I promised that when I started out writing this series, I didn’t mean to make the consideration review so long. I really am excited about the innovative possibilities of AI. ITC makes use of AI internally and for our customers. I do hope that this transformation and the innovative pace continue at the level it is. If we understand and incorporate lessons from the downsides, we really can make AI work to the betterment of humanity.
Plagiarism and Copyright | ChatGPT passes most current plagiarism tests for detecting when people wholesale borrow text as their own. Although this is an issue (especially for teachers), this isn’t strictly the kind of plagiarism we are worried about.
The articles written at happycoding.ai reflect my concerns. When you drill down, it becomes apparent that ChatGPT is regurgitating existing code without source provenance. If you asked ChatGPT, it will tell you that it can’t plagiarize as it’s a model. This is disingenuous. It is certainly possible and probable that it’s using copyrighted data. Given the scale of ingest, curation is an issue. Stable diffusion although not strictly ChatGPT (Stability AI created it and DALL-E another OpenAI product is similar with a different model), has its own issues in terms of borrowing art, manipulating it and passing it off as new without credit to the artist.
Bias |Bias comes in a lot of forms. It can be built into models, it can come from the data in obvious and hidden form, and it can come from the scientists curating the data and building the models. ChatGPT is no different in this respect.
Data is messy, it comes with confidence levels and scientists typically use error bars to denote this. Ask yourself who curated the data, who labeled it, were they paid for it, where did it come from, what was the reason for the model, who benefits from it, etc. Would it be better to train your own model on your specific data? Probably, if it’s domain specific. Given the bias in ChatGPT’s case largely has to do with their dataset, as a user what do I do since I can’t control the curation of their data set? Awareness and education is key as you need to know that there can be a biases result in anything that it produces.
ChatGPT’s founder says it will end capitalism as we know it (although at the same time taking a massive investment from Microsoft.) ChatGPT has “morality” filters applied and it will refuse to answer some controversial topics, although easily bypassed by asking it to write a book about a topic or create a fictional account. Through testing of ChatGPT, it seems to be more likely to classify statements as negative comments about demographic groups that are identified as disadvantaged. (https://davidrozado.substack.com/p/openaicms) Is this a reflection of the training set or the internet in general? Other studies have shown that ChatGPT is also politically biased. It seems to come out as establishment left with a libertarian bent. (https://the-decoder.com/chatgpt-is-politically-left-wing-study/)
Where does my data go | Is that data being used to train ChatGPT in a feedback loop? Would those paragraphs pasted into a question or used for summarization be used in other responses ChatGPT gives? Can we accidentally leak confidential information or PII? It certainly says thank you I’ll update my response when you correct it. (A different issue when you tell it 3+3=7 and not 6) In their FAQ ChatGPT explicitly says they do review questions to improve the system and although they have a data deletion process but what if it was already incorporated into the pre-trained model?
Personalized Manipulation |We’ve seen over the last couple of years bot farms cranking out foreign national sponsored mis/dis-information to create FUD in order to manipulate our election cycle. This manipulation is rudimentary but quite effective. We’ve seen an early form of this throughout history in the form of advertising and during wartime as PsyOps or propaganda. Conversational AIs such as ChatGPT combined with other AI analysis techniques like micro expression analysis enables manipulation to a new level. Coming from a psychological starting point, analysis of individual emotions wants and needs and an AI with the ability to respond with direction and bias to feed those emotional needs is extremely dangerous. On the positive side of this it could be an excellent tool to understand your organization and drive them towards a common goal and alignment but could be a slippery slope towards outright manipulation of society. A recent article from VentureBeat (https://venturebeat.com/ai/the-profound-danger-of-conversational-ai/) laid this issue out pretty well, so I’m going to quote from it directly:
“I’m deeply concerned about a different type of control problem that is already within our grasp and could pose a major threat to society unless policymakers take rapid action. I’m referring to the increasing possibility that currently available AI technologies can be used to target and manipulate individual users with extreme precision and efficiency. Even worse, this new form of personalized manipulation could be deployed at scale by corporate interests, state actors or even rogue despots to influence broad populations.”
ChatGPT in its current form should be seen as a glimpse of what is yet to come. Although it was tempting to use ChatGPT to write a large portion of this article, I decided against it. It could certainly write an article about itself, as well as some of the benefits and drawbacks of its use. If you want that view, just ask!
For more on ITC, how we are moving with the forefront of AI, ML, and RPA technologies, reach out to info@useitc.com!