I'm proposing that all educators confronting AI—even writing teachers—ask students to generate an image. Unlike ChatGPT, which comes off as some kind of robot oracle, text-to-image generators show AI capabilities and limits in vivid color 🧵 1/4
AI's error-prone nature is blatant in image generation. Students may not know Robespierre from Richelieu but they know neither had 13 fingers. Getting multiple outcomes by default and "rerolling" for a new set helps elucidate AI's probabilistic nature 2/4
Unlike ChatGPT's simple text field, an interface like Leonardo.ai pulls back the curtain on LLM machinery. Let students choose the temperature (.3 or .8?) or model (PhotoReal or Pastel Anime Dream?) and see the complexity behind the black box in action 3/4
Images also illustrate the critical concept of AI averaging. Use it to explain the Pope in Balenciaga. Notice how attractive everyone is, then explore biases in training data as well as the more subtle "midpoint hottie" problem. Way easier to see this bias in images than text 4/4
Add comment