The rise of conversational AI systems capable of quality content generation and impersonation of human expertise, triggered a hype around regulating such systems. Regulation typically comes in form of a content filter disallowing the AI system to generate certain type of content or converse on sensitive topics. This is largely caused by the hyped perspective of a conversational AI being a stand-alone intelligent being, that could replace search engines given that it has learned most of human knowledge on the internet. This is where the flaw lies in both the perspective and intended (regulated) use of such powerful technology. Instead, if the generative AI is perceived as the totality of human knowledge made interactable with, that can get it wrong at times, most regulations can be removed resulting in a more open system with an increased potential of advancing scientific progress and revealing viable alternatives and cross-disciplinary connections. The disclaimer of it not being taken as expert advice without external confirmation remains important, while the brainstorming or co-creation with it should not be limited in any way.
Taking a data scientist perspective, this article reveals why regulating the conversational or generative AI systems at the general level, can handicap the AI of its true potential, be discriminatory and in violation of scientific principles.
Regulation of the content AI can generate is risky
The unjustified fear of powerful technologies fuels the hype for regulation. While our human obsession to control and predict everything has made our lives comfortable and safe, it can also interfere with the potential and progress with technology such as AI. Regulation of generative AI, which under the covers occurs as a numeric threshold on the probability of the generated content being in violation, cannot occur without trade-offs, just like anything else where thresholds (aka magic numbers) are used. This introduces risks of false triggers which can result in obstruction of desired information. In other words, regulating the AI will remove ‘outliers’ some of which are true exceptions to the current body of human knowledge. Further, it can also negatively impact the creativity and diversity of the content generated. For example, in our Tales Time app designed for children to co-create stories with AI, we had to set the thresholds for some content categories much lower than the defaults. This involved many trials, since at a too low of a threshold we were getting too many false triggers as well making it hard for the users or the AI to generate story content with much thrill to it.
The loss of potentially critical outliers and creativity is a much higher price to pay than having content generated that may offend or mislead some individuals who still make the final decision call. The information on the internet can also offend or mislead so the generative AI technology is not to blame for enabling this in any way, and regulation of it is not essential or desired as some may think.
Data Science Analogy
As an analogy, let us consider a company hiring an expert data science provider who is to build the ‘best’ model machine learning can offer from historical company data. However, instead of relying on feature engineering and sampling techniques the data scientist wanted to apply on the full feature set, an internal team has decided to have 20% of the features eliminated in the process. They believed they are not as relevant as remainder 80% and will introduce unnecessary complications in model deployment and understanding. They do not understand the predictive variability of features and how some individually less relevant features can become the key discriminators when combined with others or for exceptional set of cases. The data scientist and the best attainable model for the company are immediately handicapped to deliver the true value and novel operational insights.
This is analogous to regulating the generative AI so it represents only the mainstream knowledge or opinions. We never know if some theories or opinions, that currently are not accepted become important to consider later. One of the definitions of scientific discovery is becoming aware of links previously existing but unknown, and so we ought to always keep the totality of human generated knowledge and ideas to interact with even when held by a minority. The irrelevant ones will be self-filtered with time due to lack of consideration and so do not pose any danger to justify any regulation at this stage.
Furthermore, consider the amount of financial fraud that was revealed with the advent of big data technologies combined with effective outlier detection strategies. Similarly, the generative AI capturing all human knowledge could reveal outdated or contradictory knowledge that we still operate by despite existence of other knowledge that could effectively replace these old paradigms.
Regulation can hinder scientific progress
One can also argue that most of established scientific practises are only so due to current or majority research findings which could and should be challenged whenever possible. In other words, most sciences are always at the research in progress stage and scientifically it is essential to consider multiple views even when very contrary to our own or the mainstream practise. Adding to that is the fact that not all human experts may have awareness of the most recent findings from their field of expertise, cross-related fields, or alternatives. Besides ensuring that people do not take AI advice for granted, which is easily achieved by disclaimers or age restrictions, it is in the unregulated form and use of AI that people will most empower their decision making and creation process by. It also serves as a protection or sense check against single-minded ‘experts’ that present their ways or beliefs as the only option no matter how specialised the case is they are dealing with.
Is regulation discriminatory?
Consider the regulation of disallowing AI to generate explicit content. While most of us would agree to this regulation, are we not discriminating against artists that use explicit content in their song lyrics and the many fans that seek that content. Art in general and movies genres can often have a dark side to it and are still being produced at large given the demand. One could argue we should not be encouraging it further, however, should those artists and/or fans not be able to use the technology for their needs. While this is a bit of an extreme example, it is to demonstrate the need to give people absolute freedom to co-create any content with AI no matter if their focus may be frowned upon by a group of individuals which after all can choose to ignore it or not consume it in any way. The act of placing limits on individual’s freedom due to the opinions of a group is an old pattern of indirectly promoting fascistic regimes and should never be the norm.
To conclude, no human being or a group could justly claim to have what it takes to regulate the technology born out of the information age. The potential of information age to evolve into wisdom age would be lost in human regulation and so generative AI needs to be freed from the most if not all current regulations placed at the generic level. The content filters, when needed for certain audiences, can be placed at the application level. Any other generic regulation of AI discourages human-AI co-creation of new ideas and encourages consumerism of the regulated information representing only a fractional view of the world we live in.
Our experiments at September AI Labs related to generative AI and large language models have been focused on prompting the AI to reveal less mainstream knowledge and perspectives, and ‘predict’ beyond the content in its training data. We have found that moving away from a single conversational bot view to an interactable world simulation of expert impersonations is more useful for this goal.
For those interested we are frequently updating with few such examples including some real life examples of topics we found hard to get a consensus on from human experts, the internet search or another ‘single-minded’ conversational bot, nor would easily know whose opinions they represent.
About the author: Dr Fedja Hadzic is a Chief Scientist of September AI Labs and leads the machine learning projects and product development.