Generative AI tools

ChatGPT is an AI language model developed by OpenAI. It is part of the GPT series, with GPT-4o being the latest version. It is designed to understand and generate natural language responses, and it is thus capable of engaging in conversations, answering questions, and providing information or assistance across a wide range of topics by following instructions from prompts by the user.

The model is trained on a diverse range of internet text, but it is important to keep in mind that it does not automatically browse the internet or access real-time information. This means that while it can provide information that is current up to its last training data in April 2023, it may not have the latest information on events or developments that occurred after that date. An option to use ChatGPT-4 with internet access for free, is Microsoft Copilot (formerly Bing Chat).

While the full functionalities of ChatGPT-4 (including Dall-E, plugins, custom GPTs etc.) can only be used via their subscription plan, there are several GPT-4 based services that you can use for free:

• Microsoft Copilot (formerly Bing Chat; possibility to change the conversation style to more creative or more precise)
• Perplexity.ai
• Poe.com
• Merlin Chrome extension

The following list is of course incomplete, it is just a selection of the most well known ones:

• Claude 3.5 Sonnet: Released in June 2024, Claude 3.5 Sonnet is Anthropic's newest generative AI chatbot version. It is free to use up to a certain limit and sharing user data is opt-in (instead of opt-out with e.g. Chat-GPT). It's training data cutoff date is April 2024; it does not have access to the internet. With the new "Artifacts" feature it is possible to create code, documents, webdesigns etc. that are displayed in a side window, making it easy to build on and change the creations in real time.
• Microsoft Copilot: Microsoft’s free for use chatbot, formerly Bing Chat, has internet connection and is able to cite sources, generate images, and write lyrics and music via plugin. It is built on GPT-4 and best in creative or precise mode.
• Gemini: Google’s chatbot (formerly known as Bard) has internet connection and integration in Google apps. A free version, as well as a subscription “Pro” / “Advanced” version are available. It is built on GPT-4 and is able to generate text, images, code etc.
• Perplexity AI: A conversational search engine that uses natural language predictive text to answer queries. While its conversational skills are less sophisticated than ChatGPT-4’s, it is able to use sources from the web and reference them. It has access to GPT-4, Claude 2, and Gemini Pro.
• GPT4All: GPT4All, created by Nomic AI, is an ecosystem to train and deploy customised LLMs that run locally on consumer grade CPUs without internet connection or a GPU. They are free and open source and can be used to integrate LLMs into applications without a subscription. Your data and chats are only saved on your local hardware, unless you intentionally share it with GPT4All to help grow their models.
• TalkAI: TalkAI is a virtual assistant / AI powered chatbot that can directly be integrated into WhatsApp and Telegram. It is built on GPT-3.5 and offers voice response, commands, and translation.
• Mistral AI: This Open Source AI model is free for everyone to use and modify and is developed by a French startup of the same name. With some know-how, the models are adjustable.
• LLaMA 2: A family of pre-trained, fine-tuned Open Source LLMs developed by Meta. They are free of charge for research and commercial use and can be downloaded in three different model sizes. With some know-how, the models are adjustable.

Yes, you can choose not to share your chat history with OpenAI to train their models via your account settings. For more information, see here.

While generative AI tools, such as ChatGPT or Claude, are not specialised in e.g. literature search, they can be helpful with other parts of the scientific writing process, for example:

• Drafting: When faced with a writer's block, generative AI tools can serve as a useful starting point. They can provide a first draft, e.g. of more narrative sections of the manuscript, a cover letter, replies to reviewers, etc. This draft can be considered an inspiration for further development.
• Paraphrasing and summarising: Generative AI tools can help paraphrase sentences to avoid plagiarism, summarise sections, or simplify text to make it better understandable for people with different backgrounds.
• Sparring partner: Generative AI tools can be used to further develop, deepen, or work out ideas and concepts. Because these tools create content using a wide variety of training data, they can provide different viewpoints and perspectives. Instead of simply asking generative AI tools for output, they can also be prompted to ask you questions. It can be a discussion between yourself and the tool, bringing up aspects that you might not have considered.
• Improving language: These tools can help native and non-native speakers improve their texts with regard to grammar, orthography, style, formality, or context. For some of these tasks, there are more specialised tools; however, generative AI can be a good starting point.
• Data analysis: To a certain extent, generative AI can help with data analysis by using statistical methods; creating charts, graphs, and tables; interpreting results; and suggesting ways to represent findings effectively.
• Citation management: Generative AI can assist with citation management and formatting and create and organise bibliographies. However, it is advisable to (also) use a specialised reference management tool for these tasks (e.g. Zotero, Endnote, Mendeley).
• Organisation: Formatting a manuscript according to journal guidelines or setting up a writing schedule are tasks with which generative AI can help.

Custom GPTs as extensions or browser plugins can be helpful or necessary for certain tasks. Be sure to always check and verify the output – use it as a sparring partner, a source of inspiration, and suggestion rather than a finished product. As the author, you are responsible for the content of the text.
 

The use of generative AI tools in scientific writing poses a number of risks and it is crucial to be mindful of them. Here is a selection of some points to consider:

• Misinformation and hallucinations: Generative AI works statistically and it is not able to judge whether something is true / untrue or ethical / unethical. It is therefore possible that the tools produce false or inaccurate information, make up references etc. all while sounding very confident and with a factual tone. Always verify outputs and check with reliable sources.
• Unidentifiable sources: Many AI tools do not cite their sources and it is very probable that they mix together content that they have encountered before. It is also unclear on which specific data the model was trained.
• Bias: Outputs are not neutral, they have a certain bias due to their training data and their algorithm and programming. Since they need to make certain classifications and assumptions for efficiency reasons, they tend to reinforce stereotypes and marginalise minority perspectives (see e.g. Feng et al. 2023).
• Interference with learning, agency and critical thinking: Generative AI tools readily offer “nice” sounding answers; they can provide a solution without highlighting the process behind it. An over-reliance on these tools can therefore interfere with the learning process of discovery and independent research. Furthermore, relying too heavily on these tools can reduce one’s own agency in decision-making during the process and critical thinking regarding which elements should be considered or included.
• Time costs: AI-based tools promise an increased efficiency. However, getting to know these tools and especially creating effective prompts that generate the desired output are very time consuming. It is possible that the time saved for writing and doing independent research could just be spent for prompting and verifying the AI’s output. Moreover, if these steps are not taken, the increased efficiency of using generative AI tools may come at the cost of attention to detail, precision and depth.
• Writing style: The use of generative AI poses different risks for the writing style: one’s personal style might be lost, creativity is limited, there might be an inconsistency of styles if different tools are used or simply because this can happen, context specific vocabulary might not be known by the tool, the tone might be off, empathy could be missing, etc. 
• Plagiarism and academic integrity: While the use of AI-based tools does not constitute plagiarism, you need to make sure you comply with all existing guidelines and laws (e.g. department or institute guidelines, research integrity guidelines, citation etiquette, declaration of originality, journal or publisher guidelines, copyright laws, etc.). It is good practice to transparently document the use of AI-based tools in a scientific work. Check, if and how you should declare the use of AI-based tools and adhere to the specific rules of your citation. Furthermore, be mindful that any text that you feed into AI-based tools can potentially be reused for training; if your text contains any sensitive or confidential information and data, do not input it into such tools.