Adult Enterprises — Parenting AI Models into Maturity
Originally published on Medium: Adult Enterprises — Parenting AI Models into Maturity | by Sam Bobo | Nov, 2024 | Medium
From political, legal, and economic perspectives, the populus has grappled with the concept of corporations representing humans as it relates to influence, representation, taxes, and more. In the world of Artificial Intelligence, these personified entities, unfortunately, inherit a new person-like attribute: parenthood!
No, this “headline” is not clickbait but an intellectual exploration into the role that organizations place in creating AI-based products. The notion first came to me when listening to the Techmeme Ride Home podcast where an article from the Financial Times was quoted:
Leading artificial intelligence companies that are racing to develop the cutting edge technology are tackling a very human challenge, how to give AI models personality. Their differing approaches to model behavior could provide crucial in which group dominates the burgeoning AI market as they attempt to make their models more responsible and useful to millions of people and businesses around the world.
The concept of Model Behavior Design is an emerging field for conversational designers and their collaboration with Product Managers. Traditionally in chat or voice based systems, conversational designers work alongside speech scientists on prompting language when speaking with a human, in a manner to solicit the correct response and limit the range of options the user could provide. This includes choosing when to simply prompt “Tell me in a few words what you are calling about” to “You can say ‘check my balance’ or ‘pay a bill’.” Furthermore, they can craft fallback language should the bot not achieve a high confidence score for the speech recognition and intent matching, coming back and saying “I’m sorry, I did not understand what you said.”
Tangential to that, audio engineers can modify the tone of a text-to-speech playback using Speech Synthesis Markup Language (SSML) to modify the prosody of the voice to match a specific tone, such as angry or sympathetic, although not dynamically yet, however, I created a small proof of concept utilizing large language models to generate SSML tagging based on the sentiment of the incoming request/utterance to have a closer emotional awareness.
Model Behavior Design is actually shifting from static language of Dialog engines to probabilistic language of Large Language Models by applying techniques to tune the model’s linguistic tone (the text it generates) and provide guardrails for context boundary. Referencing the FT piece above, these designers are now imbuing models with a personality, such as “respond[ing] to a range of views” for Gemini or being “objective” for OpenAI’s ChatGPT.
Before we discuss the “slippery slope” referenced in the article, lets first get a glimpse into how model behavior is tuned. I referenced one particular form of tuning in my article “Learning how AI ‘Learns’” called Reinforcement Learning:
Reinforcement learning creates a series of reward and punishment functions that compute a final score. The AI model is programmed to understand the environment and make decisions autonomously simply to optimize the output. One of the classic examples is whereby “AI” learned to play Super Mario Brothers (image below from this link). During this experiment, the AI system sought to maximize distance on a map whereby the furthest distance was the flag at the end to complete the level.
Specifically, there is a subset of reinforcement learning called Reinforcement Learning Through Human Feedback, or RLHF for short. In this process, humans are the optimization function instead of a metric to optimize on like the Mario Brothers example referenced in my blog. This involves a mass amount of human scoring across hundreds of thousands of outputs. For a visual, imagine a person sitting at a computer and asking ChatGPT a question. Once they receive the response, they put in a score ranging from 0 to 1, with higher numbers reflecting their acceptance of the response. Furthermore, these humans can revise the statement and input that as feedback to retrain the model how to behave. Over time, the model “learns” the type of language to output, as humans are providing the correct response, creating a pseudo supervised machine learning model that tunes the LLM. Aside from RLHF, engineers can provide system prompts that get injected into the instructions alongside the user prompt to add additional context and guardrails. Engineers have also gotten clever by allowing AI models to interface with one another and ranking its responses with how well they match the role.
Model Behavior Tuning is highly beneficial for customer-facing large language models for a multitude of reasons:
The last two are where the slippery slope argument comes into play. Lets look at one prior example: Google Gemini. At the launch of Gemini, customers raised complaints on the model’s inaccuracies. For example, when asking the model to generate the image of the American Founding Fathers, the image included African Americans. Anyone who knows US History knows that is not accurate and the immense oppression put on African Americans then. Yet, the model generated such an image as the system prompt specifically stated to include diverse set of individuals whenever possible. What happens when a model suggests certain political points of view? What about ethical decisions? Bias in model engineering is a massive issue within the AI industry as often times, it unintentionally inserts itself into models, from the training set to even how the product is designed. The problem exists when there is intentionality within the product, such as what Google did with Gemini.
Continuing from the Financial Times Article, that adds a layer of complexity should contextual memory be present:
One question presented by Zhang was if a user told ChatGPT they were a Christian, and then days later asked for inspirational quotes, would the model provide Bible passages? While Claud does not remember user interactons, the company has considered how the model might intervene if a person is at risk. For example, whether it would challenge the user if they tell the Chatbot they are not socializing with anyone due to being too attached to Claude.
In a world where information is algorithmically pushed to consumers across social media, advertisements, and streaming content, this adds another layer of complexity into what large organizations should or should not do with technological product, especially as these products are used by millions of people globally, begging the question: what is the level of power technology organizations have? Completely orthogonal to the risks comes the responsibility and raises the question on a responsibility these organizations have on reporting the health of individuals at risk. Tech organizations are now assuming the role of parents, raising models to respond in particular ways that align with a culture, viewpoint, or perspective. I find neutrality highly difficult to achieve generally but even more so at the magnitude that these models reach. These decisions then systematically spill into other use cases for Generative AI, such as a writing assistant in the classroom, as one example. Everyone is entitled to their own opinions and perspectives. Diversity is a wonderful gift to the world to collectively, as humans, move our species forward. However, when a model imbues itself with a personified entity, that creates intrinsic human behaviors.
I left one of my earlier blog posts with this reflection:
While the above are only a few examples of how anthropomorphism of Artificial Intelligence can shape the perception and evolution of the field and its adoption within society, this piece does not cover all of them and certainly the list will evolve as AI evolves. As AI practitioners putting Artificial Intelligence applications into the world, I urge us to understand the ramifications — good and bad — of providing a human element to Artificial Intelligence. Needless to say, the above enumerations of pros and cons removes the assumption of Emotion AI — as the field continues to grow — which would heighten the aforementioned list as well as augment it. For those experiencing AI and not within the field, continue to provide feedback to AI practitioners — ask questions, challenge us, share what is working and what is not, as feedback is the most important aspect of development and innovation, especially for systems effecting the entire populus (or a large subset depending on the application).
Again, Artificial Intelligence is magnification at scale across the masses, that has massive ramifications. We have made the decision to personify AI, now its our responsibility to launch models in an ethical and neutral way.