Reconstructing the most likely prompt that generated a piece of AI output — used for detection, content auditing, and understanding how models respond to specific framings.
Reverse prompting is the practice of inferring the original prompt that produced a given AI-generated text. Given an output, a model trained for the task predicts the instruction or question most likely to have caused it. The direction of inference is inverted relative to normal generation: instead of prompt → output, the system reasons output → prompt.
At DEJAN we fine-tuned a Gemma 3 270M model specifically for this task, training it to reconstruct likely prompts from AI-generated responses. The model learns patterns in how different prompt types — informational, comparative, list-based, persuasive — produce recognisably different output structures, vocabulary distributions, and framing choices. Given a candidate text, it returns the prompt family most likely to have produced it.
Reverse prompting has several practical uses. In AI content detection, it provides a complementary signal: text that maps cleanly to a standard AI prompt pattern is more likely to be AI-generated than text that does not. In content auditing, it can reveal whether a competitor's pages were written to answer specific query types, giving insight into their content strategy. In research, it helps map which prompt framings consistently surface which brands — feeding directly into model probing methodology.
Reverse prompting is probabilistic. Many different prompts can produce similar outputs, and the reconstructed prompt is a best estimate rather than a ground truth. As AI models improve and produce more naturalistic text, the signal becomes harder to decode reliably. The technique works best as one layer in a broader detection or analysis stack, not as a standalone verdict.
