1.04 Large language models (LLMs) and text generation

Large Language Models, frequently abbreviated as LLMs, represent a cutting-edge and exceptionally powerful category of Artificial Intelligence. These models are meticulously designed and engineered to comprehend, generate, and interact with human language in a manner that often appears remarkably fluent and human-like. The development of LLMs involves a process called “training,” where they are exposed to and learn from enormous quantities of text data. This training dataset can encompass millions of books, countless articles from diverse publications, a vast array of websites, and extensive collections of conversations, amounting to trillions of words. Through this intensive training regimen, LLMs learn the intricate patterns, complex grammatical structures, subtle semantic relationships, and even some of the nuanced stylistic conventions of how human language is used across myriad contexts.

It’s crucial to understand that LLMs don’t “understand” or “think” about meaning in the same conscious or experiential way humans do. Instead, they become exceptionally proficient at statistical prediction: predicting the most probable next word in a sentence, or determining how to coherently continue a piece of text based on an initial phrase, instruction, or “prompt” provided by a user.

When you engage with a sophisticated AI tool such as OpenAI’s ChatGPT (often utilising models like GPT-4o), Google’s Gemini (e.g., Gemini 2.5 Pro/Flash), or Anthropic’s Claude (e.g., Claude Opus 4/Sonnet 4), you are directly interacting with an LLM.  These versatile models can perform a wide range of language-based tasks. You can ask them complex questions on almost any topic, request clear and concise explanations of difficult concepts, get assistance in drafting professional emails or detailed reports, use them to brainstorm creative ideas for a project, have them summarise lengthy documents into key takeaways, and much more. For instance, if you provide an LLM with a prompt like, “Explain, in simple terms, the main steps involved in writing an effective Curriculum Vitae (CV) for a job application in the tech industry, highlighting common mistakes to avoid,” the model will leverage its training to generate a relevant, well-structured, and often highly insightful response, drawing upon the vast amount of information about CV writing and industry best practices it has processed.

LLMs offer a wealth of potential benefits and practical applications within the field of adult education:

• For educators: LLMs can serve as valuable assistants, helping to draft initial outlines for lesson plans, create diverse examples and case studies for learners, generate a variety of practice quiz questions with different formats, or write clear explanations of challenging topics tailored to different reading levels to suit the diverse needs of adult learners. This can significantly reduce preparation time, allowing educators to focus more on direct interaction and personalised support.
• For adult learners: LLMs can act as powerful personal learning aids. Learners can ask them for concise summaries of complex academic subjects, request customised practice questions to prepare thoroughly for an upcoming exam, obtain constructive feedback on their written work (such as a cover letter, an essay, or a project proposal), or seek explanations of difficult concepts presented in simpler, more accessible terms. This empowers learners to take more control over their learning process. The newest features of these models, such as enhanced reasoning for complex problem-solving or multimodal input/output for diverse content creation, are particularly relevant for adult education.

THE IMPORTANCE OF CLEAR PROMPTS (“PROMPT ENGINEERING“)

The effectiveness and relevance of the output you receive from an LLM are directly proportional to the quality of the instructions, or “prompts,” you provide. The skill of crafting effective prompts is often referred to as “prompt engineering.” A well-designed prompt is clear, highly specific, and provides sufficient context. It should ideally tell the LLM exactly what task you want it to perform, who the intended target audience for the output is, what style or tone of language to use (e.g., formal, informal, persuasive, technical), and any other important contextual details or constraints.

For example, instead of a vague prompt like “write about workplace safety,” a much more effective prompt would be: “Act as a health and safety officer. Write a list of five crucial workplace safety rules specifically for new employees working in a busy warehouse environment. The language should be simple English, easy to understand, and the tone should be friendly and encouraging. Include a brief rationale for each rule.”

Effective prompting is often an iterative process; you might need to try a few variations of a prompt and refine it based on the LLM’s responses to achieve the desired outcome.

THE “HUMAN-IN-THE-LOOP” IS ABSOLUTELY CRUCIAL

Despite their impressive capabilities, it is vital to recognise that LLMs are not perfect or infallible. They can sometimes make factual errors, provide information that is outdated or incomplete, or even generate text that sounds plausible and authoritative but is actually incorrect, misleading, or nonsensical – a phenomenon sometimes referred to as “hallucination.” Therefore, the principle of maintaining a “human-in-the-loop” is essential. This means that any content generated by an LLM, particularly if it is intended for educational, professional, or other high-stakes purposes, must be meticulously reviewed, critically edited, thoroughly fact-checked against reliable sources, and carefully refined by a knowledgeable human. This human oversight ensures the accuracy, relevance, appropriateness, and ethical soundness of the information. LLM-generated content should generally be viewed as a helpful starting point, a useful draft, or a source of inspiration, rather than as a finished, authoritative product to be used without critical evaluation and human judgment. This diligence is especially critical in contexts like the HER[AI]TAGE project, where cultural narratives and historical information must be handled with utmost sensitivity, accuracy, and respect.

Some of the most well-known and widely used LLMs currently include:

• ChatGPT (developed by OpenAI, often using models like GPT-4o): Celebrated for its strong conversational abilities, versatility in generating a wide array of text formats, enhanced voice interaction, video understanding, and its capacity for creative writing and problem-solving.
• Google Gemini (e.g., Gemini 2.5 Pro/Flash): Google’s flagship conversational AI service, capable of diverse text-based tasks, with strong integration with Google’s search capabilities and other services. It features improvements like “Deep Think” for enhanced reasoning and native audio output.
• Claude (developed by Anthropic, e.g., Opus 4/Sonnet 4): Noted for its proficiency in handling very long texts (large context windows), its focus on producing helpful, harmless, and honest responses through its “Constitutional AI” approach, and its strong reasoning capabilities, including “extended thinking” and tool use.

Developing the skills to use LLMs effectively, efficiently, and, most importantly, responsibly is rapidly becoming an indispensable competency for both adult learners aiming to enhance their learning and career prospects, and for educators seeking to enrich their teaching practices in today’s technologically advanced world.

PRACTICAL EXAMPLES

• An adult learner is diligently preparing for a challenging job interview in a new field. They ask an LLM: “Generate five common behavioural interview questions for a project manager role in the IT sector. For each question, provide an example answer using the STAR method (Situation, Task, Action, Result), focusing on demonstrating strong leadership and problem-solving skills.” They then use these detailed examples to structure and practice their own responses.
• An educator is developing a new module for the HER[AI]TAGE project focusing on the impact of climate change on local intangible cultural heritage. They prompt an LLM: “Act as an expert environmental historian. Summarise the key documented and potential impacts of the 1888 Drava river flood on the traditional agricultural practices and oral folklore of riverside communities in the Međimurje region. Present this in simple, accessible terms suitable for adult learners with no prior specialised historical or ecological knowledge. Highlight any lessons that might be relevant for contemporary climate adaptation discussions.” The educator then meticulously reviews, fact-checks, and refines this summary with expert sources before incorporating it into the module.
• A learner is working on improving their professional communication skills. They draft an email to a potential business partner proposing a collaboration. They then ask an LLM: “Review this draft email. Please provide specific suggestions on how I can make the tone more persuasive yet professional, ensure the value proposition is crystal clear, and strengthen the call to action. The email is intended for a busy executive who may not have much time to read it.”
• In a collaborative group setting for a community development project, participants are brainstorming innovative ideas for a local environmental awareness campaign targeting young adults. They use an LLM with a multi-step prompt: “First, suggest ten creative and low-cost activities our community group could organise to raise awareness about plastic pollution in the Mura river, specifically targeting individuals aged 18-30. For each activity, briefly outline the objective, target audience, and potential impact. Then, for the top three most impactful ideas, elaborate on potential challenges and suggest mitigation strategies.”
• A workplace trainer needs to create a comprehensive safety checklist for a newly installed piece of complex machinery. They ask an LLM: “Create a detailed, step-by-step operational safety checklist for the new ‘XJ5 Advanced Milling Machine.’ The checklist should cover pre-operation inspections (e.g., fluid levels, safety guards), operational safety procedures (e.g., proper material handling, emergency stop protocols), and post-operation shutdown and cleaning procedures. The language must be exceptionally clear, direct, and unambiguous, using industry-standard terminology where appropriate.” This AI-generated draft checklist is then rigorously verified and approved by certified safety experts and machine operators before implementation.
• An educator is designing an online adult learning course and wants to foster active participation. They use an LLM to draft a detailed rubric for assessing the quality and depth of learner contributions in an asynchronous online discussion forum. The prompt specifies: “Create a comprehensive rubric to assess adult learner participation in an online discussion forum for a course on contemporary European history. Include criteria for: 1. Timeliness and frequency of posts, 2. Quality of initial contribution (e.g., relevance, depth of analysis, use of evidence), 3. Responsiveness and engagement with peers’ posts (e.g., constructive feedback, asking clarifying questions), and 4. Respectful and professional interaction. For each criterion, provide detailed descriptors for three levels of achievement: ‘Exemplary,’ ‘Proficient,’ and ‘Needs Improvement’.” The educator then carefully adapts this rubric to align perfectly with the course learning objectives and expected standards.
• A researcher for the HER[AI]TAGE project uses an LLM to help categorise and tag a large collection of transcribed oral histories. They prompt the LLM: “Read the following interview transcript [transcript text inserted]. Identify and list the main themes discussed, key individuals mentioned, significant locations, and any references to traditional ecological knowledge or specific cultural practices related to river ecosystems.” This initial AI-assisted tagging then helps the human researcher to more efficiently navigate and analyse the collection.