Newspaper Breaking News artificial intelligence AI takeover. (© Crovik Media - stock.adobe.com)
BATH, United Kingdom — When it comes to artificial intelligence taking over the world, a new study says that everyone needs to calm down and stop fearing advanced technology. Researchers in the United Kingdom have found that AI language models like ChatGPT do not pose a serious threat to society — challenging the prevailing narrative surrounding the rise of AI-powered systems worldwide.
The researchers uncovered evidence suggesting that the much-touted “emergent abilities” of large language models (LLMs) may be less mysterious than previously thought. The study, conducted by a team from the Ubiquitous Knowledge Processing Lab at the Technical University of Darmstadt and the University of Bath and published in ACL Anthology, proposes that these seemingly new skills are actually the result of enhanced in-context learning rather than spontaneous emergence. Simply put, these machines are only learning what information humans feed into them, not learning on their own.
The AI community has been captivated by the idea that LLMs like GPT-3 and its successors develop unexpected competencies as they grow in size and complexity. This notion has sparked both excitement about potential applications and concerns regarding safety and control. However, the new study throws cold water on these speculations, offering a more grounded explanation for the observed phenomena.
“The fear has been that as models get bigger and bigger, they will be able to solve new problems that we cannot currently predict, which poses the threat that these larger models might acquire hazardous abilities including reasoning and planning,” says Dr. Tayyar Madabushi, a computer scientist at the University of Bath, in a media release.
“This has triggered a lot of discussion – for instance, at the AI Safety Summit last year at Bletchley Park, for which we were asked for comment – but our study shows that the fear that a model will go away and do something completely unexpected, innovative and potentially dangerous is not valid.”
At the heart of the research is the concept of in-context learning, a technique that allows AI models to perform tasks based on examples provided within the prompt. The study suggests that as LLMs scale up, they become more adept at leveraging this technique rather than developing entirely new abilities.
To test their hypothesis, the researchers conducted over 1,000 experiments, carefully examining whether LLMs truly acquire new skills or simply apply in-context learning more effectively as they grow larger. Their findings reveal that many abilities previously considered emergent are actually the result of improved in-context learning combined with the model’s extensive memory and linguistic knowledge.
This revelation has significant implications for the field of AI. It challenges the unpredictability factor associated with emergent abilities, potentially alleviating some concerns about the safety and control of these models in critical applications. Moreover, it provides a clearer framework for understanding and potentially enhancing the capabilities of LLMs.
The study’s conclusions offer a more pragmatic view of AI development, suggesting that improvements in LLM performance are rooted in their ability to recognize and apply patterns from provided examples, rather than spontaneously developing new competencies. This insight could reshape approaches to AI research and development, focusing efforts on optimizing in-context learning mechanisms rather than chasing elusive emergent properties.
“Importantly, what this means for end users is that relying on LLMs to interpret and perform complex tasks which require complex reasoning without explicit instruction is likely to be a mistake. Instead, users are likely to benefit from explicitly specifying what they require models to do and providing examples where possible for all but the simplest of tasks,” Dr. Madabushi explains.
As the AI community grapples with these findings, the debate over the nature of machine intelligence is likely to intensify. While the study doesn’t diminish the impressive capabilities of modern LLMs, it does provide a sobering perspective on their perceived “intelligence.”
“Our results do not mean that AI is not a threat at all. Rather, we show that the purported emergence of complex thinking skills associated with specific threats is not supported by evidence and that we can control the learning process of LLMs very well after all. Future research should therefore focus on other risks posed by the models, such as their potential to be used to generate fake news,” concludes Professor Iryna Gurevych from the Technical University of Darmstadt.
Paper Summary
Methodology
The researchers tested 20 different models across 22 tasks to assess whether these so-called emergent abilities would still manifest when in-context learning was controlled for. They used two settings for their experiments: few-shot, where the model was given examples in the prompt, and zero-shot, where no examples were provided.
The tasks were carefully chosen to include both those previously identified as emergent and others that were not. By comparing the performance of models across these tasks in both settings, the researchers could determine whether the model’s success was due to in-context learning or if it was exhibiting a true emergent ability.
Key Results
The results of the study were clear: when in-context learning was controlled for, the models did not exhibit emergent abilities. Instead, their performance on tasks was predictable based on their ability to apply in-context learning. For example, tasks that required reasoning or understanding of social situations — often cited as examples of emergent abilities — were not solved by the models unless they were provided with examples in the prompt.
This finding has important implications for how we think about and use LLMs. If these models are not truly developing new abilities but are instead relying on in-context learning, then their capabilities are more limited and predictable than previously thought. This means that concerns about unpredictable, potentially dangerous abilities may be overblown.
Study Limitations
One significant limitation is that the study only tested models on tasks in English, which leaves open the question of whether the findings would hold true for models trained in other languages. Additionally, the study focused on tasks that are commonly used to assess LLMs, which may not capture the full range of potential emergent abilities.
Moreover, the researchers acknowledge that while in-context learning may explain many of the abilities observed in LLMs, it does not necessarily account for all of them. There may still be some abilities that emerge as a result of scaling that were not captured in this study.
Discussion & Takeaways
The findings of this study invite us to rethink how we perceive the capabilities of LLMs. Rather than viewing these models as potentially developing new and unpredictable abilities as they scale, we should consider the possibility that their capabilities are more constrained and predictable, rooted in techniques like in-context learning.
This has practical implications for how we design and deploy LLMs in real-world applications. If in-context learning is the primary mechanism behind what we perceive as emergent abilities, then the focus should be on refining and controlling this technique to ensure that models are used safely and effectively. It also suggests that the potential risks associated with LLMs may be more manageable than previously thought.
Funding & Disclosures
This research was funded by several institutions, including the LOEWE Distinguished Chair “Ubiquitous Knowledge Processing,” the German Federal Ministry of Education and Research, the Hessian Ministry of Higher Education, Research, Science and the Arts, and the Microsoft Accelerate Foundation Models Academic Research fund. The researchers involved in this study have disclosed no conflicts of interest.
