GenAI functionality and policies

Artificial intelligence (AI) exists when a computer system performs tasks similar to, superior to, or beyond human capabilities or as the Oxford English Dictionary (2010) states, AI encompasses “computer systems able to perform tasks that normally require human intelligence, such as visual perception, speech recognition, decision-making, and translation between languages.” Other tasks encompass problem-solving, pattern recognition, planning, and solving mathematics and logic problems. Examples include discerning if a cat is in a photo and generating computer art such as a photo of a cat, to detecting irregular cells in a biopsy or identifying spam emails. Prediction is the most crucial function for AI, with GenAI being inherently predictive. Predictions span a wide range, from anticipating human perceptions of nature, cityscapes, or architectural designs, and modeling disease spread, to determining the subsequent word or sequence of words in a sentence. Various machine learning approaches exist for these tasks.

Machine learning (ML) is a subset of AI, where a computer system learns a task from example data and can improve its efficiency at that task without being explicitly programmed using traditional algorithms. Prior to 2012, many machine learning techniques performed poorly at human tasks like language and vision, even with Big Data to support task learning.

Big Data refers to colossal datasets that traditional machine learning or statistical methods struggle to handle. These datasets contain millions or even billions of data points and examples are datasets like those held by Google, Facebook, Twitter, or Visa. In the realm of Big Data, relationships between data points are mostly non-linear, and Deep Learning (DL) evolved as the primary technique to decipher them. DL empowers companies like Google or Facebook to offer personalized advertising experiences for each user, a feat beyond the reach of even the most experienced salesperson.

Deep Learning (DL) is a specialized subset of machine learning that was first introduced in 1989 through the work of LeCun et al. (1989), shortly after efficient methods for training multiple layers or so called ‘deep’ artificial neural networks (ANN) that were developed by Hinton and Williams in 1986. The deeper the network, the more abstract the representations become. For instance, humans and deep neural networks alike can identify abstract representations of a bird. However, these abstract representations within the deepest layers might not resemble a bird to us, much like brain scans of someone contemplating a bird would not appear bird-like. Deep learning effectively processes and examines large datasets, where non-linearity is not a hindrance but a characteristic feature. Deep learning can facilitate numerous tasks, such as image and speech recognition, as well as natural language processing. Deep learning serves as the driving force behind GenAI.

The key breakthrough in deep learning that drives today’s GenAI involves the Transformer architecture that was invented by the Google Brain team in 2017 (Vaswani et al., 2017). Transformers can easily process sequential data, such as natural language inputs, enabling a model to identify and evaluate relationships among words and even relate these to overarching themes. Likewise, in the visual arena, when working with image inputs, transformers learn to recognize colours and shapes at a basic level, while grasping objects and their spatial relationships at deeper levels. Transformer layers in a deep learning model bring a self-attention mechanism that can capture both short- and long-range relations in natural language inputs, leading to an improved ability to process and predict natural language outputs. Transformers enable GenAI models to perform at a higher level for natural language comprehension when compared to previous non-transformer (LSTM or RNN) deep learning models. ChatGPT, for example, is based on the more foundation model called GPT-3.5 (or GPT-4) which stands for generative pre-trained transformer (GPT). These more general models are also called foundation models as they provide a malleable intelligence that can be used for a multitude of tasks beyond basic conversations.  Google’s BARD is based on the foundation model called LaMDA.

Large Language models (LLMs) like GPT by OpenAI or LLmDA by Google Inc. have the most extensive capabilities of any LLM to date. These capabilities straddle numerous knowledge domains and some already exhibit human-like or better performance over a broad class of cognitive tasks. LLMs most prominent capabilities include the capacity to provide structured knowledge on almost any topic, adapt to new situations with minimal direction or training examples, perform multi-step reasoning and inference, classify emotions expressed through language, translate and respond in multiple languages and include idiomatic expressions, follow the context of a conversation, undertake complex math and logic feats and play various roles such as a ChatBot like ChatGPT, or a teacher, novelist, historical figure with good consistency. LLMs are like a polymath in many respects but with access to a large part of the corpus of total human knowledge.

Training is the process through which a deep learning model acquires the ability to process natural language inputs. Current LLMs, for instance, undergo pre-training, where the model learns from massive datasets sourced from the Internet and other places. Following that, the model is fine-tuned using techniques like reinforcement learning, which mirrors the way children learn. In this phase, human feedback is provided for natural language inputs and outputs, similar to how we guide and correct a child’s language development. This combination of techniques helps GenAI become more adept at understanding and generating human-like responses. This training strategy should also bring LLMs into alignment with human values [which values?].

Alignment refers to the degree of similarity between the responses generated by an LLM for various conversations or directives and those that a rational, average human would provide. It measures how well the AI’s responses align with human expectations and reasoning. Alignment has been a concern before today’s LLMs existed and remains a current challenge for most LLMs that are public (Hou and Green 2023; Gabriel 2020). Alignment is a very active area of research.  However, companies controlling LLMs act as gatekeepers, influencing the content these models can express. This is based on the ethical and moral guidelines of the region or population where the GenAI is located, and do not necessarily reflect the same of ideas of the common good, e.g., the alignment would be different for some of the ethics surrounding a public vs. private company (profit maximization vs. social justice and equity). Thus, until such time as a set of alignments is, say internationally agreed on, each GenAI will be largely influenced by capitalist goals and government regulation. Errors and media attention will play a major role in that arena.

Prompts guide GenAI’s responses, acting as natural language directives. For example, asking “Why is Paris the capital of France?” prompts the AI for information.  These types of prompts are the most academically contentious because the GenAI responds with coherent answers without the user needing to think about what or how the material was produced, or indeed, whether it is correct.  Such prompts were the first to foster the current concerns surrounding GenAI and academic integrity because of the proven capabilities of LLMS to write essays and similar prose or create computer code (another type of language).

Prompts are a form of natural language programming. Programming languages exist on a spectrum, from low-level like Assembly to high-level like C++ and Java, up to very high-level, human-readable languages like Python or R that are, in turn, programmed in lower-level languages. GenAI can produce computer code using high-level written language thereby allowing the developer to concentrate more on validation than coding itself. With GenAI, a scientist can describe code requirements in natural language prompts, and the AI generates the code. Although it’s not perfect, especially for complex tasks, GenAI simplifies mundane, repetitive work and reduces the time to completion. Natural language programming research aims to simplify coding. The more precise and detailed the prompt, the more accurately a GenAI can generate the desired lower-level computer code.