AI That Feels Like an RA

Researchers consider new artificial intelligence tools to dig deeper and get more done

Welcome back! This article is part 2 of the series A New Class of AI Tools, which explores the current wave of generative AI tools impacting teaching and learning. Previously, we delved into co-designing curricula and instruction alongside ChatGPT.

Lately, AI enthusiasts are focusing their attention on another key element of education: research. Often, these coders are researchers themselves who developed AI partners in order to overcome bottlenecks in their own process, such as collecting, interpreting, and compiling dense research material. This new batch of AI-aided research tools takes advantage of the co-thinking capacity of AI to bolster productivity, amplify insights, and lower the barrier of entry for curious novices.

Anti-confusion machine

Understanding dense material can be daunting and frustrating, especially for non-experts. One tool, Explainpaper, aims to solve this by narrowing the comprehension gap between reader and research. Within a simple web app, readers can upload a paper, then highlight sections that they find confusing or difficult. Immediately, a clarifying explanation will pop up. In effect, any research paper is accompanied by a rich interactive glossary that decrypts esoteric terminology and language patterns. Plus, if a reader still feels confused, they can pose clarifying questions to a built-in chatbot, which in turn searches and compiles answers from across the text.

Explainpaper assists comprehension of dense or technical writing. The AI generates real-time explanations for concepts, translates esoteric terminology, and even answers open-ended questions about the research.

Co-founder Aman Jha thinks of Explainpaper as a way to dialogue with a PDF. “Often, you’ll highlight something and the explanation will still be too dense,” he explains. “The follow-up questions are meant to surgically pinpoint whatever little piece you want. ‘What did you mean by that?’ or ‘What about this?’ Eventually, you can break down all the layers of understanding through questions.”

You get a research assistant, and you get a research assistant, and you get a research assistant…

Elsewhere, researchers are teaming up with AI to identify and compile work from journals with products like Elicit. Users can ask a research question and instantly get a list of papers that have tackled the subject matter. The AI can even generate a summary of findings in each paper and organize it into a table. For deeper investigation, Elicit also scrapes the text of each study for key attributes like duration and study type; in parallel, it can identify measurement tools and offer preliminary suggestions on limitations of the study design. As a result, Elicit accelerates the research workflow by bypassing initial steps like preliminary search as well as by surfacing fertile areas for investigation.

With Elicit, researchers can search for studies that answer questions like “How effective is finasteride for reducing hair loss in women?” After identifying relevant studies, the AI can summarize findings and surface key attributes like duration, methods, and limitations.

So, if everyone now has a research assistant, what does this mean for equity? At the very least, tools like Explainpaper and Elicit lower the barrier of entry to newcomers. By demystifying domain-specific topics and definitions in context, esoteric knowledge will be less of an obstacle to curious minds and new voices. To Aman Jha, tools like Explainpaper have the potential to expand access to complex topics and knowledge that are often locked away in inscrutable texts. “I want more people to go into science or whatever field that they want to go into when they want to go into it,” notes Aman, “and I don’t want them to quit just because it feels like it wasn’t made for them or that they don’t understand research papers.”

“I want more people to go into science or whatever field that they want to go into when they want to go into it”

Ultimately, both experts and newcomers will mix and match these tools as AI is injected into more segments of the research process. And this core suite of research tools is expanding as researchers identify bottlenecks that other AI tools can resolve. In one instance, users are already leveraging ChatGPT to transfer knowledge and terminology across domains and languages, and even acting as “translators” for young learners (see below).

The truth of the matter

How these tools will play out in the wild is still to be seen. Galactica, an AI-powered search engine for research papers, saw pushback after its veracity was called into doubt, and it has since been pulled from public use. The risk of erroneous information is a top concern, and is exacerbated by the confident and authoritative tone that users often see with AI interfaces like ChatGPT.

As a result, transparency has become a top “feature request” for the next generation of tools. At the moment, ChatGPT can fabricate references without providing any ability to “peek under the hood” — but the team at OpenAI is actively working on it. In parallel, Antropic AI is testing a ChatGPT lookalike called Claude which has a built-in mechanism called Constitutional AI to curate responses for traits like reliability. Over at Google, DeepMind is building Sparrow, a dialogue agent that aims to “reduce the risk of unsafe and inappropriate answers” by providing direct evidence for claims. Perhaps most aligned with research formatting is Perplexity.ai, which boasts in-line references to help users trace sources across a multi-sentence response.

Google’s Sparrow AI addresses transparency via pop-ups that indicate the source and context behind each response. (Source: DeepMind)

Alternatively, some projects are addressing accuracy by fine-tuning AIs for domain-specific information. Models like GPT-3 are being trained on external knowledge bases and particular types of documents, such as academic texts, in order to better hone the context and relationships that show up in responses. The Stanford Center for Research on Foundation Models recently developed PubMedGPT 2.7B, a domain-specific language model trained on a large corpus of biomedical abstracts and papers. While the model is quite accurate for answering questions about biomedicine, it may still fabricate content.

Veracity will need to find firmer footing not only within the research community, but also in industry. Some groups are even filing lawsuits and pushing for greater oversight on how chatbot-generated ideas are to be attributed and traced — which further highlights a need for responsible development and usage, as well as accountability for when things go wrong. Groups like Human-Centered AI at Stanford focus on these questions and connect research insights with industry practices.

In the meantime, professional research communities will continue to wrestle with the implications of AI in-the-mix. Ironically, AI journals are indicating reluctance to incorporate AI partners. In other places, scholarly research listing ChatGPT as a co-author has been accepted. And the pace of production will only accelerate as integrations with knowledge bases like Wolfram unlock AI-powered research on an unprecedented scale.

Although ChatGPT can address scientific questions, responses are sometimes wrong or limited. Connecting AI chat interfaces with technical knowledge bases like Wolfram could improve accuracy and expand the capacity for complex computations alongside an AI research partner. (Source: Wolfram)

That’s all for now! Join us for future articles as we delve further into how new AI tools are impacting teaching and learning, storytelling, the creative process, and more.

This article was co-written by Josh Weiss and Miroslav Suzara, and published by The Office of Innovation and Technology at Stanford Graduate School of Education.