In this last part of this Generative AI series, we will look at some of the issues Generative AI systems have raised and are struggling with. The post will in no way be exhaustive since it is too early to identify all issues at this point.
Hallucinations
The fact that generative AI systems - especially for text - produce text that looks so convincingly similar to human-generated text raises the false impression that the output of chats or dialogues with such systems produces correct information. However, such systems might produce a plausible-looking text which includes incorrect information. This phenomenon is called a hallucination.
In May 2023, the hallucination rate of ChatGPT 4 was assessed to be around 8% by OpenAI. AI companies are constantly introducing techniques to prevent or reduce the rate of hallucinations.
One of the reasons for this might be the lack of certain information or the introduction of biased data during training.
Another reason is the size of the context window for the LLM. If the context window is small, then the model can run out of space and as such could produce text that is unrelated to the previously generated text, since context is not remembered beyond the context window. However, there is some research that the hallucination rate might actually increase with a larger context size, thus the clear effect of this is still being debated.
Copyrights
Thousands of non-fiction authors sued Microsoft and OpenAI in November 2023, due to the fact that they have used copyrighted material to train the model and as such have infringed on the copyrights of these authors. This was a continuation of various lawsuits started by fiction others, some as prominent as George R.R. Martin and John Grisham about the same topic.
Although courts have not decided in favour of the copyright holders, OpenAI, as one of the main players in this area, has declared that they give owners of copyrighted works to opt out of their works being used as training data explicitly. They have also announced that they are refusing prompts to create images in the style of existing creators, thus reducing the chance of copyright infringement.
Who owns the generated text or images?
In most cases, applications state that any generated text or images belong to the user who generated them, although some more complex generative systems (such as logo and promotional material creation systems) may charge a one-time or recurrent fee to use the generated material.
Deepfakes and False Propaganda
With the advent of generative AI, the danger of “deepfakes” (realistic photos or videos of real people, generated by AI and providing false or misleading information) has increased. OpenAI has banned the use of public figures in the prompts in order to avoid this. They have also introduced measures to prevent under/overrepresentation of groups (i.e. you tend to get more minorities in the images, due to invisible prompt elements OpenAI uses in the background), due to the fact that prompts have been deliberately designed to impose the idea that a group or community is more suitable or not suitable for certain social situations, thus enabling false propaganda about these groups or communities.
There are also some tools to detect AI-generated material and I have a hunch that this will be an important business area in the near future.
Effective prompting
The quality of the responses you get from generative AI systems very much depends on the quality of your original prompt. A totally new profession called Prompt Engineering has been created to design effective prompts.
The reason why I labelled this as an issue is that different people with different (or no) prompting skills will get very different output from AI systems, thus treating low-quality output with falsities as valid information.
There are multiple systems which provide structured input capabilities to the user and as such pass the structured information through an API to the AI system. With the release of GPTs by OpenAI or similar capabilities, there will be many options to improve the quality of the input and consequently the output of these systems.
Schoolwork
It is relatively easy now to generate a plausible essay about any topic using one of the available LLMs, thus it will be a challenge for teachers and professors to detect AI-generated text within the essays delivered to them, though AI-detection tools are in development. One way out of this debacle is a revision of homework and exam question structure, to make it more creative and less prone to AI generation.
Artificial General Intelligence and LLMs
An artificial general intelligence (AGI) is a hypothetical type of intelligent agent. If realized, an AGI could learn to accomplish any intellectual task that human beings or animals can perform.
As a result of the rapid development of LLMs and the apparently plausible results they can generate, some researchers think that we are closer to the establishment of AGI, whereas some researchers do not classify LLMs as intelligent and as such declare AGIs to be hypothetical at this time.
Part of the crisis around OpenAI was due to the speed of AI improvements that were being envisioned and the difference of opinion between different parts of OpenAI. This worry was also present within the U.S. Government since Biden issued an Executive Order to regulate some developments in complex AI models.
Future Prospects
Most researchers are concentrating on Large Language Models and using pre-trained neural networks with Deep Learning, given the huge successes in this area in a very short amount of time (last 2-3 years). However, I believe this is going to be limiting the type of areas and results obtained through AI. In the early years of AI, most work was in more theoretical AI areas (symbolic AI) with more structural approaches or models, but the results were not coming due to the complexity of the models and the lack of computing resources or technology.
Similar to the improvements imposed through structured prompting, it would be possible to improve models by using structured AI models along with pre-trained neural networks normally working without any human intervention.