Summary: A new study reveals that large-scale language models (LLMs) like ChatGPT are predictable and controllable, as they cannot independently learn or acquire new skills without explicit instructions. The study allays concerns about these models developing complex reasoning capabilities, highlighting that while LLMs can generate sophisticated language, they are unlikely to pose an existential threat. However, caution is still advised about potential misuses of AI, such as generating fake news.
Key Facts:
LLMs cannot acquire new skills without explicit instruction. The study found no evidence of the emergence of complex reasoning in LLMs. The concern is not an existential threat, but the misuse of AI.
Source: University of Bath
ChatGPT and other large-scale language models (LLMs) cannot learn or acquire new skills independently, and therefore do not pose an existential threat to humanity, according to new research from the University of Bath and Technische Universität Darmstadt in Germany.
The study, published today as part of the proceedings of the Association for Computational Linguistics’ 62nd Annual Conference (ACL 2024), the premier international conference in natural language processing, reveals that while LLMs have a superficial ability to follow instructions and excel in linguistic skills, they are unlikely to learn new skills without explicit instruction — meaning LLMs are inherently controllable, predictable and safe.
This means that they are inherently controllable, predictable, and safe.
The research team concluded that LLMs, which are trained on increasingly large datasets, can continue to be deployed without safety concerns, although there is potential for the technology to be misused.
As they grow, these models may produce more sophisticated language and improve their ability to follow explicit and detailed prompts, but they are unlikely to acquire complex reasoning skills.
“The common view that this kind of AI is a threat to humanity is hindering the widespread adoption and development of this technology and distracting from the real problems we need to focus on,” said Dr Harish Tayyar Madhavshi, a computer scientist at the University of Bath and co-author of a new study on “emerging capabilities” among LLM students.
A collaborative research team led by Professor Irina Gurevich of the Technical University of Darmstadt in Germany conducted experiments to test the LLM’s ability to complete tasks that the model had not encountered before – known as emergent capabilities.
As an example, LLMs are able to answer questions about social situations without being explicitly trained or programmed to do so. While previous research had suggested this was a result of the models “knowing” social situations, the researchers showed that it was actually a result of the models using LLMs’ well-known ability to complete tasks based on a few examples that they were presented with – “in-context learning” (ICL).
Across thousands of experiments, the researchers demonstrated that a combination of LLMs’ instruction-following ability (ICL), memory, and language abilities explains both the capabilities and limitations they exhibit.
Dr Tayyar Madabhsi said: “There is a concern that as models get bigger and bigger they will be able to solve new problems that we cannot currently predict, and as a result these large models could gain dangerous capabilities such as reasoning and planning.”
“This has generated a lot of debate – for example, we were asked to comment at the AI Safety Summit at Bletchley Park last year – but our research shows that fears that the models will go off and do something totally unexpected, innovative and potentially dangerous are unfounded.
“Concerns about the existential threat posed by the LLM are not limited to non-specialists but have been expressed by some of the leading AI researchers around the world.”
However, Dr Tayyar Madabshi argues that this concern is unfounded as the researchers’ tests clearly showed that law students lacked complex reasoning skills.
“While it is important to address existing possibilities for the misuse of AI, such as the creation of fake news and increased risks of fraud, it would be premature to enact regulation based on perceived existential threats,” he said.
“The point is that it is likely a mistake for end users to rely on LLM to interpret and perform complex tasks that require complex reasoning without explicit instructions. Instead, users are likely to benefit from being explicitly told what they want the model to do, and from providing examples, whenever possible, for all but the simplest tasks.”
Professor Gurevich added: “…our findings do not mean that AI is not a threat at all. Rather, they show that claims of the emergence of complex thinking skills linked to specific threats are not supported by evidence, and that the learning process in LLMs can ultimately be very well controlled.”
“Future research should therefore focus on other risks posed by the model, such as the possibility that it could be used to generate fake news.”
About this AI research news
Author: Chris Melvin
Source: University of Bath
Contact: Chris Melvin – University of Bath
Image: This image is provided by Neuroscience News
Original research: The findings will be presented at the 62nd Annual Conference of the Association for Computational Linguistics.
