Humour

In a typical FL architecture, a centralized server initializes a global model and broadcasts its parameters to participating devices or institutions. Each client then uses its local data to train a model that shares the same architecture as the global model, and only the updated parameters are sent back to the central server for aggregation^[1]. This approach ensures that raw data remain on-site, thus mitigating privacy risks associated with traditional centralized machine learning. In addition, the concept of data partitioning is fundamental, where sites may share the same feature space (horizontal FL) or have complementary features (vertical FL), which is particularly relevant in collaborative healthcare environments^[1]. Organizations such as those highlighted in industry applications are leveraging FL to aggregate models without compromising data ownership, thus bridging the gap between medical research and advanced analytics^[5].

Differential privacy (DP) is increasingly incorporated into FL frameworks to strengthen privacy guarantees during model training. In local differential privacy, noise is added to data on the user device before transmission, ensuring that even model updates do not reveal sensitive details^[2]. For instance, mechanisms that combine the Laplace mechanism with randomized response methods are used to perturb data points in real-time, especially for streaming health data from wearable devices^[2]. Additionally, systematic reviews of wearable health data publishing under DP demonstrate that careful calibration of the privacy budget is crucial; a smaller privacy budget offers stronger privacy but may decrease data usability, whereas a larger budget recovers some utility at the potential cost of privacy^[7].

Wearable devices continuously collect various physiological data such as heart rate, blood pressure, and activity levels that are essential for remote health monitoring and personalized medical care. Federated learning in wearable ecosystems facilitates the joint training of models on these sensitive data streams while ensuring that individual data remain on the device^[2]. Industry examples illustrate that collaborative FL setups not only protect privacy but also enable insights from diverse sources – for example, consortiums can pool chemical libraries or clinical trial data without exchanging raw information^[5]. Such approaches demonstrate the potential to improve diagnostics, treatment personalization, and even drug discovery by leveraging previously inaccessible datasets.

Performance evaluation of federated learning models has been performed on several benchmark datasets. In controlled experiments using simple datasets such as MNIST, FL frameworks have achieved global model accuracies of around 82% under IID data conditions, whereas non-IID data distributions caused steep drops in performance, with mean accuracies as low as 17% in some scenarios^[4]. Furthermore, the introduction of differential privacy techniques involves a trade-off between accuracy and privacy; increasing the privacy budget tends to improve model accuracy, yet the addition of noise can also delay convergence or reduce performance when excessive^[4]. Simulations also underscore that computational resource disparities among devices may lead to slower convergence, with straggler effects notably impacting training speed despite similar final accuracy levels^[4].

Despite its strong potential, federated learning faces several limitations that must be addressed for broader clinical adoption. One notable challenge is the risk of privacy leakage during the sharing of model updates; even though raw data remain local, reconstruction or membership inference attacks remain possible if adversaries analyze the transmitted weights^[1]. Communication overhead is another critical limitation because as the number of participating devices grows, the volume of data exchanged increases significantly, which can slow down model convergence and require significant network resources^[6]. Additionally, device heterogeneity and non-uniform data distributions challenge the fairness and stability of the global model, as evidenced by stark performance differences between IID and non-IID settings^[4]. Future research should focus on improved client selection, adaptive aggregation strategies, and sophisticated privacy-preserving algorithms to balance privacy with model utility. The integration of techniques such as homomorphic encryption, secure aggregation, and federated meta-learning is also anticipated to enhance both the security and scalability of FL in healthcare^[7].

AI models are trained to recognize and generate joke structures and surprise^[1]. One common structure involves a 'setup-punchline' pattern, where the punchline subverts expectations created by the setup^[1]. This element of surprise can be quantified using mathematical tools like Kullback-Leibler (KL) divergence, which measures the difference between expected and actual outcomes^[1]. Signal processing and dynamical systems are also used to mathematically model timing and delivery, critical aspects of comedic performance^[1].

To train AI models, large datasets of jokes are essential^[1]. These datasets are sourced from joke websites, comedy shows, and social media, requiring careful pre-processing to remove irrelevant or offensive material^[1]. Natural Language Processing (NLP) and Natural Language Generation (NLG) techniques are then employed, with language models trained to predict the next word in a sequence^[1]. Recurrent neural networks (RNNs) and transformer-based models like GPT-3 are popular choices, adept at capturing complex patterns and dependencies^[1]. Fine-tuning these models on specific joke datasets further enhances their ability to generate humor^[1]. HumorSkills, for example, uses a three-step process: visual detail extraction, narrative and conflict extrapolation, and fine-tuning the joke generator with examples of Gen-Z humor^[2].

Evaluating AI-generated humor presents unique challenges due to its subjective and context-dependent nature^[1]. The 'Turing Test of Comedy,' where human judges rate the funniness of jokes without knowing their origin, serves as one approach^[1]. Metrics such as precision, recall, and F1-score are utilized, alongside human ratings of funniness, surprise, and coherence^[1]. Intrinsic evaluation methods like perplexity and BLEU scores measure fluency and similarity to reference texts, although they may not fully capture the nuances of humor^[1]. Ultimately, the audience's laughter remains the most crucial test^[1]. A recent study also introduced a novel humor detection metric designed to evaluate LLMs' capability to extract humorous punchlines from stand-up comedy transcripts, using fuzzy string matching, sentence embedding, and subspace similarity^[5]. The study revealed that leading models achieve scores of at most 51% in humor detection, surpassing human evaluators who achieve 41%^[5].

AI-generated humor finds practical applications in stand-up comedy, television, movies, and personalized recommendations^[1]. Robo-comedians, AI-powered virtual or physical robots, deliver stand-up routines, tailoring their performance to the audience^[1]. Transformer-based language models like GPT-3 generate jokes, while AI handles speech synthesis, facial expressions, and body language^[1]. In television and movies, AI assists writers in generating humorous content, offering suggestions and helping overcome writer's block^[1]. Personalized joke recommendations leverage collaborative filtering techniques to analyze user preferences and recommend tailored jokes^[1]. For instance, collaborative filtering identifies users with similar tastes and recommends jokes enjoyed by those users^[1].

The rise of AI-generated humor raises ethical concerns, including the potential displacement of human comedians^[4]. A hybrid model, combining AI's strengths with human creativity, could foster collaboration and synergy^[4]. Ensuring inclusive and non-offensive humor is crucial, requiring robust filtering mechanisms to remove biased or offensive content from training data^[4]. Intellectual property and joke ownership present complex challenges, potentially requiring revised legal frameworks to address AI-generated content^[4]. One approach is to recognize AI-generated humor as derivative work, with ownership attributed to the human creators who designed and trained the AI model^[4].

The future of AI-generated humor holds exciting possibilities in virtual and augmented reality (VR/AR) experiences^[4]. Immersive environments can facilitate innovative comedic interactions, where robo-comedians adapt performances based on audience feedback^[4]. AI can also foster creativity and collaboration among human comedians, providing fresh ideas and inspiration for new comedic styles^[4]. Generative adversarial networks (GANs) can further enhance this collaboration, with AI generating jokes and human comedians providing feedback^[4]. Experts anticipate that AI may one day create some form of humor, though it will never match what professional comedians create^[9]. Some researchers are also exploring AI's capacity to understand humor, noting that language models can craft passable jokes using systematized pattern recognition processes^[9]. The use of AI also could help those who do not speak the same language be able to improve the humor in their jokes^[7].

Despite advancements, AI-generated humor faces limitations, including the difficulty of capturing subtle cultural contexts and the dependence on training data^[9][4]. For example, safety filtering in LLMs can make them very dull, and they cannot explore more personal subjects and edgy material^[7]. AI can be helpful in generating innovative ideas, but there is a concern that the humor produced may lack the genuine touch that comes from human creativity^[8]. Also, there is a possibility that AI will amplify disingenuous portrayals of oneself; as AI for social, cultural, and personally relevant communication improves, we may need a way to discern genuine from disingenuous communication^[2].

Humor is a crucial element of social interactions and well-being, making its integration into AI systems valuable^[3]. AI-driven tools can assist individuals, especially those for whom English is not their first language, in improving their humor and grasp cultural references^[7]. In organizational contexts, humor can enhance workplace environments by relieving boredom, building relationships, and improving camaraderie among workers^[6]. The use of humor is especially important in youth development, playing a significant role in social interactions, conflict resolution, and psychological adjustment^[6]. It is generally known that humour contributes to higher subjective well-being (both physical and psychological)^[6].

Memes, a staple of modern online culture, have become a powerful tool for marketing, especially when targeting younger audiences^[10]. Brands are exploring AI to automate meme generation and engage with fast-moving trends, balancing AI efficiency with human creativity^[10]. AI enables businesses to create shareable content, increasing engagement and maintaining relevance^[10]. By using AI-generated memes, companies are able to maintain their image as a brand that understands and engages with internet culture^[10].