i tested basically every AI research tool for my engineering capstone. most are complete garbage.

A senior engineering student working on a complex capstone project involving VLSI fault models and LTE diversity architectures has conducted a brutal, real-world test of popular AI research tools. The goal was to cut through the 'SEO-farmed vendor ads' that dominate Google results and find actual technical specifications. The findings reveal a stark divide between tools that are useful for serious research and those that are 'paywalled garbage' or 'hallucination engines.'

For general technical search, the open-source, privacy-focused Scira (scira.ai) scored 4/5 and became the daily driver. Its key strength is forcing strict, clickable inline citations to real PDFs, preventing the user from being 'gaslit by fake references.' For finding strictly peer-reviewed academic literature, Consensus (consensus.app) was rated 'god-tier' but is limited to that domain. Other specialized tools filled crucial niches: Elicit (elicit.com) for extracting methodology and p-values into spreadsheets, Scispace (scispace.com) for decoding dense math in papers, and ResearchRabbit (researchrabbit.ai) for building visual citation graphs from a single seed paper.

The major disappointments were raw chatbots like Claude (claude.ai), which was 'banned for raw search' due to inventing fake IEEE DOIs, and Perplexity (perplexity.ai), which was criticized for being 'nerfed' and lazily scraping the top SEO blogs instead of digging deeper. The student's final stack recommendation is to avoid raw chatbots for research, use specialized tools for decoding and data extraction, and rely on Scira to bypass Google's SEO ecosystem without falling victim to AI hallucinations.