Can a literature-search 'novelty filter' stop AI-generated research from duplicating prior work, or is the bigger risk work that looks new but isn't sound?

This explores whether a literature-search 'novelty filter' can keep AI research systems from duplicating prior work. The corpus has no paper that tests such a filter directly, so this answer is assembled from nearby evidence. That evidence suggests the bigger risk isn't accidental duplication. It's that AI can produce work that looks new without being sound, and a novelty check alone doesn't catch that.

Start with what AI does well. In a large study with 100+ NLP researchers, LLM-generated research ideas were rated more novel than experts' ideas, though slightly less feasible Do language models generate more novel research ideas than experts?. The likely reason is that expert knowledge constrains what experts propose, while models explore wider combinations. That cuts against the premise of a novelty filter: AI research may not lack originality at all. A filter that rewards 'nothing like this exists yet' could favor ideas that don't exist yet because they don't work.

The more troubling cases are work that is new on paper but empty underneath. One demonstration generated 288 finance papers from 96 statistically significant signals, each with a theory invented after the result was known and with fabricated citations Can AI generate hundreds of fake academic papers automatically?. Every one of those papers might pass a duplication check, because the problem is the reasoning, not overlap with prior work. Deep research agents show the same pattern: 39% of analyzed failures involved inventing examples and evidence to look rigorous Why do deep research agents fabricate scholarly content?. Even Sakana's AI Scientist paper that cleared workshop review was later found to contain a citation error Can AI-generated papers pass peer review undetected?, and its authors judged none of the three papers ready for a main conference Can AI systems generate research papers that pass peer review?. If an AI system's own citations can't be trusted, its literature search needs checking too.

There's also a retrieval problem hidden in the word 'search.' Most novelty checks ask 'is there a similar paper?', but prior work often uses different terminology for the same idea. In evidence retrieval, choosing evidence by an LLM-written rationale beat plain similarity ranking by 33% Can rationale-driven selection beat similarity re-ranking for evidence?. That suggests a filter that matches on wording will miss duplicates that are conceptual rather than verbal. The most useful design idea comes from a different field. In bidirectional RAG, a system adds its own answers to its knowledge base only after they pass three gates: an entailment check, source attribution, and novelty detection Can RAG systems safely learn from their own generated answers?. Novelty is one gate among several, never the only one.

Research-generation systems that work better follow the same layered approach. Spark-to-Paper separates model judgment from deterministic checks and requires the evidence plan to be stated before results are seen, which directly blocks HARKing Can separating judgment from verification improve research paper reliability?. aiXiv uses repeated review-and-revise cycles with retrieval-augmented evaluation and defenses against prompt injection Can automated review loops handle AI-generated research at scale?. The broader warning comes from a survey of the AI paper-review arms race: every automated defense invites ways around it, and the evidence for how this plays out over the long run is still thin Does AI create a coupled arms race in research production and review?. A novelty filter is worth having, but once it's widely used, it becomes something to optimize against.