That's a great question. Since it's the first in a long list, I'm going to give some definitions and context before giving a direct response. Skip what you already know and enjoy learning what you don't.

Artificial Intelligence (AI) is ubiquitous now. But here's the thing: AI isn't a new presence in our computer and communication systems — the messaging about its value has just been hyped up significantly over the past two years.

While a fountain pen and a good notebook can readily sit alongside a solid pair of boots for curious historians, digital technology is also a major part of a present-day historian's output. If that's true for you, you're likely already using AI in your research. AI is part of digital humanities (DH), a critical branch of humanities research. Output associated with DH draws on key insights from languages and literature, history, music, media and communications, alongside computer science and information studies. Combining these approaches and technologies into new frameworks produces a range of output — avatars, interactive maps, interconnected record-keeping, digitised records, chatbots, APIs — for public, professional, academic and amateur histories.

Recently, digital humanities has widened to include critical engagement with research processes such as machine learning, data science, and AI. It's worth noting that AI is a very broad term for a range of automated processes designed to identify patterns and make statistical inferences. AI models perform tasks or produce output that normally requires human intelligence, by applying machine learning techniques to large collections of data and identifying patterns within them.

For example, if you're monolingual or have limited capacity to learn other languages, you've likely used machine translation on records to better understand another language in the archive or an interview. Machine translation converts text from one language into another using a computer. There's a distinction between purpose-built translation tools (Google Translate, DeepL, Microsoft Translator) and general-purpose chatbots (ChatGPT, Claude, Google Gemini, Microsoft Copilot). A chatbot is a computer program that simulates human conversation, written or spoken, allowing researchers to interact with digital devices.

Google has used AI to shape search results for over a decade. The recent introduction of AI Overviews is changing how users receive those results. Programmed like a chatbot, AI responses often appear fluent, which increases users' trust in the answer. But after trawling the internet — and factoring in other parameters set by the programmer — these AI answers trend toward being incomplete at best and hallucinating at worst. In an AI context, a "hallucination" refers to a large language model (LLM) generating partially or entirely false answers, often supported by fictitious citations. AI's fluency makes these responses deceptive, because the inaccuracies, omissions, and additions aren't always obvious right away — and that's how AI-generated content ends up spreading mis- or disinformation.

So, after all of that: the answer is that you start using AI according to your own needs. If you're working with archival or library systems, maps, images, audio files, digitised material, or handwritten documents, there's a paid or open-access AI tool suited to it. You just have to find it.

I hear you. Paying subscription fees for numerous products to process data is expensive. And while there are paid versions (ChatGPT Plus, Claude Pro, Gemini Advanced), here we are — lucky enough to exist at a moment when companies are giving us access to this new technology for free. Why not use it?

Let's think about the future before we upload the past: what's at stake here? The "free" use of AI is like those "free" toiletries in a fancy hotel. The cost is there — it's just embedded somewhere in the pricing where you can't easily see it. In this case, you're paying for the use of AI (Gemini, Claude, Copilot) by uploading material for processing, whether that's pattern recognition, translation, or analysis. When you upload material for image processing, translation, image generation, or map-making — just to name a few — you're handing corporations information about yourself, or worse, about others who don't know you're sharing their information, ideas, or creations.

There are services, like Transkribus, that protect your data. But often those "free" tools gather data — personal, business, community, social — without clear structures around privacy, intellectual property, security, or digital governance. Before you upload material to a chatbot, ask yourself:

• Is it your information to share?

• Are you happy for the company to use that information in any way it sees fit?

• Are you using someone else's data ethically?

• Does paying for a license actually guarantee data protection, privacy, and security?

If you answered no to any of these, stop and rethink how to process the material. If you do want to use AI, read the data policy carefully before sharing historic information.

Transkribus is a program specifically designed to deal with the complexity of working with handwritten texts that have been digitised. AI is used innovatively on this platform to improve Handwritten Text Recognition (HTR — when a computer receives and interprets handwritten input from images) and Optical Character Recognition (OCR — the process of converting an image of text into machine-readable text). If you upload a digitised image to Transkribus and select the correct language and period, Transkribus says the processing time for historical documents, colonial records, diaries, letters, newspapers, and memos can decrease.

What makes Transkribus an excellent time-saver is that it can be trained by users and tailored to their particular needs, cutting back on processing time in the long run. Programmers, and some users, have trained the platform using Deep Learning and Neural Networks for pattern recognition — teaching the system to understand that certain shapes of ink represent particular letters, which form particular words, and so on.

Users need to acknowledge that Transkribus has bias. It's trained on a select range of already-converted images, and "high resource" languages — French, German, Dutch, and English — dominate the platform. The AI uses what it "knows" from previous uploads to transcribe your material, making it searchable, editable, and generally more workable. The catch is that your author (the person who created the original document) must have clear, conforming handwriting that the Convolutional Neural Networks (CNNs) and Recurrent Neural Networks (RNNs) behind Transkribus can recognise. If the AI can't process — "read" — the uploaded material, you do have the option to train it to do so. But that takes, you guessed it, time. And a lack of time is what got you here in the first place.

TL;DR: If you're working with colonial office records or standardised handwriting, use Transkribus to save time. If you're working with an author who has unique handwriting, or studying "low resource" Pacific languages, build extra time into your project plan — to locate a range of data, consult with community, and train Transkribus on "ground truth." Yes, "truth" is a contentious term for historians, but in this context it refers to information acquired through direct observation of the data. Ground-truth documents are human-verified transcriptions used as the gold standard for AI to learn from.

‍ ‍An avatar is an icon or figure representing a particular person in a video game, internet forum, or other virtual space. In the digital realm, an avatar acts as a bridge between the present and the past. While gamers know them as digital skins, in history they are embodied forms of archives. This electronic image can be moving or still, typically representing a figure — person, animal, or otherwise — standing in for a real-world entity. Avatars can also be manipulated by a computer user. If you play video games, you already know avatars as a virtual representation of self. If you're not a gamer, hang on to your notebook: avatars, with the right research, programming, and audience refinement, can be a fun way to share stories from the past.

Avatars have long been used to communicate stories from the past to a viewer. Because their foundational characteristic is representing a person, avatars are an appealing way to convey one person's experience of an event or period. They can be static — think of an image of a person with information delivered through a dialogue box — or they can incorporate movement, whether full-body or simply facial expression. Avatars with movement are more likely to be used for education. Unlike static images or simple animations, these avatars are designed to mimic human behaviours and expressions, making them appear more lifelike. An interactive avatar runs on two components: a Large Language Model (LLM) knowledge base, paired with system instructions.

Active and emotive avatars can be tailored to a particular curriculum goal, such as Indigenous history or ANZAC history. Creating avatars of specific people — especially Indigenous people — raises important ethical questions. When an avatar is based on a real historical person, consultation with community or descendants is essential. If you're using oral histories, diary entries, reports, or letters to shape an avatar's answers about someone's lived experience, you also need to consider intellectual property, copyright, and collaborative insight. Merging informed consent and ethical engagement with historical records with technical expertise is crucial to creating a good avatar.

Using virtual avatars for education — to deepen a user's engagement with the story being told — can be valuable. [Citation needed] research on non-realistic avatars in education programmes suggests they enhance curiosity, reduce social barriers, and foster playful learning, while realistic avatars promote empathy, relatability, and deeper emotional investment. High video quality, including the avatar's expressiveness, is key to this kind of storytelling success — realistic expressions from the avatar's programmers make users more likely to engage with the content.

The combination of virtual avatars and AI, especially chatbots, has changed how history is shared with particular audiences — school groups at primary, secondary, and tertiary levels, for example. A chatbot is a computer program designed to simulate conversation with human users, especially online. You've likely encountered one while shopping, logging a query with a service provider (electricity, gas, insurance), or doing online training. History and cultural education outreach avatars include Charlie the Virtual Veteran (Queensland State Library) and King David (ACTS Education — when you book a demo, ask if you can trial their Charles Bean avatar too). Rather than the avatar talking "at" the user, AI lets the user ask the avatar questions about their experiences in the past.

For example, an avatar of Emily Caroline Creaghe should base its responses (and I hope more avatar creators hire historians for content creation) on primary sources — her diary — and secondary sources, depending on the associated researcher or historian's choices. These documents set the parameters for what the avatar knows. It's worth remembering that an avatar is only as good as the research behind it.

A good avatar also needs firm guardrails. These directives limit what the avatar will engage with — it shouldn't, for example, discuss sexuality (there's no data on this, and it's not the focus of a lesson plan about women explorers) or respond to prompts asking it to assume the persona of a lead singer in a Riot Grrrl band (a chatbot's priorities can shift without carefully applied guardrails). Guardrails let the chatbot gently steer the user back to the task — an in-depth discussion, grounded in primary and secondary sources, of a female explorer's life in the late nineteenth century. Clear guardrails preserve the historical integrity of the persona and prevent "jailbreaking" (in this case, turning the avatar into a singer). These boundaries also reinforce to the user that they're not speaking with a "person," but a simulation of one, based on specific records.

Firm guardrails can't solve every issue with AI-powered avatars, though. The underlying LLM can still hallucinate — inventing information to satisfy the chatbot's drive to give the user a plausible answer. That's a real risk with avatars. But I also see it as an opportunity: it lets users learn about the limitations of AI avatars through the lens of archival gaps and silences — an excellent discussion point for history students and the historically curious, and a way into critiquing both the past and the tools we use to tell it.

Hopefully I've answered this week's question. Got one of your own? Send it to colourfulhistories@gmail.com.

Deborah

This one's actually her strongest piece — real citations, a concrete case study with specific numbers (130 changes down to 70 per page), and it does real intellectual work explaining why supervised learning matters for historians specifically. The content and argument are excellent. But the mechanical issues are still there, plus one factual slip:

Notable issues:

• "Sarah Conner" — should be Sarah Connor (Terminator). Minor, but it's a factual/spelling error in a specific reference, not just a typo.

• A genuinely tangled run-on: "over time the general model will struggle with the nuances and variability of a person writing in a diary over the course of 12 months uniform spelling, letter shaping and sentence structures are rarely conforming" — this needs to be split into two sentences; as written it's hard to parse.

• A broken sentence: "What we want to process such records with something like Transkribus, an application that uses..." — missing "to" (should be "To process such records, we want something like Transkribus").

• "giveaway" should be "give away."

• "so to reassure" should be "but to reassure" — as written it doesn't logically connect to the clause before it.

• Inconsistent dash spacing and a stray comma ("The ability of machine learning, offers an opportunity").

Here's the cleaned-up version:

This is an excellent question! There are a few stages to the response: what machine learning is, some examples, and then the historian's role itself. Grab your beverage of choice — hopefully you'll be inspired to get involved with supervised learning, or in discussions about the importance of critical humanities to model development, after reading this.

To understand supervised learning, we first need to understand machine learning. Researchers have developed algorithms — termed machine learning — that imitate, and sometimes exceed, human cognitive ability at complex tasks in the history profession. Machine learning refers to an algorithm in which a computer recognises patterns and relationships between variables based on given data. Each algorithm develops a model to output an answer to a specific problem.

There are currently three types of machine learning: supervised, unsupervised, and reinforcement learning. In history, there are no universal or regional guidelines yet for using AI models in the workplace, unlike, say, in medical research. Historians aren't averse to guidelines — just as we follow archival protocols to ensure integrity, we could use supervised learning as an opportunity to "label" data so others can recognise our professional standards. We need guidelines for using AI models not to give away our expertise and be replaced (Sarah Connor taught me better than that), but to reassure our readers and peers that our research follows particular standards when using digital tools. By subjecting models to supervised learning, we ensure the historiographical tradition, and its conceptual styles and frameworks, remains recognisable to others.

Supervised learning is when the researcher modifies content by tagging information to create a "ground truth" document. While the concept of "truth" is contentious for historians, in this context "ground truth" borrows a computer science definition: the document gives the AI program an authoritative baseline to work from, providing a sense of what counts as a correct or incorrect output. For computer scientists, if we're using a handwritten diary, the ground truth is the accurate transcription of its entries — historical truth remains a separate debate for historians. From this ground truth, AI runs training sessions behind the scenes and learns to make predictions, which are then applied to a digitised image of text, such as a diary.

In the world of historians (welcome, 'tis a fine place to be), machine learning can be an option when confronting huge swathes of information about a topic, event, or time period. Machine learning offers a way to bring those beautiful, complex, heart-pounding piles, metres, stacks, boxes, and tied bundles of manuscripts — or gigabytes and terabytes of digitised archive and library images — down to a manageable scale. Right now, the models most relevant to historians' work focus on palaeography (the study of old handwriting), to improve access to cursively handwritten documents such as diaries, letters, and colonial-office records.

Before we get to realising the (impossible?) dream of wrangling copious amounts of historical data in a timely manner, we need to build a model specific to a particular collection. Why? I could take a model like Claude, Gemini, ChatGPT, or even the specialised ChatGPT model "Historian's Friend," give it a digitised image of 19th-century diary handwriting, and prompt it to "translate, transcribe, contextualise" — and it would. Wonderful. But the result might not be usable, because these are generative, probabilistic models designed to give an answer through pattern recognition. Because of the chatbot's fluency, I'd get a response that sounds and reads very nicely. It would make sense. But — and this "but" is why I'm so critical of such models — over time, a general-purpose model struggles with the nuances and variability of a person writing in a diary over twelve months. Spelling, letter shaping, and sentence structure are rarely uniform. I work with collections created by people across a range of institutions — missions, schools, libraries, maternal health care — and their writing presentation and style vary accordingly. That means we need a model that can handle the unique elements of these records, which is what prompts the creation of a model using supervised learning. To process such records, we want something like Transkribus, an application that uses Handwritten Text Recognition (HTR), a specialised form of supervised computer vision, to process handwritten documents.

This supervised learning process requires historians to correct and tag documents, giving the model clear parameters. For instance, to create a document suitable for training an AI model in Transkribus, I took the following steps:

1. Have a "supermodel" (an epic, pre-trained model containing millions of words and "knowledge" of character images) process a digitised image of a diary.

2. Compare the diary page's content with the AI output to correct presentation errors, while preserving original spelling mistakes and variations. For this pilot project, producing "ground truth" for the Strathfieldsaye diaries required an average of 130 changes per page.

3. Mark up or tag particular names, places, or dates of interest, to show the model what information matters for later data mining (an AI process that uncovers patterns and other valuable information from large datasets). The AI tries to reproduce what it "sees" as accurately as possible — when I see "rain," AI sees "Ram." With faded ink, digitisation artefacts, and individual handwriting quirks, that's not "wrong," exactly, but it is woefully inaccurate. This is why supervised learning is so important to producing usable digitised records.

Once steps like these were completed, the marked-up material was uploaded to Transkribus to build a model suited to this particular series of records. In a pilot project for the University of Melbourne using the Strathfieldsaye station diaries (1872–1875), the best results came from correcting and tagging around 50 pages of data — approximately 24,000 words — for model training and supervised machine learning.[i] The correction rate needed almost halved, to around 70 changes per page, down from the initial average of 130.

It's worth noting that in these cases, the AI isn't trying to capture spelling and writing conventions — its primary goal is to reproduce an author's writing style accurately. Correction and tagging are essential to supervised learning, and to exporting trustworthy data for collection discovery, data mining, and metadata creation.

Supervised learning of AI by historians serves a dual purpose: training more accurate AI models opens up collections to a wider range of users, while archivists and historians continue to safeguard historical records — and address long-standing historiographical concerns about the gaps and silences within archives.

By taking part in supervised learning, historians support multiple organisations and users of historical records. We can help librarians and archivists tackle the sheer scale of data their institutions face.[ii] If historians are excluded from — or reject involvement in — supervised learning, AI-augmented output addressing historical content is more likely to contain content and contextualisation errors. Our involvement helps ensure the standards and practices of our discipline are consistently revised, adapted, and practised in both the analogue and digital worlds we live in.

Hopefully I've answered the question.

Got a question about history and the digital humanities? Send it to colourfulhistories@gmail.com.

Deborah

[i] Disher, Harold Clive, Strathfieldsaye Estate Diary February 1872–1875 (February 1872–31 October 1875), [UMA-IT-000147344]. University of Melbourne Archives, https://archives.library.unimelb.edu.au/nodes/view/637676, accessed 28 October 2025.

[ii] National Archives Australia, 'Information management: Outsourcing digital storage', National Archives of Australia, 2025, https://www.naa.gov.au/information-management/storing-and-preserving-information/storing-information/outsourcing-digital-storage; Indigo Holcombe-James, '"I'm fired up now!": digital cataloguing, community archives, and unintended opportunities for individual and archival digital inclusion,' Archival Science, 22 (2022), 521–538, https://doi.org/10.1007/s10502-021-09380-1; Hider, P. (2024). 'At a Crossroads: Cataloguing Policy and Practice in Australian Libraries.' Journal of the Australian Library and Information Association, 74(1), 53–72, https://doi.org/10.1080/24750158.2024.2403165.

This one's more polished than the others overall — the jokes land, the structure holds together well, and the citation examples are genuinely useful. Fewer major issues, but still some rough patches:

• GLAM is used without ever being spelled out (Galleries, Libraries, Archives, Museums) — fine for insiders, but worth defining once for general readers.

• A tangled closing sentence: "By recognising the complexity of AI, accounting with authority how we are using it in professional and academic research, we have made an important step..." — "accounting with authority how" isn't quite grammatical; needs "accounting, with authority, for how."

• Minor formatting inconsistencies: stray double spaces, inconsistent bullet dash spacing, "systems' security settings" bundled oddly into a longer list of things to check before starting.

• "I am acknowledging that AI can be used..." — slightly stiff phrasing, reads more naturally as "AI can also be used..."

• A few places where sentence rhythm is uneven from run-on clauses ("While there is potential to keep doing historical research and writing without AI, a little bit of careful and critical use can be a great benefit" is fine but could be tighter).

Here's the cleaned-up version:

The answer to this question begins with "why should I document my use of Artificial Intelligence in my research?" Historians love showing off their primary and secondary sources — that's a core focus of any first-year introductory class. Show me the evidence! I want to evaluate it for myself.

Historians don't seem so keen on showing their methods or tools, though. That's usually a line or two in the introduction, maybe a conference paper, and then it's on to (trumpets, please) THE ARGUMENT.

Nice. I love a good argument (argument: the collective noun for historians), and arguments are important — but we can learn a lot from sharing our methods too.

The ubiquity of AI in our laptops, writing programs, recording equipment, and visual programs means we should be actively seeking ways to show authenticity and authority in our research. We need to distinguish our own work and ideas from AI output, which is a researcher's tool, not a substitute for one. I'm not saying AI should be used to generate content — an introduction or an argument in a journal article, blog post, or grant proposal. There's real concern about plagiarism and AI, and that concern is warranted. But generating whole papers isn't the only thing AI can do. AI can also be used effectively to transcribe handwritten material, process an oral history transcript, generate metadata for an archival collection, create a map for a piece about exploration, or produce an image of written material from a diary. Institutions like Te Papa and the University of Melbourne Archive are testing AI to improve access and processing times in their collections. They're not implementing AI in their workflows yet, but the potential for future use is there. When our GLAM institutions (galleries, libraries, archives, and museums) do adopt AI, historians should be able to account for it and explore its impact — if any — on our research.

We need to hold ourselves, and others in our discipline, to certain standards. As always: before you start, check your system's security settings and your institution's policies, and make sure you have the required permissions to work with the material on whichever AI platform you've chosen. Some people in our digital world are engaging in data fabrication, intellectual property theft, plagiarism, and violations of data sovereignty — but that doesn't mean historians have to. Part of our responsibility as educators and researchers, when we do use AI, is being clear and precise about the platform we chose, the prompts we used, the output we got, and our own critical, specialist response to that material.

One way to demonstrate disciplinary authority and authenticity is through references (sigh — I love a good footnote). There's still plenty of scope to keep doing historical research and writing without AI, but careful, critical use can be genuinely valuable — and we can pass that knowledge on to our students, improving their digital literacy and helping them push back against the unchecked spread of AI use.

Citation standards are an important and underused tool for regulating AI use. Carefully sidestepping the endnote/footnote debate to give one example: the Chicago Manual of Style does have guidelines for referencing AI use in academic writing.

A footnote might look like this:

1. Text generated by ChatGPT, OpenAI, March 7, 2023, https://chat.openai.com/chat.

Note that CMS is assuming here that you used ChatGPT to generate content outright. In some journals, this could read as ChatGPT standing in as "author" of the content. If you've denied yourself the fun (read: hair-pulling) of actually writing the thing, you'll want to know that many publishers — journals and books alike — won't recognise AI as an author. Best to delete that paragraph or sentence and try again.

I have faith in you.

Say you used AI as part of your research method — for example, to create a transcript from handwritten material. You can note this in your methods section and cite the AI use in a couple of ways:

Sample acknowledgement (in-text):

Super AI-5000 provided translation assistance for 34 archival documents from Spanish to English, representing roughly 40 per cent of the translation work. All AI translations were reviewed by the lead author, with particular attention to historical context, legal language, and culturally specific concepts. Approximately 25 per cent of AI translations required substantial revision to maintain accuracy.

Sample acknowledgement (footnote):

Super AI-5000. OpenAI. Accessed 15 July–22 August 2024. Used for initial interview transcription: Spanish-language materials (8 hours of audio). Human verification: complete manual review and correction of all transcripts against original recordings.

To make these references comprehensive, include:

• Specific AI tool names and versions used (e.g., "ChatGPT-4, GPT-4o, Claude 3.5 Sonnet") and the date range of use.

• Specific tasks performed by the AI tool(s) — translation assistance, transcription support, literature review support, editing assistance (limited to wording or formatting changes; excluding generative editorial work and autonomous content creation).

• The extent of AI involvement in these tasks, plus a discussion of how human oversight and verification were applied at every stage of AI use. Keeping a research diary from the start of your project will make this much easier down the line.

For multi-author manuscripts where AI use varies between team members, individual contributions should be specified. Where the same tools and workflows were used collaboratively, a single consolidated statement is enough. One word of caution: on interdisciplinary projects, have the conversation about AI use before the hard work begins. In some disciplines, having AI summarise and write up findings is common practice. In others, it's a career-ending move.

AI, generally speaking, has a lot of problems. By recognising that complexity, and accounting with authority for how we're using AI in professional and academic research, we take an important step toward holding the broader research community to a higher standard of transparency.

Want to talk about this post further, or have a question about history practice, digital humanities, and/or AI?

Email me at colourfulhistories@gmail.com. I'd enjoy the chance to continue the conversation.

‍Welcome back to another instalment of Historians and AI! This week's question: what can AI do to synthesise written material produced over the course of a career? Can AI locate the methods, and the changes in method, used over time across a range of written materials?

Gosh, this was a great question to work through with a colleague. At first glance it looks like it's about summarising material and picking a product — but it's bigger than that. It touches on methodology, disciplinary norms, secondary sources, security, privacy, open access, generative AI, agents, chatbots, and prompt design. Let's work through the steps, and temper our curiosity about AI along the way.

1. Create a project abstract

The first step is writing a project abstract covering your research topic, argument, and goal. What do you want to achieve by integrating AI into your workflow? Writing this out will help you identify the platform, plug-in, or app you actually need. Is the task better done manually — for small-scale quantitative data or small transcription projects, say? Large Language Models generally need large amounts of data to work well.

Remember: AI is like an archive, or an oral history — it is not neutral. It's a tool that can help with research, but the platform you choose has to align with historical method and your sub-discipline's norms.

2. Question parameters and obligations

Taking the time to write the abstract also helps surface ethical and capacity issues. Do you own the copyright to everything you're thinking of uploading for summarisation? You may also be looking for a program that summarises PDFs — one of the most hyped options is TLDR. If you're a historian, this platform isn't ideal for you: TLDR was trained to summarise computer science papers, which use different disciplinary language and conventions than history does. If you want to summarise history articles, look for a platform trained on history articles specifically. Historians rely on narrative argumentation, implicit causality, and fragmentary evidence — disciplinary characteristics that LLM summarisation tends to flatten.

3. Assess your materials

There are more questions to work through before uploading anything:

• What kind of documents are you using — PDF, Word, video, audio, websites? Some platforms handle multiple formats; others handle just one.

• Can you reshape the source to fit the platform's requirements? PDFs may need breaking into smaller chunks before being fed into the AI. Some AI-driven platforms struggle to process even small amounts of information well enough for an undergraduate class discussion.

• Does the platform run online or offline? Some AI tools only work online — but closed-loop, offline AI does exist. If you're working with sensitive data, that may be the better choice.

• Do you own the copyright on these documents?

• Is the platform secure? Is there evidence of encryption? Where are the servers based? These answers will shape how you handle governance, privacy, and security — governance over AI in the EU, for instance, is quite different from the US.

• How will your data be used after uploading? Yes, you have to read the terms and conditions to find out. Yes, I am the type of geek who reads T&Cs.

• How long will the platform retain your information, and who decides when it's deleted?

Once you've worked through the documents, security, privacy, copyright, and ethical questions for your project, you're ready to choose a platform to process the material.

4. Choosing and using a platform

AI assistants

You might prefer using something like Claude or Gemini to summarise material. In that case, you'll need to know how to write effective prompts — taking the time to build custom prompts noticeably improves the quality of the summaries you get. This is where your abstract pays off again. You'll need to define:

• Context: your research topic and objectives.

• Role: the AI's persona (e.g., "act as a research methodologist").

• Methodology focus: whether you need qualitative, quantitative, or mixed-methods guidance.

• Output format: the format you want (table, bullet points, etc.).

Once you've set those parameters, tell the model what you actually want it to do. Some sample prompts:

• Detailed extraction: "Read this [article/abstract] and identify the specific research design, data collection, and data analysis techniques used. Present the findings in a structured table."

• Methodological comparison: "Compare the methodology of [Paper A] and [Paper B] regarding their approach to [e.g., oral histories]. Highlight which study's approach is more robust and why."

• Limitation analysis: "Analyse the methods section of the attached paper. Identify potential methodological limitations. Identify the primary source base used and how the author addresses gaps in the archival record."

• Finding the "why": "Based on this article, explain the rationale behind using a [e.g., qualitative analysis] approach instead of a quantitative survey."

Agents

Maybe you're looking at your reading pile and thinking, no, Claude and co. aren't for me — in which case, you might consider agents instead. An AI agent is a system or program capable of autonomously performing tasks on a user's behalf, by designing its own workflow and drawing on available tools. AI agents can go well beyond natural language processing, into decision-making, problem-solving, interacting with external environments, and executing actions.

You could build an agent to process your material under strict parameters for research and output. Personally, I don't recommend this: agents still tend to lack precision, and their security processes are often weak. Historians should favour tightly constrained, inspectable workflows over autonomous agents — historical argument depends on interpretation, contextual judgement, and an awareness of what sources can't tell us, and agents aren't yet well equipped for that. Workflows where you stay in control of each step are the safer bet.

Platforms

A few platforms are built to summarise a range of genres and formats:

• FileReadyNow condenses PDF files into concise summaries.

• Genei.io helps you find useful sources for your topic, or upload your own webpages and PDFs. Documents can be organised into projects and folders, and its AI extracts key information from articles instantly.

• NotebookLM (Google) lets you upload PDFs, websites, YouTube videos, audio files, Google Docs, and Google Slides; it summarises them and draws connections across topics, powered by Gemini's multimodal capabilities.

• Scholarcy.com lets you bulk-import files in any format to quickly build a collection of summaries, with interactive highlights to extract key facts and findings. Unlike TLDR, it's trained to respect general academic writing conventions, though it may struggle with older styles of writing compared to 21st-century articles. It's built on machine learning trained mainly on academic articles and individual book chapters — it's not a writing tool and won't write your lit review for you, but it can help you read and synthesise a collection of papers and structure your thinking beforehand. It can also generate flashcards from video, though results vary depending on the transcription quality and content.

• LightPDF AI Summarizer is an online AI-powered tool for editing, converting, OCR, signing, annotating, and chatting with PDFs, accessible across desktop, mobile, and web. Key features include instant summarisation of lengthy documents, AI chat with PDFs for asking specific questions, and broader functionality like OCR and format conversion.

All of these platforms come with the same caveat: their output tends to flatten historiographical debate, argument, and case study nuance. If you want broad trends, use the AI. If you want the unique perspective and rich detail, plan the time to do it yourself.

Hopefully that walkthrough gives you a clearer sense of whether to use AI for summarising and synthesising large bodies of work.

You can keep the conversation going, or submit a new question for the series, by contacting me at colourfulhistories@gmail.com.