Ethical and Societal Concerns Regarding Large Language Models (LLMs)

Key Insight

Early on, an internal paper titled 'The Malicious Use of Artificial Intelligence' cautioned against AI's growing capacity to spread misinformation online and facilitate other harmful activities. This report played a pivotal role in an AI lab's decision to withhold the full code for its advanced language model, GPT-2, due to concerns about its potential for malicious applications. This precedent-setting move, intended to prevent misuse, was met with public criticism and negative headlines, yet it underscored a foundational apprehension within the research community about the societal implications of powerful AI technologies.

A significant critique of large language models (LLMs) emerged in a paper titled 'On the Dangers of Stochastic Parrots: Can Language Models Be Too Big?', introducing the term 'stochastic parrots' to describe LLMs' tendency for mimicry and guesswork over genuine understanding. This paper dissected key concerns, including the immense carbon footprints of LLMs due to their intense computational demands. It also highlighted how LLMs 'encode biases potentially damaging to marginalized populations' and introduced 'value-lock,' a concept explaining how the process of scraping and reconstituting existing content by LLMs can 'reify older, less-inclusive understandings,' thus perpetuating historical biases.

The 'Stochastic Parrots' paper specifically identified training data sources like Reddit and Wikipedia as potential origins for bias in models such as GPT-2 and GPT-3. Citing a 2016 Pew Research Center survey, it noted that Reddit's US users were predominantly young men aged 18 to 29, and similar surveys found only 8.8% to 15% of Wikipedians were women or girls. Furthermore, the authors warned of the risk of malicious exploitation, where 'bad actors' could leverage LLMs to 'produce large quantities of seemingly coherent texts on specific topics,' exploiting the human tendency to trust intelligible information to promote dangerous conspiracy theories or other harmful narratives. These concerns resonated with earlier internal fears, validating the caution exercised in the initial limited release of advanced language models.