Disclosing Use of AI for Writing Assistance in Genealogy

Exploring the use of artificial intelligence chatbots for creating written content is the new frontier – full of uncertainties and unknowns. Organizations that once prohibited the use of LLMs for research papers are now allowing them, as long as authors are fact-checking the output and attempting to cite potentially copyrighted materials included in the output. In 2023, the International Conference on Machine Learning prohibited the use of LLMs in their call for papers. In 2024, they are allowing LMs as a general-purpose writing assist tool, but authors should take full responsibility for the contents of their papers, including content that could be construed as plagiarism or fabrication of facts. They also suggest that LLMs are not eligible for authorship.[1]

Is it okay to use an AI chatbot to help with our genealogy writing? The answer may depend on the audience and the use of your written product, and whether or not it’s being published. Editors of journals typically have policies about the use of generative text models.

Students using AI Chatbots

Students need to consider if their teachers or professors allow the use of AI in their writing. If a student is learning how to write an essay, it defeats the purpose of the writing assignment to use AI. If the student is taking a test or having their writing abilities evaluated, using an AI chatbot is probably cheating.

Genealogy Writing that is Being Evaluated

If you are writing a three-generation report for accreditation, or a case study for your certification portfolio, you may make a different decision about using AI than if you are writing a report for yourself. For example, BCG asks that portfolios are the applicants’ own work and not reviewed by anyone. If an LLM generates much of the text for you, then it’s not your own work anymore.

Taking Credit

Genealogy Standard #62 states that we should not “assume credit for others’ words and ideas.”^[2] AI chatbots are not  human, and cannot be authors or contributors, yet the question then arises – do you need to give the AI chatbot credit for contributing words and ideas to our writing? Mohammad Hosseini, David B Resnik, and Kristi Holmes, the authors of “The Ethics of Disclosing the Use of Artificial Intelligence Tools in Writing Scholarly Manuscripts” say that if you don’t give credit to the LLM, you are claiming credit for work you did not do.^[3] Elizabeth Shown Mills says, “When we use generative AI to create text-prompted images or narratives, the same expectations exist. We do not take credit for what we did not create.”^[4]

Genealogy Writing Commissioned by Ourselves

Most of the time, genealogists are not writing for teachers (except those in family history degree programs and certification courses). Usually, genealogists are writing in order to keep track of what they have found, share with family members, and preserve family information for future generations. Perhaps you are writing a report to yourself or a proof argument to be posted online. Reports to ourselves and proof arguments posted for the information of others are not commissioned by clients or being tested by anyone – so the use of AI generative text models for these is probably fine, provided that we are giving the models the information we want them to use, instead of asking them to come up with information on their own.

Genealogy Writing Commissioned by Others

When writing client reports, handouts, syllabus materials, and so forth, our writing is commissioned by a client or genealogy society/group. They asked us to create the research report or syllabus because of our qualifications and skills in writing. Our personality, skills and experience give our writing a unique style and tone. Large language models have their own style and tone – sometimes described as generic and robotic.^[5]  If we hand over the creation of our reports and syllabus materials to the AI chatbot, our unique authorship, knowledge, style, and tone can be lost.

Authenticity and Personal Style

While LLMs can be trained to use our own style, they cannot be experts to the level that we have become experts in our area. In this category of writing, we must be careful that the use of AI chatbots doesn’t water down our expertise or cause our unique style and tone to be lost. It is imperative to review the paragraphs carefully to ensure that they reflect the unique experience and knowledge we have developed over time. We can also use prompts or GPTs that we have trained in our style.

Example of Writing a Video Description for YouTube

For example, for my YouTube channel, I use ChatGPT to summarize videos and make the process of writing video descriptions more efficient. I uploaded the video transcript and went back and forth with the chatbot refining the summary. I asked for the summary to be more concise, and the result stripped out specific details and kept general statements. The title of the video was “Documenting Lines of Descent From a Most Recent Common Ancestor to a DNA Match.” The over-simplified description generated by ChatGPT was:

This video is from the ‘Research Like a Pro Question and Answer Series’ and talks about how to show family connections using DNA tests. It explains why it’s important to keep good records of family history, especially when using DNA. The video uses examples to show how to make your family tree research believable and accurate by including proper evidence. It stresses that good evidence is key to understanding family links through DNA.

Although I liked the simplicity and length of the paragraph, I realized that it was so watered down that it didn’t accurately and specifically describe the video’s contents. If I had written a four-sentence summary of the video, it would have been much more specific about suggestions and methods discussed in the video. I refined the paragraph through additional prompts asking for specific examples of tools and methods used, then edited the first sentence to include that the video is about documenting the line from DNA matches to the MRCA couple. Here is the final result:

In this presentation from the Research Like a Pro with DNA study group, Diana Elder, AG, talks about how to document family connections from DNA matches up to the MRCA couple. She explains the need for proof arguments and tables to show these connections clearly, like in Nicole’s study on Barsheba Tharp Dyer. Diana points out that showing parent-child links requires a lot of documentation. She suggests including this information in reports or appendices to manage the large amount of data. She also mentions using Ancestry trees or the researcher’s website to share these documents, which helps save space and makes it easier for readers to see the evidence. Diana emphasizes the importance of solid evidence in making genealogical research believable and trustworthy.

Lindsey Wieck, a professor of history at St. Mary’s University, found this same result when using an AI tool to edit sentences and paragraphs. If she did not provide careful oversight, meaning was lost.[6] This did not stop her from using the AI tool, but she came to think of it as a “very patient and thorough editor,” rather than a one-and-done tool where you upload a rough draft get a perfect final draft as the output.

Appropriately Giving Credit

If we use LLMs in our genealogy writing, how can we appropriately give it credit? Experts generally agree that listing an AI chatbot as an author or contributor is not appropriate, since an AI chatbot isn’t human and can’t take responsibility for the writing.[7] Another consideration is that the training data for the generative text models came from other others’ work published online. Often the model gives us ideas or text but we need to cite the underlying works used as the models’ training data. Some models can provide sources for where their information came from. We can also mitigate this issue by providing the information and citations ourselves, and simply asking the model to assist with putting it into coherent paragraphs.

Many editors of scholarly journals have come together and created a statement on the responsible use of Generative AI technologies in scholarly journal publishing. They suggest that authors who use generative AI to summarize literature, form ideas, make outlines, write drafts, and revise text should disclose this to their editors, reviewers, and readers; yet because the technology is so new, they have decided not to recommend rules for how to disclose it. They say,

Since generative AI is constantly changing and the scholarly community is only beginning to experiment with it, it is not prudent at this time to promulgate hard and fast rules for how generative AI should be disclosed. We recommend, however, that disclosure should describe how the AI was used and should identify AI-generated content. Authors should err on the side of too much transparency rather than too little: when in doubt, disclose. Some ways of disclosing the use of generative AI could include describing the use in a paper’s introduction, methods section, appendix, or supplemental material or citing the generative AI tool in the notes or references.[8]

Additional ideas for disclosing use of AI tools in published papers comes from a computational linguistics conference. At the 2023 Association of Computational Linguistics conference, the program chairs created a policy on AI Writing Assistance.[9] They defined six different ways AI can be used in writing assistance, and proposed methods for citing the AI, if necessary. Their categories are as follows:

1. Language assistance: use of Grammarly, spell checkers, dictionaries, etc. – use of these tools does not need to be disclosed.
2. “Short-form input assistance:” smart compose in Google Docs and other predictive tools – use does not need to be disclosed.
3. Literature search: use of generative text models to help identify articles and literature to review is like using a search engine. Authors should read the suggested literature and cite the references appropriately. Use of the model should be disclosed.
4. “Low-novelty text:” use of generative text models to produce descriptions of “widely known concepts.” Authors should check the work of the model and if the model copies verbatim from an existing article, acknowledge it with a citation. Use of the model should be disclosed.
5. Novel ideas: if the generative text model produces text that gives the author new ideas, that should be acknowledged. Authors may also want to check for publications that contain those new ideas in order to cite them. Use of the model should be disclosed.
6. “New ideas + new text”: the use of generative text models to contribute ideas and write about them requires fact-checking and finding/adding citations. ACL discourages this use. They suggest that authors who do use models for this purpose should disclose its use and let the peer reviewers know that they have done their due diligence to cite any other works that the model may have been trained on.

 

In their 2024 call for papers, the ACL adopted their 2023 AI Writing Assistance Policy and added a seventh usage, one that needs to be disclosed – assistance with code writing.[10]

Creating Source Citations for AI Chatbots

The Chicago Manual of Style recommends citing content generated by AI by including the model name, prompt or topic, the company, and date.^[11] Others recommend adding the name of the person using the model, since the tool can become customized to the user’s preferences, as well as which version was used (since ChatGPT could be either 3.5 or 4, depending on whether or not the user is paying).^[12] Some style guides suggest different styles for reference note citations depending on whether or not the prompt was mentioned in the text.^[13]

Elizabeth Shown Mills suggests adding another layer to our citations describing any source we provided to the language model, using the citation format for that type of source.^[14]

Drafting Text and Disclosing This Usage

Generative text models are helpful tools for getting started – drafting text. The initial paragraph(s) generated by chatbots are usually not exactly what we want, so we will probably refine it quite a bit. The process of using a generative text model might look like this:

• you provide a prompt with your own ideas of what the paragraph should say
• the AI drafts a paragraph
• you refine the paragraph through additional prompts
• you edit the paragraph for final use in a written product

 

In this example, the generative text model was a tool you used to help you write “low-novelty text” (#4 in the list of possible uses above). The model was not coming up with new ideas, simply taking your idea and putting it into words. Similar to the ability of a family tree program that is capable of generating narrative reports from the facts we provide, the AI generative text model is a tool to help us take structured data and create a readable narrative output. However, unlike family tree programs, where the output usually looks and sounds like a report generated by a family tree program, the text generated by LLMs can sound very humanlike and often avoid detection. To follow the suggestions from ACL, this usage should be disclosed. To follow the recommendations from the statement on the responsible use of Generative AI technologies in scholarly journal publishing, we can disclose this use in the introduction, appendix, supplemental material, or reference notes.[15]

When writing a research report for myself, I would disclose the use of an AI tool to draft text in the limitations paragraph. This reminds myself and other readers that the report may contain errors due to not being written by a machine and not by a human. As the human using AI tools, it’s my responsibility to fact check the information. Although sometimes I feel like the chatbot generates writing more superior to my own, I’ve lately realized that my own unique style, voice, and experience is better than what the chatbot can produce, even if it has a larger vocabulary. This is why I’ve decided to put the disclosure of the use of AI in my limitations section.

I would only feel comfortable using AI tools for client reports for simple uses, like summaries of long transcriptions and generating a results summary after I’ve already written the report myself. I would cite my use of the tool for this in a reference note citation. However, I could potentially see a day when a professional genealogist gets permission from a client to use AI to write a report in order to save time and spend more time researching and logging the research. Writing a report from a research log can make the professional much more efficient and save the client’s valuable time and money. In this case, the use of the AI tool should definitely be mentioned in the limitations section and cited in a complete reference note citation.

Example Report Written by AI

For the NGS workshop I taught yesterday, I used my previous research on Baldy Dyer as an example of generating text from timelines and research logs with AI. I completed most of the research a few years ago, but never wrote a report. In the meantime, I received a bounty land warrant application, since Baldy died in the War of 1812, and added the results from that file to my log.

Using Claude, I uploaded a CSV file of my starting point timeline and asked it to generate paragraphs for the beginning of the report.

Then I uploaded a CSV file of my research log and guided the chatbot through creating sections of my report row by row of the log, with footnotes from my citations column. I asked Claude to use my own transcriptions and comments from the log as much as possible. Claude helped smooth out some of the text and added topic sentences.

Several rows in my log were for searches with negative results. I guided Claude through creating a bulleted list of the negative searches. My log didn’t have complete citations for those searches, just the browsed image collection titles, but Claude made them consistent by adding the website, URL, and date accessed.

I used Writage (https://www.writage.com/), a plugin for pasting markdown into Word, to paste the generated paragraphs while preserving the section headers, bold text, and footnotes. I added an objective to the top of the report and saved it as a PDF. I uploaded this file to Claude and asked it to write a conclusion by summarizing the findings section and to list progress made on the objective (to trace Baldy’s children).

In Research Like a Pro and in our client reports, we usually add a results summary section at the top of the report with a bulleted list of what was accomplished in this phase of research. I asked Claude to read the PDF and make a bulleted list with action verbs about the Findings section.

This report was the result:

Baldy Dyer research report from log – Claude AI April 2024

I haven’t fully checked every sentence of this report, but I hope you’ll see how valuable it can be to keep a detailed research log with complete citations for generating a research report with AI. You can also see how I mentioned the use of Claude in my limitations section.

Various Ways to Disclose the Use of AI Tools

Now that I’ve studied this topic and determined that it’s almost always a good idea to disclose the use of AI tools when you’ve used it to generate text, I feel much more confident about using AI tools when writing. I know that for many uses, I can simply note at the end of an article that I used ChatGPT to help draft some of the text. Here are some other ways I have and will disclose my use of AI tools for writing:

• Video and podcast descriptions: acknowledgement at the end of the summary
• Blog posts: reference note citations and acknowledgement at the end of an article
• Reports to self: reference notes and limitations section
• Client reports: limitations section, reference notes for summaries of long transcriptions; reference note for generating bulleted list of results; reference notes for explanations
• Syllabus materials: reference notes and note at in the footnote section of the first page
• Published journal articles: reference notes and note at in the footnote section of the first page – if the journal allows AI writing assistance

 

Conclusion

Hopefully this discussion gave you some ideas for when to disclose the use of AI. I found the six categories of AI use in writing assistance from the Association of Computational Linguistics conference to be especially helpful in determining when disclosure is needed. Spell-checking and grammar checking are clearly different from an LLM generating whole sections of a research report – and therefore should be treated differently. Let me know in the comments if you have ideas for how to do this as we all navigate the world of AI tools in genealogy writing together.

 

AI tools were not used to generate text for this blog post. 🙂

Learn more about Using AI Tools in our 4-day workshop, Research Like a Pro with AI, July 29-August 1, 2024.

 

Reference Notes

[1] Program Chairs for the International Conference on Machine Learning, “ICML 2024,” icml.cc, 2024, https://icml.cc/Conferences/2024/CallForPapers. See also Program Chairs for the International Conference on Machine Learning, “International Conference on Machine Learning (ICML) 2023,” icml.cc, 2023, https://icml.cc/Conferences/2023/llm-policy.

[2] Board for Certification of Genealogists, Genealogy Standards, 2nd ed. (Nashville: Ancestry.com, 2019), 36.

[3] Mohammad Hosseini, David B Resnik, and Kristi Holmes, “The Ethics of Disclosing the Use of Artificial Intelligence Tools in Writing Scholarly Manuscripts,” Research Ethics 19, no. 4 (June 15, 2023): 449–65, https://doi.org/10.1177/17470161231180449.

[4] Elizabeth Shown Mills, Evidence Explained : Citing History Sources from Artifacts to Cyberspace, 4th ed. (Baltimore, Maryland: Genealogical Publishing Company, 2024), 691–92.

[5] AIContentify Team, “The Ethical Dilemma of AI Writing Assistants: Balancing Authenticity and Automation,” AIContentfy, 26 July 2023 (https://aicontentfy.com/en/blog/ethical-dilemma-of-ai-writing-assistants-balancing-authenticity-and-automation : accessed 28 March 2024).

[6] Lindsey Passenger Wieck, “Revising Historical Writing Using Generative AI: An Editorial Experiment,” www.historians.org, August 15, 2023, https://www.historians.org/research-and-publications/perspectives-on-history/summer-2023/revising-historical-writing-using-generative-ai-an-editorial-experiment.

[7] Ibid.

[8] Gregory E Kaebnick et al., “Editors’ Statement on the Responsible Use of Generative AI Technologies in Scholarly Journal Publishing,” Hastings Center Report 53, no. 5 (September 1, 2023): 3–6, https://doi.org/10.1002/hast.1507.

[9] Jordan Boyd-Graber et al., “ACL 2023 Policy on AI Writing Assistance,” ACL 2023, January 10, 2023, https://2023.aclweb.org/blog/ACL-2023-policy/.

[10] Association for Computational Linguistics (ACL) Rolling Review, “Call for Papers,” ACL Rolling Review, 2024, https://aclrollingreview.org/cfp.

[11] “Citation, Documentation of Sources,” The Chicago Manual of Style Online, n.d., https://www.chicagomanualofstyle.org/qanda/data/faq/topics/Documentation/faq0422.html.

[12] Mohammad Hosseini, David B Resnik, and Kristi Holmes, “The Ethics of Disclosing the Use of Artificial Intelligence Tools in Writing Scholarly Manuscripts,” Research Ethics 19, no. 4 (June 15, 2023): 449–65, https://doi.org/10.1177/17470161231180449.

[13] University of Wisconsin – Whitewater, Department of Reference & Instruction, “Research, Citation, & Class Guides: How to Cite Generative Artificial Intelligence (AI),” University of Wisconsin – Whitewater, March 7, 2024, https://libguides.uww.edu/AI.

[14] Elizabeth Shown Mills, Evidence Explained : Citing History Sources from Artifacts to Cyberspace, 4th ed. (Baltimore, Maryland: Genealogical Publishing Company, 2024), 692.

[15] Gregory E Kaebnick et al., “Editors’ Statement on the Responsible Use of Generative AI Technologies in Scholarly Journal Publishing,” Hastings Center Report 53, no. 5 (September 1, 2023): 3–6, https://doi.org/10.1002/hast.1507.