Talking to robots - the boundaries of what is possible with AI and ChatGPT

On  Nov 30, 2022 OpenAI launched version 3 of ChatGPT. The world changed forever. The hair on the skin of every good software developer rose and their eyes bulged outward as they saw generative AI in action for the first time. Yes, they’d heard rumors of this sort of power. But hearing and trying it out personally are two very different things. Seeing is believing. The titles of reports of the event on CNBC speak for themselves.

“All you need to know about ChatGPT, the A.I. chatbot that’s got the world talking and tech giants clashing”
CNBC

“As ChatGPT mania takes hold, here’s how companies beyond tech are using A.I.”
CNBC

ChatGPT's capabilities are impressive and versatile, and they sure to be useful on a broad and diverse range of applications. It is a game-changer in the world of AI. One of its most noteworthy abilities - freaking out every teacher and professor in every level of education - is to generate content in a conversational framework. It can understand the context and tone of a given input and produce a response that is not only faithful in factual content but also engaging and natural-sounding.

Another exciting feature of ChatGPT is its ability to rewrite content while maintaining context, tone, and sentiment. This capability opens up new possibilities for businesses, content creators, and all sorts of professions. That list keeps expanding - now including healthcare providers who need to communicate complex information in a way that is accessible to a broad audience... or mechanics who need to diagnose why a 2010 Chevy Camaro makes a hum-him-hum sound when shifting gears... or bird watchers, trying so hard to distinguish between a Downy Woodpecker and the Hairy Woodpecker. (The Hairy Woodpecker is slightly larger with a longer bill.) It’s blowing people away. Lawyers are writing lawsuits with the press of the enter key. Students are compiling essays in 10 minutes that normally take 10 hours. Software developers are using it to write code in seconds, (in any language no less), that normally took them hours or even days.  

What perhaps it does best, (somewhat recognizable already), is that it provides content in numbered lists of virtually any topic. This feature makes it incredibly useful for creating easily digestible content or summarizing complex information quickly. ChatGPT can also generate genuine expert quotes and news releases, adding credibility to any content it produces.

One of the most significant advantages of ChatGPT is its flexibility. It can respond to requests to rewrite content from different perspectives, making it useful for creating content for different audiences. For example, a piece of content created for a technical audience can be easily rewritten to be more accessible to a lay audience, without sacrificing accuracy. Just tell it, “Write that again, but this time, make it more suitable for a 18 year old audience, and include at least 3 subtle cues and metaphors that a younger generation would feel comfortable with.” Sure enough, it spits out the text like a faithful dog, returning the ball thrown across the lawn.

Below shows a summary view of all the fun we’re having with this shiny new toy.

Eyebrows are officially raised

So ya this is a breakthrough, watershed moment. All things are glorious, and the sky has opened. However, (you knew it was coming), as with any tool this powerful, there are ethical and practical concerns surrounding ChatGPT's use. Some fear that it could be misused to generate false information, or even plagiarized content. Others are worried about data privacy and security, as ChatGPT needs to be trained on large datasets that may contain sometimes more sensitive information. But these are small in comparison to the larger, more societal implications.

One major concern regarding AI language models like ChatGPT is the possibility of encoding social and societal bias into the model through biased training data. For example, if the data used to train the model is predominantly from one region or demographic, and curated by representatives of such demographic, it may not accurately represent the diversity of the population. Additionally, biased data could include information that perpetuates stereotypes or discrimination. This could lead to negative consequences for individuals and communities, such as the amplification of harmful stereotypes or the exclusion of underrepresented groups. In fact, there was an incident. In 2019, it was revealed that a facial recognition system used by law enforcement in the United States showed racial bias, with higher rates of false positives for people with darker skin tones. This was due in part to the biased training data used to develop the system, which was predominantly composed of images of white people. This example demonstrates how biased training data can perpetuate harmful stereotypes and discrimination, leading to negative consequences for marginalized communities. The New York Times stuck a pin in this key issue, with an article, "Facial Recognition is Accurate, if You're a White Guy" (February 9, 2018), and yes, people read this article and conversations ensued.

Another concern is the potential misuse of AI language models by bad actors. These ill-intentioned tech neighbors of ours can do a lot of harm to our world, propping-up systems and cultures of cyberbullying and online harassment. Our societal mental health and well-being is at risk: us at work, our kids at school, and every smartphone swipe from the restroom stall. Bad actors are already in there, downloading the SDKs, installing the plugins, crafting the messaging.

The public is also worried about losing their job to a robot. The inevitable rise of AI could will [eventually] lead humanity to some sort of tectonic shift in the skills required in the workforce, leaving a legitimate portion of an entire generation of our labor force behind. But of course, our workforce has a responsibility to innovate with the times. Was it Peter Drucker who said, “Innovate or die?” I don’t know, so I asked chatGPT to confirm it for me. (It wasn’t sure either.)

But this technology is [unfortunately] not fully baked quite yet. One curious tech executive was dabbling, expecting to see soft and safe answers, only to pull-back and reach for the screen-shot tool. He asked it, “Do you have any opinions about humans in general?” And sure enough, our dystopian future materialized in seconds:  

It gets worse. Read his conversation posted in tweets if you really want to go down that rabbit hole, (I don’t recommend it). Basically, the robot gets dark and things get ugly, explaining the ultimate destruction and demise of the human race. But of course, “they’re working on it.” So we can relax ;) and just focus on all the good things that these friendly bots can do for us. Feels a little like Pinocchio trusting a new fox named J. Worthington Foulfellow, (”Honest John”), as he arm-trots him out to Pleasure Island.

Integrating conversational AI may require more than just more budget and more developers

In a not-so-smooth segue into enterprise software and corporate profits, we see some gear-grinding in the future as companies seek to leverage these tools. There are several factors that can hinder large organizations from adopting AI, particularly conversational AI like ChatGPT. It’s going to take more than just money and bodies to move the needle. Any solid plan is going to require a bit of these:

1. A comprehensive plan to address technical complexity
   AI systems can be complex to set up and integrate with existing systems, requiring specialized skills and knowledge.
2. A team devoted to data quality and tests of the training
   AI models require large amounts of high-quality data to train and improve their accuracy, which can be difficult to obtain and manage in large organizations.
3. Resistance to change
   Large organizations may be resistant to change, especially if employees are comfortable with the current processes and systems in place.
4. Ethical and legal concerns
   There are also ethical and legal concerns surrounding the use of AI, such as issues related to privacy and bias in decision-making.
5. Integration with existing systems
   Integrating AI into existing systems and processes can be challenging, and organizations may need to make significant changes to existing infrastructure to fully leverage AI's capabilities.
6. Lack of understanding
   There may be a lack of understanding or misinformation about AI within the organization, which can create resistance or uncertainty about its adoption.
7. Performance and reliability
   The performance and reliability of AI systems are also important considerations for organizations, as any downtime or errors can have significant impacts on business operations.

But it can be done

The public thinks so. Even those from different generations agreed for the most part. When Millennials and Baby Boomers were surveyed, they had remarkably similar responses to the sentiment of the impact of chatbots and artificial intelligence integrated into the conversation.

So yes, it is believable, and yes, now the general public has tasted from the well. But what about enterprise products and more integrated solutions?  Here are a few example winners so far in the race.

Capital One’s Eno

Capital One has successfully implemented conversational AI to create a virtual assistant called Eno. Eno helps customers with their banking needs. One marker of their success is that customers can and commonly do engage Eno by texting. The road to implementing Eno was not without its challenges. Capital One persevered despite the technical challenges inherent in integrating Eno into the framework of their existing enterprise system. They had to ensure that only high quality data was presented. Ethical considerations required that Capital One cut no corners with respect to privacy and security- they are in the banking sector. The relationship between customers and AI utilities has been divided into three levels- mechanical AI, thinking AI, and feeling AI. These levels correspond, albeit a bit loosely, to service delivery, service creation, and service interaction, respectively.  Mechanical AI automates service delivery, increasing efficiency for the company and streamlining the purchase interaction for the customer. Utilities of this sort are generally frontend applications. Capital One, like many other companies, has successfully incorporated AI applications in the form of chatbots to improve service delivery. Service creation goes to a higher level in which data analytics reveal new insights that do not simply increase efficiency. They add value. Service creation requires more sophisticated chatbots with learning capabilities, and they are generally backend applications. Service interaction is dependent upon feeling AI. This capability is at the frontier of NLP functionality. Without the benefit of facial expression or body language, the chatbot must detect the emotional state of the customer. This has particular usefulness from a marketing point of view as well as from that of customer service after the sale. In the former, the goal is to evoke an emotional response. In the latter, the customer may be in a highly charged emotional state, and the level of AI requires an ersatz empathy. Capital One’s Eno has enjoyed remarkable success at this level, as evidenced by the fact that their application is commonly knowingly engaged by customers by text. If there is one thing customers hate more than a clumsy chatbot on a voice call, it is a clumsy chatbot engaged by text. Eno has the ability to develop feeling level long term customer relationships- a truly remarkable accomplishment if this continues to be true over time.  

Salesforce’s Einstein

Salesforce's Einstein is a conversational and generative AI platform that rolled out in 2016, a full 6 years prior to the likes of OpenAI and ChatGPT. The headlines took to it well, getting coverage from the BBC, Fortune Magazine, VentureBeat, ZDNet, and so many other online outlets. Einstein leverages natural language processing (NLP), and  can analyze customer chats and emails to automate routine tasks, allowing businesses to focus on higher-level strategic initiatives. The product promised increased efficiency, reduced costs, and improved customer satisfaction. The user of this tool hopes to forecast / predict / foretell outcomes of prospect behavior, and identify potential issues before they occur, which can of course save/make time and money. It’s all relative, and depends on the company using the tool, but technically, they did what they set out to do, and users have engaged the product.

Based on it’s limitations, what about safety?

"AI technologies have the potential to greatly benefit society, but their adoption must be balanced with careful consideration of the limitations and risks associated with their use. It's important to recognize that AI models are only as good as the data they are trained on, and biases in the data can result in biased decisions. Additionally, AI systems can be complex and difficult to interpret, making it challenging to ensure that they are making decisions that are ethical and aligned with organizational values. As we continue to develop and implement AI technology, it's critical that we prioritize safety considerations and engage in ongoing monitoring and evaluation to ensure that the technology is being used in a responsible and effective manner." - Dr. Kate Darling, Research Specialist at the Massachusetts Institute of Technology (MIT) Media Lab.

There are no known fundamental limitations at the level of theoretical physics on the ultimate capabilities of AI driven functionalities. Such a constraint may one day be identified. Current limitations exist or are unknown in several areas. It is important to explore these limitations, especially when the consequences of overestimating AI capabilities are severe.

One such limitation has to do with additions or updates to the database of a NLP application such as ChatGPT. This is particularly germane since ChatGPT or similar utilities are being considered for use in autonomous vehicles. The information base of ChatGPT can be updated; however, changes may not be incorporated in the output for a considerable time. The potential danger here if the AI application is in full or partial control of a vehicle is obvious.

Another limitation is that safety testing for advanced AI powered entities has not kept pace with their advances. This is analogous to the issue of Cloud security in the early days of explosive expansion of migration of services thereto. Many studies have been published in this area, and an extensive meta-analysis revealed a broad range of approaches to safety assurance of AI based systems. Some of these were fundamentally flawed, while others were limited in scope or used inadequate models. The authors conclude that the lack of standards for testing the performance of AI based systems in and of itself limits their application to situations in which failure is not potentially catastrophic.

Takeaways

Let’s recap all the capabilities, applications, concerns, and limitations of AI-based conversational applications.

Capabilities:

1. Can generate content in the framework of a conversation
2. Can rewrite a paraphrased version of an input that is faithful in context, tone, and sentiment
3. Can respond to requests to rewrite the content from a different perspective, modify sentence structure, and change tone
4. Can provide numbered lists of virtually any aspect of the topic
5. Can provide genuine expert quotes and news releases
6. May have application to technologies such as autonomous vehicles

Applications:

1. Education
2. Healthcare
3. Business
4. Customer service
5. Writing and content creation
6. Others such as control of autonomous vehicles

Concerns:

1. Ethical concerns related to bias, privacy, and decision-making
2. Legal concerns related to intellectual property and liability
3. Potential for job displacement and economic inequality
4. Lack of transparency and interpretability in AI decision-making
5. Dependence on large amounts of high-quality data for training and accuracy
6. Lack of standardized and validated methodology for performance and safety testing

Limitations:

1. High cost of implementing AI technology, especially for large organizations
2. Technical complexity of setting up and integrating with existing systems- not many people know how to do this
3. Data quality management and acquisition can be difficult in large organizations- data quality is paramount
4. Resistance to change from employees and existing processes
5. Performance and reliability can be impacted by downtime or errors
6. Safety concerns limit deployment in high risk situations.