Do advanced AI systems have what it takes to do high-stakes tasks in industries like computing and technology?

With recent public discussions of AI adoption occurring across industries, the answer is not as simple as whether AI is capable of performing tasks but also whether allowing AI to perform tasks introduces too much risk.

To examine this question, an interdisciplinary team of IGPA scholars in the Science and Technology Working Group just published a new preprint examining the economic risks of the use of AI systems in computing-centric careers in comparison to AI capabilities.

The main findings of the study indicate that while some risks of using AI may decline as models improve (e.g., AI hallucinations), more malicious use and security risks may actually persist and increase. As a result, the study observes that any gains won by increasing AI system capabilities can be uneven, or “jagged” in comparison to the risk if deployed in industry settings.

Higher levels of observed risk in using AI across several job functions highlights the need for serious consideration from both organizational and policy perspectives as AI adoption continues across industries.

Read the paper here: https://ggndpsngh.github.io/files/jagged.pdf

Deepika Chawla, Gagandeep Singh, Elham Buxton, Meicen Sun, Lav R. Varshney, Jeremy Riel, & Craig De Voto