Skip to content
Start main Content

Five AI Research Tools That Referencing Genuine Sources

Hallucinations, or the generation of inaccurate or fabricated information, are a significant concern surrounding the use of AI. In this post, we introduce five AI research tools that can generate answers based on real sources. These tools are designed to ensure that students and researchers can rely on accurate and reliable information when citing sources.

Limitations of GPT3.5/GPT4 in Referencing Evidence

GPT3.5 and GPT4 lack the ability to reference evidence. Although they provide seemingly authentic and confident responses, the accuracy of their arguments and assertions is highly questionable.

A recent study examined the performance of GPT3.5 and GPT4 by asking them 50 questions from diverse domains and requesting the top five sources or citations for each answer. The study revealed that 72.5% of the citations provided by GPT3.5 were fictional, and for GPT4, the figure was 71.2% – Quite an alarming result!

Citations to answers given by LLMs. Each row represents 5 sources / citations for a single answer. Overall, 72.5% of citations provided by GPT3.5 were fictional. This figure was 71.2% for GPT4

Citations to answers given by LLMs.

 

Scite Assistant

https://scite-ai.lib.ezproxy.hkust.edu.hk/assistant

  • Scite.ai offers a conversational AI Assistant that provides answers referencing real papers with real DOIs.
  • Sentences of the referenced part are highlighted from their full-text so we can see where and why they are relevant.
  • Users have control over the Assistant’s setting, such as specifying reference sources (from certain journals or Scite dashboards), reference year range, and the number of references to include in an answer.
Scite AI Assistant Setting

Users can specify their reference requirements in the settings of Scite Assistant.

 

Scite Assistant Answer Sources

Scite Assistant provides real sources, and when clicking on the sources in the answer, users can view relevant citation statements, as well as the source’s Smart Citation metrics.

scite dashboard papers research themes

I have a Scite dashboard of around 900 papers which are HKUST publications on SDG#3 “Good Health and Well-Being” between 2019 and 2022. Scite Assistant helped to summarize the papers into ten research themes and highlighted key papers.

SciSpace

https://typeset.io/

  • Similar to Scite Assistant, SciSpace allows users to refine results based on publication type, year, or towards a specific PDF.
  • The SciSpace Copilot Chrome extension is a convenient tool for answering questions and explaining concepts in a paper while browsing the web.
Scispace Lit Review Table

A mini-literature review on a research topic, where referenced papers are summarized into a structured table.

 

Scispace Copilot explaining concepts

The copilot plugin can answer questions from a paper and explain highlighted text, including mathematical concepts.

Elicit (Beta)

https://beta.elicit.org/

  • The beta version of Elicit.org supports more research workflows and offers more robust results.
  • Users can find papers related to their research question, extract information from PDFs, and discover concepts across papers.
  • Elicit provides structured tables summarizing relevant papers and allows users to customize and add additional details to the table.
Elicit Finding Concepts

Elicit can identify concepts from relevant papers. As shown on the left-hand of the screen cap, Elicit is relatively transparent by showing the number of papers analyzed to generate the results.

Elicit answers questions based on top relevant papers, presenting them in a structured table. Users can add more columns to the table and indicate key papers. Afterwards, the generated summary may be re-computed accordingly.  

 

Petal

https://www.petal.org/

  • Petal is a document analysis platform that goes beyond analyzing a single PDF.
  • It can answer questions and facilitate multi-PDF chat, allowing users to generate structured AI tables summarizing multiple PDFs.
  • Petal offers flexibility in customizing the columns of the AI table and allows users to edit cell information if the AI generates inaccurate results.
Petal Chatbot

Similar to SciSpace Copilot, the Petal PDF AI chatbot that can answer questions from a paper, even citing the specific page numbers.

More flexibility in generating AI tables for documents added to the library. Users can edit columns and cell information to correct any inaccuracies.

Consensus

https://consensus.app/

  • Consensus positions itself as an academic search engine rather than a chatbot.
  • Users can ask research questions, and Consensus synthesizes results from top relevant research papers.
Consensus Driver Question

In this example, it recognizes the complexity of the question “are men better drivers than women?” and acknowledges its limitations. However, it still finds five papers and highlights relevant findings.

Consensus Entrepreneurial Intention Question

In this example, Consensus does give a synthesized summary base on top 10 papers. It is concise and brief compared to other tools discussed in the post.

Remarks

This post focuses on highlighting the tools’ capabilities and is not an exhaustive technical review. It is important to refer to the tools’ respective documentation for detailed information about the underlying language models used, data sources, privacy policies, and other specifications.

Finally, let’s conclude this post with a table of key features of these tools:

Scite Assistant SciSpace Elicit Beta Petal Consensus
Can accept full text PDF upload?
No Yes Yes Yes No
Can summarize papers in structured table (lit review matrix)?
No Yes Yes Yes No
Can refine/customize reference sources?
Yes Yes Yes Yes Yes
Can ask chatbot specific questions on multiple papers?
Yes Yes Yes Yes No
Pricing Model
Monthly/
Yearly
Monthly/
Yearly
Token Monthly/
Yearly
Yearly
FAQ / About / Support
FAQ FAQ FAQ (to elicit) Documentation  FAQ How it Works

 

– By Jennifer Gu, Library

Hits: 38780

Go Back to page Top

Tags: , , , , , , , ,

published August 31, 2023
last modified January 22, 2024