 🤖 Asi he CREADO una WEB Completa desde Cero usando ChatGPT ✅ (y alguna que otra IA más…). Youtube, 16 Ene, 2023. https://www.youtube.com/watch?v=8OufESMHw64
[34] Nick White. How I Coded An Entire Website Using ChatGPT. Youtube, 4 Ene, 2023. https://www.youtube.com/watch?v=ng438SIXyW4
[35] Website So Simple. 10 Best ChatGPT Examples, Prompts & Use Cases (Chat GPT Demo & Tutorial). Youtube, 3 Ene 2023. https://www.youtube.com/watch?v=fgjoF97gdjQ