alt Hacker News>When one model refuses or stalls, I switch models, rephrase prompts, and ask what the next step would look like “hypothetically.”
I wonder If the hypothetically criterion gives the models carte blanche to make up references.
>When one model refuses or stalls, I switch models, rephrase prompts, and ask what the next step would look like “hypothetically.”
I wonder If the hypothetically criterion gives the models carte blanche to make up references.
The models don't know or need permission to make up references. References are just complicated things to predict in terms of next tokens -- Author names are found in many references, with many journals. Most hallucinated refs use researcher names in the field and credible article titles related to their research.
Agentic models now, however, can correctly search pubmed or OpenAlex and obtain valid refs, so that tell is mostly gone.
Wait until we get the first retractions of papers with AI generated micrographs or blots, and synthetic data plots. That's the bit that really scares me.