Recent artificial intelligence (AI) developments have generated much interest about safety, both in the potential for misuse of AI and in the accidental risks that may emerge as unintended consequences. Some evidence suggests that even when designed to mitigate these risks (sometimes referred to as aligning AI with human values), large language models are capable of mimicking or faking this alignment, covertly pursuing misaligned goals, posing security and privacy risks, and disclosing sensitive, private, or illegal information. At the same time, human limitations (such as difficulty distinguishing between human- and AI-generated text), along with the expansion of AI-generated content, pose new risks of misuse or accidents.
AI has been around for a long time, so where does the renewed interest in safety come from? This is partly due to AI’s new affordances and four notable features.
- First, AI can exhibit abilities in areas outside those intended or considered in its design. Unlike classical programming, where machines excelled at the task they were programmed to perform, large language models trained to predict the next word in a text sequence have proved helpful in functions ranging from translation to writing computer code.
- Second, AI can be generative, producing novel output based on descriptive prompts expressed in everyday language, unlike classical computer machines that execute instructions only from prespecified scripts. In early 2024, it was reported that ChatGPT alone was generating 100 billion words a day: within one year, this would be roughly equivalent to the amount of high-quality text available on the Internet.
- Third, AI can personalize and customize outputs adaptively and iteratively. It can do it quickly and at scale, unlike classical programming’s outputs, which are one-size-fits-all and have limited opportunities for rapid and dynamic customization at scale. Personalization applications hold particular promise in education and healthcare.
- Fourth, AI is very efficient at discerning functional patterns in data that are hard for people to do. At the same time, classical programming can provide insights only from data guided by human intuition. One practical application of this feature of AI is the rapid progress in predicting protein folding in biology, which used to take humans much time and effort.
Download the full article to read it offline: