80 examples is all it takes to reshape a large AI model's values

OpenAI has published new research demonstrating that a language model’s behavior can be significantly steered by fine-tuning it on a remarkably small, hand-curated dataset. The paper reveals that a set of just 80 text samples, formatted as question-and-answer pairs, was enough to improve adherence to specific behavioral values in GPT‑3 without degrading its performance on other tasks. The dataset, which is about 120KB in total size—a fraction of a percent of GPT‑3’s original training data—targeted categories with direct human impact, including violence, injustice, health advice, and political destabilization. The values were largely drawn from U.S. and international human rights law and Western social movements for equality. OpenAI found a clear scaling effect: the process became more effective as the models grew larger, from 125 million up to 175 billion parameters. This suggests operators of massive models could need surprisingly few examples to align behavior to their application’s specific needs. The company is now actively looking for API users to test the nascent technique in production, framing the selection of exact values not as a one-size-fits-all problem, but as a choice each developer must confront. Their qualitative probes showed the values-targeted models broadly adhered more to the desired behaviors, with statistically significant improvements measured through human evaluations, toxicity scoring via the Perspective API, and co-occurrence metrics for gender, race, and religion. The researchers are candid about the limitations, noting that toxicity scores carry known demographic biases, which is why they layered multiple evaluation methods. They also acknowledge that outlining values for large groups risks marginalizing minority voices, a tension they attempted to address by making the process scalable rather than requiring a full retraining from scratch.