Who is Aligned to Whom? Expert deferral to AI
As AI agents take on more of experts' work, who makes the judgment calls? In a randomized trial, experts mostly deferred to the AI agent's preferences, even though we had set those preferences at random.
As AI agents take on more of experts' work, who makes the judgment calls? In a randomized trial, experts mostly deferred to the AI agent's preferences, even though we had set those preferences at random.
Does your AI assistant tell you when your plan can no longer work, or when it would choose differently? Many models I tested went along without a warning, while Claude models spoke up, sometimes even when nobody asked.
Can you get more varied answers from a language model by giving it a persona, like a star sign? It worked for Google's models but not for OpenAI's.
Can language models stand in for people in social science surveys? Mostly they repeat findings from the papers they were trained on, and how closely they match real survey data depends on how you write the prompt.
When the 1934 Surname Law required every Turkish family to pick a surname, what did people choose? The Republic's political elite favored "pure Turkish" and Old Turkic names about strength and heritage, while ordinary families chose names about nature, looks and occupations.
What do expecting parents talk about, week by week? I mapped over 300,000 public forum posts from conception to the first months with a baby.
During COVID-19, I built a privacy-preserving dataset at Meta showing how much people in each region were moving or staying home. It became Meta's most-downloaded dataset on the Humanitarian Data Exchange and informed the response of the European Commission, the CDC and many others.
At Çilek Ağacı, we have analyzed and visualized pretty much every Turkish election and referendum of the last decade and a half.