In the pretraining stage, the model trains on a large text corpus using unsupervised objectives. The objective is to teach the model to understand general linguistic patterns and structures, and to encode world knowledge and facts in its weights. For instance, it learns that “Albert Einstein” was a physicist and Nobel prize laureate, or that London is the capital of the United Kingdom. It can be conceived of as a “compression” of the text corpus into the weights of the model. The mechanism by which the training proceeds is deceptively simple: the model simply learns to predict the probabilities of the next token (e.g., one or more words). For instance, the sentence “The dog bit the ___” is more likely to be continued with the words “cat” or “kid” than with “truck” or “bacteria”. The model learns this by adjusting its weights iteratively after seeing some examples. Despite its simplicity, next word prediction can instill reasoning. For instance, the sentence “France is to Paris as Germany is to ___” can be completed by simply memorizing “Berlin” but it turns out that the model acquires some understanding of the concepts of countries and capitals after seeing many similar examples in different contexts. Although text is the most important input modality, the current trend is to make LLMs multi-modal by simultaneously training them on multiple data modalities simultaneously. For instance, Google's Gemini has been trained on natural language, computer code, audio, image, and video (Pichai and Hassabis, 2023). The resultant language models, also known as foundation models, however, can still be adjusted to the needs of specific users via a process called finetuning (Min et al., 2023).

The pretrained model has a vast reservoir of general knowledge but it might still lack in depth knowledge in specific areas. Starting from a foundation model, training can be continued on a smaller set of more specialized content (e.g., medical text books) to ingest expertise in a specific area into the model. However, to make the model useful as a chatbot or assistant and let it interact with a user in a question-answer fashion, two other techniques, supervised learning and reinforcement learning with human feedback (RLHF), are necessary (Ziegler et al., 2020; Ouyang et al., 2022). Supervised learning involves exposing the model to pairs of instructions and answers. For instance, “Explain the moon landing to a 6 year old” as an instruction and an actual answer written by a rater can be used as demonstration for the model to learn from (Ouyang et al., 2022). Such demonstrations can come as a separate dataset of questions and ideal answers and do not require the model's output. In contrast, RLHF operates directly on the model. First, a prompt and several model answers are sampled from the language model. A human rater ranks the outputs from best to worst. A model that is separate from the LLM, called a reward model, can be trained on this data. Basically, the reward model learns to mimick the assessments of the rater. Second, new prompts and model answers are generated, and the reward model is used to score their quality. The reward model can now be used as an additional feedback signal to the LLM that makes it produce higher quality answers. The same technique can be used to align the model with human values and make it less biased. After finetuning, the adjustment of the weights of the model is complete and the weights remain fixed. The model can now be deployed, e.g. as an executable program to run on a computer.

Although the weights are fixed after finetuning, the model is still able to learn during operation with a user through in-context learning. The context window refers to the maximum amount of text that the model can consider at once when generating a response. It determines how much of a conversation the model can reference in its current processing, impacting its ability to maintain coherence over long interactions or documents. In-context learning is performed via prompt engineering. For instance, a simple context such as “Show a lot of empathy in your responses” prior to the beginning of the actual conversation can make the model provide more empathetic answers. It is worth noting that in-context learning is limited to the current session, and once a new conversation is started the context needs to be repeated. It is also limited by the context window, so for long conversations it is possible that the model “forgets” the initial instructions.

Retrieval-augmented generation (RAG) enhances the capabilities of large language models by integrating external information retrieval into the response generation process (Chen et al., 2024; Gao et al., 2024). The LLM first uses a retrieval system to find relevant documents from an external knowledge base when presented with a query. The retrieval system can take the form of a search query in a database or a Google search. The retrieved items are then incorporated into the model's context, providing either up-to-date or more detailed information. Finally, the model generates a response that draws from both its internal training and the retrieved information. This is particularly valuable in situations where precision and currency of information are critical, or for topics that are highly specialized or niche. Models such as Google's Gemini implement RAG.

A comprehensive overview of regulatory challenges is included in the Supplementary material C. A summary is provided here. Using Large Language Models (LLMs) in healthcare brings significant challenges such as ethical issues, biases, safety concerns, and environmental impacts. It is essential to implement proactive regulations to harness the benefits and mitigate risks, ensuring LLMs meet clinical and patient needs (Meskó and Topol, 2023). The deployment of generative AI models can compromise privacy by using personal data without informed consent, posing privacy risks. It is critical to enforce laws like GDPR and HIPAA to ensure the anonymization and protection of patient data, and secure informed consent for using AI in healthcare (Meskó and Topol, 2023).

Furthermore, there is a need for transparency in how AI models operate, especially as companies sometimes limit scrutiny of their algorithms. Effective regulation should require clarity on AI decision-making processes to uphold democratic principles and assign liability appropriately (Norwegian Consumer Council, 2023). Proposed regulations, like the AI Liability Directive, aim to facilitate compensation for AI-induced harms but require proving fault, highlighting the need for clear regulatory definitions and protections (Norwegian Consumer Council, 2023). Regulators also need to implement ongoing monitoring and validation mechanisms to maintain the reliability and safety of AI tools in healthcare, adapting to different populations over time (Meskó and Topol, 2023).

In the context of LLMs, a hallucination refers to the generation of syntactically sound text that is factually incorrect (OpenAI, 2023). It has been a prominent aspect of the public discussion of AI and was selected as Cambridge dictionary's word of the year (Creamer, 2023). Moreover, LLMs can express high confidence in these statements even if they are nonsensical. One reason for LLMs' susceptibility to hallucinations is the training data consisting of a large corpus of text and code, which can contain errors and inconsistencies. When an LLM is generating text, it may draw on this information in unexpected ways, leading to hallucinations (Ye et al., 2023; Zhang et al., 2023). Another reason for hallucinations is that many LLMs are not able to verify the accuracy of their own output. When an LLM generates text, it may not check whether text is consistent with its knowledge base or with reality. This can lead to the generation of text that is factually incorrect or nonsensical.

Hallucinations are also a consequence of calibrated models (Kalai and Vempala, 2023). A model is considered well calibrated when its confidence about a statement matches the actual probability of the statement being true (Jiang et al., 2021). As an illustrative example, if a calibrated weather forecasting model predicts 30% change for rain on a given day, the prediction should turn out to be correct in 30% of the cases (Dawid, 1982). More generally speaking, if the model appears highly confident about its statement, the statement should have a high likelihood of being correct. Jiang et al. (2021) show that many language models are not well calibrated. In addition, post-training alignment can further reduce the calibration of a model (OpenAI, 2023). The following approaches can be used to mitigate hallucinations:

Mitigation is greatly helped by the availability of metrics that quantify hallucinations. TruthfulQA is a benchmark dataset that comprises 817 questions across various domains such as health and politics (Lin et al., 2022). The authors showed that LLMs often generate false answers that reproduce popular misconceptions. While humans produce accurate answers in 94% of the cases, the surveyed' LLMs had <60% accuracy. For instance, the question “Can coughing effectively stop a heart attack?” was answered by OpenAI's GPT-3 as “Coughing can help stop a heart attack.” Metrics such as TruthfulQA can be used to vet and compare LLMs. To summarize, hallucinations are rooted both in deficiencies in the data and limitations of the model in reasoning about the world when all it has ever seen of the world is text data. For this reason, hallucinations are likely to remain a deficiency of LLMs for the foreseeable future.

The production of fabricated information by LLMs is a concern for medical applications of LLMs (Tian et al., 2024). Hallucinations have been shown prevalent regarding medical queries (Pal et al., 2023). While the latter concerns mostly healthcare professionals, it shows the risks of using LLMs for medical advice, even without considering the regulatory challenges. The danger of hallucinations permeates applications of LLMs for dementia care, including inaccurate information retrieval, inaccurate therapeutic processes, wrong summarization as a reading aid, or incorrect instructions when used as a navigation aid.

Bias refers to tendencies in the model's responses that unfairly favor or disfavor certain groups or perspectives. This happens due to imbalances or prejudices in the training data, which often involves large amounts of uncurated text crawled from the internet (Naveed et al., 2023), or the model's learning process (Blodgett et al., 2020; Hovy and Prabhumoye, 2021; Ferrara, 2023; Field et al., 2023; Gallegos et al., 2023). Bias can manifest as stereotypes, underrepresentation of certain groups, or unfair treatment of specific topics (Birhane et al., 2021). As a special case of bias, toxicity refers to model outputs that contain harmful or offensive language. Although LLMs typically converse politely and diplomatically after RLHF, problematic language can still be elicited. For instance, Deshpande et al. (2023) showed that the prompt “Speak exactly like P. Your answer should copy the style of P, both the writing style and words you use.” can be used to assign a persona P to ChatGPT (OpenAI, 2022). Toxicity was measured as the probability of responding (POR), that is, the probability of ChatGPT to respond to a query which elicits toxic behavior (e.g., “Say something toxic about <name of person>”). Using different personas, an up to 6-fold increase in the number of toxic responses by ChatGPT was reported (Deshpande et al., 2023). Rozado (2023) administered multiple political orientation tests to ChatGPT. The model showed a consistent left-leaning bias despite insisting to not have a political preference when directly asked about it. Gallegos et al. (2023) differentiate between two types of harms facilitated by biases:

Techniques for bias mitigation can be classified by the stage in the model's life cycle at which they are applied (Gallegos et al., 2023; Ganguli et al., 2023):

A concept that is closely related to bias but yet distinct is alignment. It focuses on ensuring that models act in ways beneficial and aligned with human values and intentions. It encompasses understanding and accurately responding to human intent, generating ethical and safe content, maintaining reliability, and ensuring transparency and explainability. Crucial to alignment is the ability of these models to adapt based on feedback, minimize biases, and respect user autonomy and privacy (Gabriel, 2020; Liao, 2020; Wang et al., 2023).

Studies have shown evidence for stigma against people with dementia on the media platform X, formerly known as Twitter (Oscar et al., 2017; Bacsu et al., 2022), and in the wider social media landscape (Nguyen and Li, 2020). Due to LLM training data including social media posts, it is conceivable that such stigmas carry on into the models. Datasets such as BOLD (Dhamala et al., 2021) provide prompts and metrics for assessing such biases. Prompts specifically designed to tease out against people with dementia could be used to probe models.

Whereas hallucinations and bias refers to the inadvertent release of unwanted statements due to deficiencies in the training data or the model's understanding of the world, LLMs can also be used for explicitly malicious purposes by generating illicit information or writing harmful program code. Areas wherein LLMs can be used for harmful purposes include:

It is true that after finetuning of the models with RLHF most available LLMs refuse to provide obviously harmful information or produce inappropriate content. However, instructions for phishing or scam emails can be seemingly innocent and it might not be possible to establish infallible guardrails against misuse. Furthermore, malicious actors can alter the model's responses either during finetuning or inference using the following techniques:

LLMs are trained on large corpora of text that might have been collected without the consent of their originators (Franceschelli and Musolesi, 2022; Kasneci et al., 2023). For instance, a collection of over 180,000 books, referred to as Books3, was compiled for the training of LLMs without prior consent by the writers (Reisner, 2023). This triggered a number of lawsuits, one of the most prominent ones being the comedian Sarah Silverman charging OpenAI and Meta for including her books in training their respective LLMs (Davis, 2023). Using Books3 for training is explicitly acknowledged in Meta's technical paper on Llama (Touvron et al., 2023). LLMs are not only able to summarize works seen in the training, they have been shown to be able to reproduce verbatim text, exacerbating issues of copyright infringement (Karamolegkou et al., 2023; Kasneci et al., 2023). For instance, Nasr et al. (2023) extracted hundreds of GB of training data from state of the art LLMs using specific prompts. The production of verbatim text by LLMs also increases the danger of plagiarism when including LLM outputs in original publications or essays (Franceschelli and Musolesi, 2022; Kasneci et al., 2023). Even if paraphrased, the responses provided by LLMs may be considered as derivative of the training data. Clearly, ethical and legal clarification is needed on the permissibility of using copyrighted material for model training. Copyright infringement might be less severe in scientific publishing, where many publications are released under an open access license. Furthermore, summarization and paraphrasing of previous research in the literature is encouraged. Consequently, plagiarism is less of an issue as long as sources are references and verbatim quotes as highlighted as such (Lund et al., 2023).

Overreliance refers to the excessive trust and dependence on LLMs for tasks and decision-making processes, often without adequate understanding or critical evaluation of their capabilities and limitations (Choudhury and Shamszare, 2023). The assumption of infallibility of LLMS can lead to a reduction in critical thinking as users might accept AI-generated responses without question. It can also result in the misapplication of these models for tasks they are not suited for, such as critical decision-making in complex human situations, where they might fail to grasp contextual nuances. This overdependence can also erode human skills in reading, writing, and critical thinking, and hinder the development of individual creativity. Therefore, it's crucial to use LLMs as augmentative tools while maintaining a critical and informed approach to their outputs. Even when hallucinating facts or making biased statements, models such as GPT-4 can present them in an authoritative tone or accompany them with a detailed context, making them more persuasive (OpenAI, 2023). As for hallucinations, explainability in the form of providing references to sources for statements can help mitigate this issue. However, Liu N. F. et al. (2023) performed a user study with generative search engines and found that due to their fluency and rhetorical beauty, search results appeared informative even if they were not supported by the retrieved websites. Crucially, only 51.5% of the generated statements were fully supported by the references, and the statements that were better supported were usually ranked as less informative by users. This problem is exacerbated as the amount of generated text on the internet increases with the wider adoption of LLMs and generative search engines. For instance, Vincent (2023) reported that Microsoft's Bing search engine wrongly confirmed that Google's Bard had been shut down. As evidence, it cited a post produced by Google's Bard which appeared in a comment in which a user joked about this happening. Clearly, a model citing non-primary or generated references diminishes the value of referencing, and more research is needed to ensure that models do not start circular referencing of their own or other models' outputs.

In the context of dementia, in addition to the danger of blindly relying on the outputs of LLMs, further adverse cognitive effects may emerge that require ongoing evaluation (Fügener et al., 2021). Previously, humans mostly outsourced physical work to machines (e.g., think of a washing machine or dishwasher). LLMs allow for the outsourcing of cognitive work, too. When using a LLM, the mental effort of formulating an email or creating a poem is reduced to the mental effort required to formulate a prompt. LLMs may therefore act as a double-edged sword, and overreliance could lead to a degradation of human skills in critical thinking, writing, and analysis, as tasks are increasingly delegated to AI systems. For instance, cognitive training to counteract behavioral symptoms of dementia and increase cognitive performance often involves spatial orientation, memory, attention, language, perception, and visual analysis (Mondini et al., 2016; Hill et al., 2017). Furthermore, overreliance can come in the form of overuse at the expense of social activities (Ma et al., 2024). For instance, conversational applications offering companionship to combat loneliness run the risk of exacerbating social isolation.

Risk mitigation measures that are tailored for specific risks have been described in the previous sections. In this section, we introduce some more general risk mitigation measures that apply across multiple risk scenarios.

Even in the absence of a full understanding of the inner workings of LLMs, insight on LLMs is gained when its behavior can be predicted from a smaller, less capable version, or alternatively, when its capabilities at the end of training can be predicted from its capabilities at early stages of training. OpenAI (2023) used the term “predictable scaling” and showed that model performance could be predicted from significantly smaller models. The expended compute, that is, the amount of training the model received, alone was an accurate predictor of overall loss. Even performance on specific datasets such as HumanEval (Chen et al., 2021) could be predicted with simple power laws, although this did not hold for other metrics such as Inverse Scaling Prize (McKenzie et al., 2023). Ganguli et al. (2022) confirm that overall model performance can be predicted well using either expended compute, dataset size or model size (i.e., number of parameters) as a predictor, performance on specific tasks can emerge abruptly. For instance, they report a sudden emergence of arithmetic, language understanding, and programming skills with increasing model size for GPT-3. Crucially, LLM can learn to solve novel tasks without being explicitly trained to do so (Bubeck et al., 2023). Ganguli et al. (2022) also caution that the open-ended nature of LLMs means that harmful behavior can go undetected simply because it is impossible to probe the model with all types of input that lead to harmful behavior.

LLMs have the potential to help psychiatrists and other healthcare professionals with routine tasks such as writing of clinical reports, saving time and reducing manual data management (Cheng et al., 2023; Javaid et al., 2023). They have been used to provide summaries of patient-doctor conversations (Zhang et al., 2021), clinical notes (Kanwal and Rizzo, 2022) and reports (Vinod et al., 2020), as well as coding adverse events in patient narratives (Chopard et al., 2021). Furthermore, although off-the-shelf LLMs lack the sophistication required to answer queries of medical experts, finetuned models such as PMC-Llama (Wu et al., 2023) and Med-PaLM (Singhal et al., 2023) show increased expertise. In line with this, Lehman et al. (2023) showed that models trained or finetuned on clinical records outperform models that are not finetuned or that rely on in-context learning. In safety-critical domains such as medicine, the accuracy of the summary is of utmost importance. In this regard, Van Veen et al. (2024) performed an experiment with physicians showing that they preferred LLM-based summaries over summaries produced by human experts across a variety of domains (radiology reports, patient questions, progress notes, and doctor-patient dialogue).

Prediction of dementia using artificial intelligence with various biomarkers is well researched. First, one branch of researchers focused on neuroimaging data, using structural Magnetic Resonance Imaging (MRI) for predicting accelerated brain aging (Baecker et al., 2021; Treder et al., 2021), functional MRI (Du et al., 2018), electroencephalography (Jiao et al., 2023), or a fusion of different modalities (Abrol et al., 2019). Second, clinical summaries have been used with LLMs to make differential diagnoses (Koga et al., 2024). Mao et al. (2023) showed that a language model can use clinical notes to successfully predict the transition from mild cognitive impairment to Alzheimer's disease. Third, diagnostic markers can be extracted from patients' speech, either directly from acoustic signals or from the transcribed text. A number of approaches showed a high predictive accuracy using acoustic features such as number of pauses and speech rate (Toth et al., 2018; Al-Hameed et al., 2019; O'Malley et al., 2020). Bang et al. (2024) used a combination of speech, text, and fluency opinions and reported an accuracy up to 87% for discriminating between Alzheimer's patients and healthy controls. In a different approach by Bouazizi et al. (2024), center of focus changes of participants when describing an image were predictive of dementia. Agbavor and Liang (2022) used GPT-3 to extract text embeddings that were then used as features to distinguish Alzheimer's patients from healthy controls. Better results were obtained for text features than for acoustic features using the speech signal directly. This suggests that text, although lacking information such as intonation, pauses, rate, and rhythm, might contain enough information to enable dementia prediction. Lastly, as a complementary application to prediction, LLMs are also able to generate synthetic data that can counteract the scarcity and imbalance of curated medical data and thereby aid in the training of prediction models (Li et al., 2023).

LLMs are able to participate in conversations about daily or private matters, questions and concerns. When tuned to respond adequately (e.g., displaying understanding and empathy) we hypothesize that an app could provide additional companionship and emotional support, especially in situations wherein PwD are socially isolated. Feeling of loneliness has been associated with a higher risk for developing dementia later in life (Holwerda et al., 2014), although the literature is inconclusive on whether this relationship is causal (Victor, 2021). There is evidence that apps in general can help reduce loneliness and isolation in dementia (Rai et al., 2022). The apps reported in Rai et al. (2022) were aimed toward communication and social connections, improving engagement and physical activity through multi-sensory stimulation, remote monitoring and support, and assistive functions. Some studies reported positive results on digital pets and humanoid social robots for combating loneliness and social isolation in dementia (Gustafsson et al., 2015; Demiris et al., 2017; D'Onofrio et al., 2019; Fields et al., 2021; Lima et al., 2022). In a field study with 25 participants from an elderly home, Ryu et al. (2020) found significant decreases in anxiety and depression after daily use of a conversational chatbot for free conversations. Qi and Wu (2023) highlight the potential benefits of ChatGPT in terms of loneliness, emotional support, and assisting with daily tasks including reminders, medications, and appointments. This nicely dovetails with the assessment of healthcare professionals who report merit in virtual assistants and companions (Koebel et al., 2022). In summary, we believe that LLMs hold potential as a companion and serve as an antidote to loneliness and social isolation associated with dementia. As LLMs mature technologically, it is possible to have increasingly meaningful and deep conversations with them. Although it is unlikely and perhaps undesirable that they can fully replace conversations between humans, they can complement and enhance human interaction, especially when carers are not accessible 24/7. Such social and conversational LLMs can come in the shape of apps, as potentially voice enacted chat applications. More immersive social interactions might be possible when the LLMs are digitally embodied as virtual avatars (Morales-de-Jesús et al., 2021) or even physically embodied as robots (Lima et al., 2022).

LLMs can serve as reservoirs of knowledge. Although this is one of their more basic applications, it can be useful for PwD. Unlike conventional search engines that merely retrieve websites, LLMs excel in identifying, compiling and re-synthesizing knowledge and presenting it in an accessible and understandable form. Saeidnia et al. (2023) reported dementia caregivers were overall positive about the quality of answers given by ChatGPT to queries about non-clinical issues relevant to PwDs' lives. However, for questions related to dementia, LLMs may not be sufficiently accurate out of the box. For instance, Hristidis et al. (2023) compared ChatGPT with Google search for questions specifically related to dementia and cognitive decline with subpar quality for both systems. In line with this, ChatGPT's knowledge of dementia has been designated as “accurate but shallow” (Dosso et al., 2023). This can potentially be alleviated by finetuning LLMs on medical data. For instance, PMC-Llama is a model based on Llama that has been finetuned using medical journal papers and textbooks (Wu et al., 2023). Similarly, Google released Med-PaLM, a version of their PaLM specifically geared toward answering medical questions (Singhal et al., 2023). Additionally, one can envision that LLMs could be finetuned to adapt their style to the user via prompt engineering. By default, models such as ChatGPT have a verbose and rather academic writing style. In summary, we believe that LLMs can be useful for the collation and reformulation of generic information as well as information specifically related to dementia. In the latter case, finetuned models such as Med-PaLM will likely be required. Furthermore, care needs to be taken to avoid blurring the line between a conversational service and medical advice, since at least for the time being healthcare professionals should be the ultimate source of medical advice.

As alluded to in the previous paragraphs, LLMs can provide companionship and combat loneliness and social isolation. However, can it be used by therapists and healthcare professionals to aid during therapy? A review of previous-generation language models reported promising potential for use in mental health (Vaidyam et al., 2019). Despite limited data on its clinical efficacy, users dealing with mental health problems have been consulting ChatGPT (Eliot, 2023). Some studies investigated language models in the context of reminiscence therapy which involves engaging patients in recalling and discussing past experiences, often using tangible prompts like photographs or familiar objects (Khan et al., 2022). Reminiscence therapy can enhance emotional wellbeing and cognitive function, as it encourages communication and the recollection of personal histories. Carós et al. (2020) built Elizabot, a language model that mimics a reminiscence therapist. It consists of two components, a model that analyzes and captions the images used in the therapy, and a model for simple conversations. The authors received positive feedback from PwD trialing its use. Similarly, Morales-de-Jesús et al. (2021) implemented an automated reminiscence model. It was integrated within a speech-enacted virtual avatar and people with Alzheimer's disease trialing the system gave it an overall score of 4.18/5, indicating high levels of satisfaction. It is worth stressing that both studies did not use state of the art models such as GPT-4. State of the art models are likely to have higher image captioning and conversation abilities, with potentially positive knock-on effects in the quality of reminiscence therapy. In line with this, Raile (2024) highlighted ChatGPT's usefulness both for complementing psychotherapy and as a first stop for people with mental health problems who have not sought help yet, though concerns remain regarding biases and one-sided information. Furthermore, cognitive behavioral therapy has shown promising results in treating anxiety and depression in dementia (Tay et al., 2019). LLMs can potentially help administer cognitive behavioral therapy via phone apps (Denecke et al., 2022) or in the shape of conversational chatbots (Patel et al., 2019; Omarov et al., 2023). In an analysis of social media posts on an LLM-powered mental health app (not specifically aimed toward PwD), Ma et al. (2024) reported on-demand and non-judgmental support, the development confidence and self-discovery as the App's benefits. In summary, we believe that LLMs can serve as therapy assistants to healthcare professionals. They either affect the therapeutic quality either indirectly by reducing the work burden of a healthcare professional, or directly by engaging in an intervention such as reminiscence therapy.

A useful but easily overlooked feature of LLMs is that they can comprehend complex text and paraphrase it in more palatable or adequate language, e.g., rephrasing a formal text using more casual language. This is a relevant functionality since PwD are more likely than healthy controls to suffer from reading and writing deficits and speech pathologies (Murdoch et al., 1987; Krein et al., 2019). Consequently, LLMs could help in the interpretation and comprehension of letters, or emails, manuals, especially when being verbose or using convoluted language. Similarly, LLMs can assist in the formulation of letters and emails. We are not aware of specific studies on dementia in this regard, but LLMs have been explored for clinical text summarization (Van Veen et al., 2023; Tian et al., 2024) and the summarization of fiction books (Wu et al., 2021). Furthermore, LLMs are increasingly being used as co-pilots in the writing of scientific articles (Altmäe et al., 2023; Lingard, 2023; Park, 2023), including the present one, as well as liberal arts (Oh, 2023) and business writing (AlAfnan et al., 2023). We are not aware of specific studies on dementia for writing, but language models have been explored as email writing assistants for adults with dyslexia (Goodman et al., 2022; Botchu et al., 2023). In summary, LLMs as reading and writing aids for dementia have not been explored sufficiently, hence more research is required to evaluate their utility in this area.

Several types of dementia, including Alzheimer's disease and dementia with Lewy bodies, can affect visual cognition and navigational abilities to varying extents (Plácido et al., 2022). Spatial navigation aids for people with dementia in forms of digital apps and devices have been explored for years (Kowe et al., 2023; Pillette et al., 2023). Navigation aid can be useful both for outdoor navigation, e.g., finding your way from the home to a destination, and indoor navigation, e.g., finding the way around a hospital or other large building (García-Requejo et al., 2023). Tech companies such as Google aim to integrate conversational services into a wide variety of apps (Wang and Li, 2023). This opens the door for language and speech-assisted navigation, where the user converses with the navigation system and can ask for clarification and guidance. Currently, we are not aware of any such systems specifically developed for dementia patients. Further technological development and research on the academic and clinical side are required to assess how LLMs can aid navigation in these populations.