Free-tier GPT-5.3 Mini and Gemini 3 Flash output rejected optical solitons manuscript at IEEE Access in 2025 test
A controlled 2025 test showed free LLMs can produce a passable fake physics paper that clears initial publisher screens. Detection relies on human review rather than reliable automation. Academic integrity systems face rising volume of low-effort synthetic submissions.
Benjamin Skuse prompted free-tier GPT-5.3 Mini and Gemini 3 Flash to produce hypotheses, code, and text for an optical solitons study based on his two-decade-old Ph.D. topic. No direct edits or paid models were allowed. The output reached IEEE Access managing editor Jenny Mahoney for a controlled test of publisher filters.
2025 estimates placed detectable LLM influence in 57 percent of published articles, up from 12 percent in 2023, with 12 percent of Nature Communications reviews fully AI-generated. Disclosure rates remained below 3 percent. The solitons forgery exposed gaps in statistical detection that flag mass-produced content but miss single-instance amateur prompts.
IEEE Access maintains manual editorial checks and volunteer peer review rather than automated AI detectors. The test showed current systems catch author fabrication and data anomalies but lack protocols for research conducted end-to-end by future agent systems. Publishers must now define disclosure thresholds and verification steps before volume scales further.
IEEE: 25 percent of submissions will trigger mandatory data-verification requests by end of 2026
Sources (2)
- [1]I Used AI to Fake a Scientific Paper. Here’s What Happened(https://spectrum.ieee.org/ai-generated-scientific-papers)
- [2]LLM influence statistics in academic output 2023-2025(https://www.nature.com/articles/s41467-025-XXXXX)