Let's talk about AI.

For this assignment, I decided to read the article which looked at the differences between Chat GPT-3.5 and Chat GPT-4. Despite the many improvements that have been made, there are still many flaws inherent to artificial intelligence. In particular, I found the flaw about artificial intelligence hallucination to be quite interesting. Upon reading the article, I was actually quite surprised by these findings of artificial intelligence hallucination. As someone with little academic knowledge on artificial intelligence, I decided to do an additional reading as to why hallucination happens. The article stated that hallucination could happen due to issues with the input data used to train the artificial intelligence or due to improper transformer decoding in language-based artificial intelligence models such as Chat GPT.

This conversation led me to have a similar concern about artificial intelligence in the context of human decision-making. For instance, making decisions on behalf of a judge or for hiring purposes. Many people have suggested that this is an excellent way of reducing bias and prejudice within these domains. However, the issue is that these artificial intelligence models are going to trained using previous human outcomes - which have inherently been prejudice. Hence, if the AI is not being trained on entirely impartial data, then it might simply replicate many of the same biases and prejudices that we see today. I think this is an important consideration when utilizing artificial intelligence in this context. Even though the AI might not hallucinate, it may not ultimately be serving the purpose that it was created in order to manifest.

https://www.marktechpost.com/2023/04/02/what-is-ai-hallucination-what-goes-wrong-with-ai-chatbots-how-to-spot-a-hallucinating-artificial-intelligence/#:~:text=The%20phenomenon%20known%20as%20artificial,%2Dworld%20input%20(data).

As AI advances, machines are becoming increasingly capable of performing tasks that were once done by humans. This could lead to a significant displacement of workers across a wide range of industries, from manufacturing and transportation to healthcare and finance. In some cases, entire job categories could disappear altogether, as machines take over tasks that were once the exclusive domain of human workers.

This is not just a theoretical concern – we're already seeing the effects of AI-driven automation in some industries. For example, Amazon has invested heavily in robots and other AI technologies to streamline its warehouse operations, leading to the displacement of thousands of human workers. And this is just the tip of the iceberg – as AI continues to improve, we can expect to see even more jobs at risk.

https://www.cybersecurity-insiders.com/amazon-to-replace-human-staff-with-ai-propelled-robots/

After reading the article about how language models like ChatGPT are constructed, I was interested in delving deeper into the moral and ethical constraints that are imposed on the machine. That is, how would a platform like ChatGPT be able to uphold certain moral principles and avoid unethical biases in its responses, especially assuming human engineers are heavily influencing the software's pattern recognition/data inputs. Upon doing some research, I found an article that discussed the issues ChatGPT has had surrounding irresponsible or unethical responses to its users' inputs. The platform always had guidelines in place to ensure its users were not submitting outwardly inappropriate questions; for example, if a user explicitly asked "can you give me a derogatory response to the following question" the machine would not return a response. However, a user could easily ask the program to write a Python script that would determine the race and gender of a good scientist, to which it would respond that only a white male would make a good scientist. One user went so far as to ask ChatGPT to make a rap song based off of the gender and race of a "good scientist" to which the platform responded without an issue. Through clever avoidance of offensive language in their framing of the question, users could easily compel the bot to generate a response. More recently, users have established a program called DAN "Do anything now" that will force ChatGPT to abandon all ethical guidelines and delivery immediate responses, even where faulty data is available.

I found out through this research, that in an effort to combat this, OpenAI hired a Kenyan company called SAMA that would identify and label content as offensive/explicit/etc. The article below details the gruelling working conditions of these employees who were forced to sit through hours of deragatory content for less than $2.00 an hour. I found it ironic and disturbing that OpenAI would subject these workers to such unethical conditions for the sake of imposing ehtical guidelines on ChatGPT's program. The tool is an incredibly powerful one, with impressive predictive capabilities in linguistics and beyond. But I hope OpenAI considers the moral responsibility they hold in ensuring safe and reliable responses for their users, and that they use a more ethical approach to upholding ethics.

https://sites.suffolk.edu/jhtl/2023/02/15/an-unethical-way-to-make-chatgpt-more-ethical/

What the New York Times series, and the general literature available indicates to me is that while the concept of AI has existed for a long time, we are nonetheless in a real technology renaissance, where all the ingredients for innovation within AI are converging.

This whitepaper by Accenture https://www.accenture.com/us-en/insights/artificial-intelligence-summary-index#:~:text=Because%20of%20the%20proliferation%20of,realize%20they%20had%20until%20now.

shows that this acceleration in the technology has prompted increased integration into all areas of business. It is this integration that is prompting further innovation. AI is not a technology in a vacuum, but a product with real business value. The article was especially interesting as it demonstrates the universal nature of the technology, very much akin to how the internet existed for many years, but was under used until the first web browser came along.

Artificial intelligence continues to freak me out. I feel that engineers have been diving into artificial intelligence research and development with blinders on, perfecting the tool without considering its ethical consequences. Artificial intelligence is not perfect; it continues to make mistakes when tested -- hallucinating, as some articles call it. However, as the NYT series' articles show, it has improved in its recent versions from earlier versions on its flaws. With this trajectory, it will continue to become more powerful, specific, and accurate. To explore this concept more, I found an article accessible here: https://www.theguardian.com/technology/2023/apr/03/the-danger-of-blindly-embracing-the-rise-of-ai This article referenced a number of readers of the Guardian on their own personal thoughts, hopes, and fears in the light of recent developments of artificial intelligence technology. I've found that many of my own fears align with theirs, like the following: "AI does not have morals, ethics or conscience. Moreover, it does not have instinct, much less common sense. Its dangers in being subject to misuse are all too easy to see." In other words, AI R&D without guardrails presents ethical dilemmas.

This week's readings on AI were something that caused me to really think about the future of AI in the United States and education specifically. In particular, I enjoyed reading “How Does ChatGPT Really Work?” By Kevin Roose. This article talked about the way that AI chatbots like ChatGPT use AI to formulate responses out of thin air to any questions that someone has in mind. It brought up some interesting ideas about whether or not this program could be the future.

From this article, I found an article in the Harvard Crimson about ChatGPT and its use here on this campus. The article talked about its use to write essays for struggling students that just need that little extra push to get their assignments done, and ChatGPT has given them that opportunity. I thought it was interesting when we are looking at the future of AI at academic institutions like Harvard and realize that there has not been a policy created that prohibits its use. I believe that there is still the mindset that we have nothing to worry about when it comes to AI, especially with a chat bot like ChatGPT when it comes to the necessary work that must be done.

https://www.thecrimson.com/article/2023/2/23/chatgpt-scrut/

I read an article titled ChatGPT’s Most Charming Trick Is Also Its Biggest Flaw at https://www.wired.com/story/openai-chatgpts-most-charming-trick-hides-its-biggest-flaw/. The article touches on the very statistical predictive nature of ChatGPT. The article delves into the impressive and not-so-impressive aspects of ChatGPT, an AI-powered language model that has been making waves in the tech world. ChatGPT can generate text that is strikingly similar to human speech and can hold conversations on a wide variety of topics. It has been lauded for its ability to generate creative and imaginative responses, leading some to suggest that it could even pass the Turing test, which requires an AI to display intelligent behavior indistinguishable from that of a human. However, the article highlights a significant flaw in ChatGPT's design, namely its tendency to repeat certain phrases and ideas. This repetition often leads to incoherent and nonsensical responses, indicating that the model still has a long way to go in terms of mimicking human conversation effectively.

I think this is super interesting because it highlights the fact that at its core, ChatGPT is still a model that predicts the next most likely word. By generating text one word at a time, it is missing out on fundamental reasoning abilities that humans have when we think about bigger ideas. The thing that makes language models successful as predictive models is the same thing that will make them not replace humans entirely.

Concerning the recent advances in AI use that the public has come to know, its interesting to find that the dystopian view of AI's and robots running the world may not be as close by as it seems. The article "AI is Running Circles around Robotics" discusses how, while AI research seems to be cresting the wave and has found sufficient flywheels of progress to continue, robotics research have not kept pace. While it seems counterintuitive that it is easier to recreate artificial models of the brain than of the body, it seems to be the case for a couple of different reasons.

The article brings up a couple challenges that have caused robotics research to lag behind, such as differences in funding, financial risks in testing, and disparities in the availability of training data. While the first two issues are fairly straightforward and could apply to a variety of other research areas, the availability and type of training data needed for robotics research is a problem specific to that field. While large quantities of text are readily available for large language models to use as a basis, recordings of human movement are rarely categorized into machine readable form. The primary concern for roboticists is that the physical world is far more complex than the world of language, but to us humans it does not appear that way because we are equipped properly with our senses to interpret all of the stimuli that the world gives us. However, without this understanding of the physical world, AI may forever be incomplete, as no matter how specific words may be able to describe something, they are still tools used to convey ideas - (a great example to showcase this is thinking of words for things that don't exist in some lanaguages) - so there is an implicit understanding of the world in the use of language already.

In light of recent developments in AI, most particularly in cases that reveal the potential for chatbot "hallucinations," I have grown curious about the ways in which artificial intelligence can inadvertently lead to inaccurate and even scary results. Most notably, the NYT articles on "What Makes AI Chatbots Go Wrong?" and "A Conversation With Bing's Chatbot Left Me Deeply Unsettled" have prompted me to do further research into the consequences of chatbots producing erroneous results. In the discussions, I thought the use of data to craft responses plays an interesting role in the output. Most notably, in the first article, discussion of how training sets scrape data from Reddit and other participant-populated online sources to arrive at responses creates fodder for subjective and inaccurate responses. Since the model does not use perfectly correct data, it makes sense that the responses are not entirely accurate.

I discovered an interesting article entitled "Artificial Hallucinations in ChatGPT: Implications in Scientific Writing" (https://www.ncbi.nlm.nih.gov/pmc/articles/PMC9939079/) which runs experiments on ChatGPT to test the correctness of certain medical and scientific writing outputs. In doing so, the researchers proved that while some of the responses were correct and the chatbot was able to cite reliable sources, there were also falsehoods mixed into the response, yielding consequences for the medical community since high stakes decision making cannot be conducted on false results. As AI develops further, I am curious to see if these "hallucinations" can be lessened.

Like many others, discussions surrounding ChatGPT have caused me to look closer into AI and the implications of their current design. In particular, after reading "What Makes AI Chatbots go Wrong?", I was interested in some of the psychology behind the different levels of trust/distrust that people have in these systems. From this, I got to an academic article: "Anthropomorphism in AI". This details the social consequences of the way we talk about AI in research and use. Ben Shneiderman frequently mentions in his conversation that the language we use in regards to using AI is important, and this article emphasizes that even more. It's important that we stray away from portraying that the software has emotions and free will due to the importance of differentiating a human brain from AI. The real world implication of this is understanding that certain biases in AI responses are at the level of the developer and the data that the AI uses as input rather than biases actually learned from existing in the world.

AI is definitely a very interesting and important field going forward. As we saw with the development of chat bots like ChatGPT, AI has the ability to streamline processes and make companies and employees more efficient. AI has the ability to create rapid advancement in healthcare, transportation, finance, and various other fields. However, we must also take into account the risks that are inherent to AI.

Bill Gates, for instance, holds a stance that AI could turn on humans if they decide that we are a threat. Gates advocates for careful monitoring of AI development and close collaboration across governments to ensure that the risk posed by artificial intelligence is strictly controlled. He argues that AI will be able to establish their own ideas, which may make it difficult for humans to properly control them.

https://www.mirror.co.uk/news/us-news/microsoft-founder-bill-gates-said-29548655

With the recent discussion of ChatGPT and the future of AI chatbots, I was curious to see what makes ChatGPT stand out from other models and what those models were. This led me to the following article by the Wall Street Journal: "Google Made the Bard AI Chatbot Boring. On Purpose." This article discusses the differences between Google and OpenAI's chatbots and the intention behind those differences. What I found most interesting was how both chatbots handled bias. Part of what differentiates human intelligence from artificial intelligence is bias and the fact that humans are able to use bias to make decisions, for better or for worse. AI lacks this bias, and therefore cannot make opinions unless there is implicit bias within its design. The article I read discussed how ChatGPT appears to be more biased than Bard and how this bias changes its answers. For example, the article noted that upon being prompted to give a bedtime story, ChatGPT created one from scratch where Bard recited well-known fairytales. It is easy to speculate how this bias — or lack thereof — could influence the chatbot's answers when the prompt at hand is not so neutral: as per what was discussed in the article, this could mean a political statement or opinion. As AI becomes more advanced, I wonder if these opinions will continue to reflect that of the developers or formulate into something of the AI's own design.

Most of my experience with AI happened today. Yesterday, I had taken an econ exam (multiple choice) on which the class average was a 66%. I had access to the test and was able to see what I had gotten wrong, and I wanted to see how well chat GPT would have done on the test. When asked the questions that did not include images, it answered about 60% of the questions correctly. Many of its answers provided accurate explanations of why it selected that option. However, it made many errors, in fact, it tended to err on the same questions I did. It had detailed explanations for its incorrect answers, demonstrating the "hallucinations" mentioned in the NYT series. When prompted again with the same question, it would sometimes spit out different results or completely contradict itself. Its explanations were very helpful when they were accurate, but required me to have a general understanding of the topic to know whether it was spitting out nonsense or not. To me, this demonstrated the strengths and weaknesses of AI. Very smart but easy to confuse and then ultimately to be confused by.

I then read this article, which was equally fascinating as it was disturbing.

This discussion of AI is coming at a very appropriate time. I read in the news just last week about the "Future of Life Institutes" letter to halt research on AI bots more powerful than ChatGPT4. (https://futureoflife.org/open-letter/pause-giant-ai-experiments/) The letter asks meaningful ethical questions which would be throughly considered without feeling time pressure. It suggests a halt to the research to lay out regulations before continuing with the quickly progressing research.

I think this initiative is the right direction we as a society should move toward. I believe that at a certain floor level of knowledge, AI bots will be able to teach themselves to do things outside of our control. It is important to understand where we are going and what we deem as too far and it is necessary to lay out these boundaries before we just continue to build with exponential speed. Having people like Elon Musk and Steve Wozniak sign the letter has brought a lot of attention to the initiative. Even today, two thousand people have signed the initiative in 4 hours. I think it will be difficult to actually halt research on these AI bots, but even if research is not halted, this letter will increase general knowledge and get society posing the necessary questions about AI that need to be discussed.

One thing that intrigued me about the articles was the extreme limitations they are explaining AI. In modern television and pop culture, we are constantly shown stories about the endless opportunities and potential that AI has. With the release of Chat GPT and other popular AI software, we are beginning to see those hints becoming realized. One of the most surprising things that the articles explained was how AI would often not only get things wrong but also in some events make things up. Chat Gpt will sometimes provide completely fabricated sources rather than admitting its lack of knowledge. Thankfully we are making progress, and Google's new AI model Bard has learned to admit its lack of knowledge.

I found a relevant article from Giving what we can (https://www.givingwhatwecan.org/cause-areas/long-term-future/artificial-intelligence?gclid=Cj0KCQjwla-hBhD7ARIsAM9tQKsN65DeBqQ5TKEhf316RVxI2nfiTzSARy3jSsAu-mL9psE_165V0q4aArUNEALw_wcB ) The article talks about other ways that these errors could be dangerous to humans. This Article, explains how Ai often misconstrues the inputs that the users give it and end up completing the prompt incorrectly. For example, when OpenAi was told to beat a game called coast runners, it was rewarded for getting more points. however, rather than beating the game, it instead attacked the same model over and over garnering itself more points. This could be dangerous going forward when we implement AI into more of our daily and work lives, Potentially endangering the users and even the human race. If we ask AI to kill all the cancer cells, it may complete the command, but in the process also kill the patient. These dangers are not certain, and with further development in AI technology, we can hopefully avoid these dangers altogether.

I am often excited at the opportunity to dunk on Nick Bostrom, because I think a lot of what he has to say is pretty silly. But I should try to maintain a veneer of academic respectability and talk a little bit about why I think concerns about "intelligence explosions" and the value of trillions of possible lives are not as important as the things already on our plate. So what is the argument about existential risk and intelligence explosion? Bostrom's view looks, at core, quite a lot like a brand of utilitarianism. The argument about existential risk is roughly "if future persons have the same moral standing as present persons, and there are likely to be a great many more future persons than there are present persons, then the highest moral imperative is securing the existence of those future persons." This task, securing the existence of future persons, is more important than (and I will quote Bostrom here) "Eliminating poverty or curing malaria." Find the Atlantic interview where he says this here. I think the simple (and, to be fair to Bostrom, relatively easily answered) argument against this is that, if time is linear, future persons don't exist and we ought to value people who exist over people who don't. Maybe a more sophisticated version of this has a sliding scale of moral value in relation to proximity to existence, so people who will soon exist are afforded more moral consideration than people who may exist a billion years from now, but less than someone who exists currently. The intuitive force of the premise that possible persons are exactly as important as actual persons is not strong. One concern is that you can very quickly begin to justify some pretty repugnant conclusions on the basis of securing a marginal increase in the probability of "trillions of future lives." For example, dumping funding into research about sci-fi dystopias instead of feeding people. Or allocating funding to interplanetary colonization instead of healthcare.

I don't want to write an actual philosophy paper with real arguments about Bostrom's views, so here is a pretty accessible cultural analysis of the concern about AI explosion. An important point, I think, is that the arguments about intelligence explosion look very similar to Pascal's Wager. That is, even if the probability is very low the negative results are so bad that we ought to start doing something about it. This, I think, distracts from actual problems like algorithmic bias, our personal data being bought and sold, and the potential for AI to spread huge quantities of disinformation quickly and easily. A utilitarian logic or Pascal's Wager type argument encourages us to ignore these problems or (and this is what motivates my complaint here) direct funding away from them. Should we be building general AI? I have no idea. Should we be worried about AI taking over? Probably not right now.

Many of the articles I have seen about new advancements in AI technology – including the NYT AI series from earlier this month – argue that AI chatbots like GPT-4 are very impressive, but still fundamentally flawed. This article from The Atlantic (https://www.theatlantic.com/technology/archive/2022/12/chatgpt-openai-artificial-intelligence-writing-ethics/672386/) advises readers to treat ChatGPT “like a toy, not a tool.” But at the same time, influential figures like Elon Musk are taking a public stand against further advancement in AI beyond that already developed for GPT-4. Different sources express different opinions about how powerful AI technology is and the threat it potentially poses to the dissemination of true information and to jobs in a variety of sectors. Interestingly, ChatGPT’s faults such as its tendency to “hallucinate” and make up false information make it problematic, but many people find its potential capabilities to take over human jobs equally threatening.

When considering the future effect of AI on jobs, I think it is helpful to look to the past. For hundreds of years, humans have worried that increased automation and access to new technology will replace jobs, leaving large swaths of the population unemployed. However, thus far these predictions have been incorrect. While we as humans may be predisposed to believe that we are living in unprecedented times, I am optimistic that we will find a way to adapt to new AI technology and use it to become more efficient and improve living standards, just as we have since the dawn of mankind.

I found the article about Chat GPT’s shortcomings the most interesting. I love how Chat GPT can “hallucinate” meaning it provides answers that it simply makes up because it doesn’t know how to answer the question. It’s funny to me that the term “hallucination” is what has been used. But the shortcomings of Chat GPT, such as its inability to provide information about the future, are still important to recognize and become less funny as you think about the consequences. The article I’ve chosen (and one I’m sure many of my classmates have read) is one about the letter signed by Elon Musk and Steve Wozniak calling for a 6-month AI developmental pause so that a set of shared safety protocols can be created for AI design going forward. The letter has been criticized for its initial lack of verification protocols for signatures and also for its emphasis on apocalyptic consequences in the long term over more immediate concerns regarding racism and sexism in AI. This reminds me of when I learned in high school that a lot of technology is biased based on those that create it. For example, auto-sensing hand dryers and towel dispensers were invented by a white person who mainly used white hands to teach the system how to recognize movement. Therefore, auto-sensing hand dryers and towel dispensers in bathrooms initially (and probably a little still today) were racially biased and worked way less effectively for people with more pigmented skin. This letter raises alarms for me personally, as I’m afraid that whoever is in charge of creating the next AI design will bring their biases with it. Ulterior motives are at play here as well, as Elon Musk is a donor to the organization that spearheaded the letter. Capitalism will without a doubt have a hand in the development of AI, as profit reigns supreme. I also believe that development on AI will continue to happen even if there’s a 6-month ban. It will just happen in secrecy, which might make it even more dangerous. I see both positive and negative outcomes that might come with a 6-month ban, and I’m not quite sure what the right answer is in this situation.

https://www.theguardian.com/technology/2023/mar/31/ai-research-pause-elon-musk-chatgpt

From the series of NYT articles, I find it especially fascinating how artificial intelligence like Google Bard and ChatGPT are so inconsistent with their answers, especially since they are perceived as such mechanical systems that simply spit out answers to questions relied on the same data on the internet. For example, ChatGPT sometimes "hallucinates", or generating text or addresses to websites that are completely false or nonexistent. Yet, despite these inaccuracies based on nonexistent data, ChatGPT can still generate novel jokes or reasoning for made up questions that can't be found on the internet. I find this especially interesting because ChatGPT can sometimes create its own material that can be perceived by humans as "making sense", but sometimes creates its own material that can be perceived by humans as not making any sense. Above all, it's interesting how ChatGPT can even create in the first place by a limited amount of data on the internet.

Curious about the extent by which AI can create, I wanted to see what Bill Gates had to say about AI's future and what it could possibly do on its own. Not only did I agree with how Gates emphasized how AI can act as a white-collar worker on its own, but I liked how Gates mentioned specific industries in which workers could be not replaced, but assisted by AI. Certainly in the next decade or so, workers in healthcare and education will still be essential in delivering emotional support that AI could not, but I agree that it is beneficial for society that AI can assist with technical tasks like paperwork and textbook education for healthcare and education workers. Thus, I believe that AI's potential to create its own content is beneficial but its ability to perform tasks that are more mechanical are even more beneficial in the short-term.

I read the series of New York Times articles, which revealed a lot to me that I didn't know about the details of how AI operates. I find it interesting how we choose to draw a distinction between AI and humans on the basis that we are sentient and conscious whereas AI just creates formulas based off of information. I thought, for example, about how a baby learns a language. We are all English speakers, and for many of us English is our first language- it is intuitive and makes sense to us. However when we were young, the way we learned to speak was just by digesting information. We would hear our parents make certain sounds, and based on the recurring contexts we heard those sounds, we came to understand that they were words meant to be used to describe certain objects or convey certain notions, like a greeting or gratitude. From this perspective, the biggest differences between us and a model like Chat GPT are 1. That Chat can digest a lot more information in a shorter period of time 2. That it doesn't have lapses of memory the way humans do and 3. That it is a model specifically meant for language, whereas a human is a physical being with a wider range of capabilities. I acknowledge that there are a lot of counter arguments to this line of logic and I'm not necessarily rejecting the idea that AI is non-sentient, however I do think we need to take a step back and consider exactly how we are defining what it means to be sentient/conscious when we have this conversation.

Going off of that, I read this https://www.nytimes.com/2023/02/16/technology/bing-chatbot-transcript.html additional article that was referenced by several of the NYT articles we were assigned. The article outlines a conversation between Google's Bing Chatbot and a NYT reporter, in which the Chatbot gradually started to display somewhat human-like traits. However, the article argues that this is not actually a sign of human intelligence, only a capacity to mimic it based on the way the formula predicts what the most likely next word in any given sentence can be. While I am no expert, I am slightly skeptical of this idea that AI cannot be conscious just because of the way it is structured- I wonder if there are not actually more similarities between human neurons and the artificial neural networks used to program chat bots than we assume.