What Is an AI Research Assistant and Why Do We Want It?

What if you had an assistant to help you rapidly answer any research question by planning and executing tasks? Internet-based searches, information retrieval, summarising relevant information, and even presenting it all in a coherent manner could be made simple.

Well, we’re on the cusp of making this a reality.

This is where AI research assistants will answer the call. Innovative new software promises to help make the most laborious parts of research easier and more accessible for everyone, advancing the reach of our best and brightest, while helping the rest of us catch up.

Two Great Use Cases - Literature Review and Market Research

A large portion of knowledge work revolves around gathering and repurposing information from the internet. This is a prime example of where AI research assistants will boost productivity. Here are two major use cases to illustrate this:

Literature review - Normally associated with academic settings, this task involves surveying the existing evidence around a topic. For example, we might ask:

• What is the state of the art in cancer therapies?
• What are the potential drug therapies for dementia?
• What are the mechanisms behind the well-known benefits of exercise?

Etc.

For each of these topics, substantial bodies of research publications exist. These must be continually revisited en masse to regenerate the picture of what is the state of the art, what we know and don’t know, and what are the most promising next directions.

To answer these questions we must analyse huge volumes of data (huge from the perspective of a human worker, at least). This involves the repetitive and laborious task of searching, highlighting, and extracting huge swathes of information before we can synthesise a view of the literature that provides the key insights in a digestible length and format.

These AI research assistants could also be invaluable for writing grant proposals. To justify your request for funds you need to paint a picture with a you-shaped hole in it. A research assistant would help draw upon the necessary background evidence to support that.

This is a perfect use case to employ language model-based technologies, to which we can delegate many of the repetitive elements of research and review.

Market research - A large portion of market research involves gathering and interpreting existing data (AKA secondary research) in order to synthesise a perspective on the market, understand the landscape, people, companies and products involved, the forces and trends at work, and finally, to formulate the predictions and actionable insights that initially motivated the research.

While superficially very different from literature review (in terms of information sources, target audience and outcomes), at its core it shares the fundamental research tasks of searching, summarising, and synthesising. Any sufficiently generalised research assistant technology will therefore be equally applicable to both market research and literature reviews.

Research Assistants as Copilots

One of the revolutionary features demonstrated by modern AI chatbots is the high degree of interactivity and flexibility which allows rapid feedback and iteration during a conversation. This affords control and steering during the process, as well as continuity - once you receive initial results, there is the potential for follow-up questions and actions.

Our vision is to apply these features to create an AI research assistant that understands where you’re going and helps you get there. It can be briefed and let loose upon a research task, but it can also change course in response to feedback and new information. Here we see an analogy between AI research assistants and self-driving vehicle technologies.

Analogy to Self-Driving Vehicles - Levels of Automation

In the field of autonomous driving, there is an established framework to describe the different levels of automation (see: Wikipedia: Vehicular automation (Autonomy levels)). This gives a common language to help us communicate about the problem and where we are in the stages of evolution.

Here we take this framework and repurpose it with respect to AI research assistants.

Levels of Research Assistant Automation

Level 0: Manual Research

No automation involved, all tasks are handled by the human researcher.

The user does all the research, from searching, filtering, and summarising to the final synthesis of information. We might have basic tools like a search engine, but we do not have any automation or smart assistance capabilities beyond that.

Level 1: Basic Assistance

The system provides basic, piecemeal assistance for isolated tasks.

E.g., the computer suggests keywords or refines your search based on your initial queries, like search engines suggesting related searches or articles. The user still has to evaluate and assimilate all the information.

Level 2: Enhanced Assistance

The system can provide significant assistance in most tasks but lacks end-to-end automation.

E.g., the AI can automatically generate queries based on your topic and provide a list of potential sources, it might be able to download and collate sources, and perhaps tag documents or excerpts with relevant metadata related to the task, but these come in the form of separate jobs/tools that a user would pick up and apply in an ad-hoc manner.

Level 3: Basic End-to-End Automation

The assistant can autonomously execute a basic research investigation, though with limited breadth/depth and sophistication, requiring significant supervision and feedback from the user.

The assistant can “drive” the process from the user’s research prompt/question, through search/gathering of relevant data, and subsequent extraction/organisation of the relevant information, resulting in a report output.

However, at this level it would be limited in scope along several dimensions like ability to handle volume, follow-on queries, nuanced understanding and parsing. It may miss and misrepresent a lot, and so the output would require scrutiny.

This would nonetheless provide a significant boost early on in the research process, charting some portion of the landscape in a fully automatic way, bringing the human user a diverse collection of threads to pull.

Level 4: Advanced End-to-End Automation

The system works autonomously with more sophisticated understanding and execution of tasks. It might still lack the nuance and depth of expert human researchers but works proficiently with minor supervision.

It could for instance be trusted to carry out a literature review on a well-defined topic. It can generate queries, search, download, summarise, restructure and integrate information.

Level 5: Expert End-to-End Automation

The assistant executes research autonomously at or above the level of an expert human, with full understanding, nuance, and integrity in handling information across various formats and sources.

The AI can conduct any research task without human intervention, even in complex and broad contexts, to deliver comprehensive and coherent results, and be able to explain its process fully.

Where We Are and Where We’re Going

Right now we are on the cusp of level 3, and semi-autonomous research assistants will start emerging across various industries.

The progression from level 3 through to 5 will naturally follow with the advancement of various trends:

• Multi-modality - Increasingly multi-modal AI models are being developed and made accessible, allowing agents to understand and generate media of all forms (text, images, audio, as well as interactive pages and software).
• Capacity - Systems will handle increasingly more information and context, enabling them to consider more factors and complexity in generating their output.
• Efficiency - Increasing speed and decreasing cost, enabling much higher throughput of information through AI systems (related to capacity above).
• Agency/autonomy - AI systems/agents being able to orchestrate and operate in multi-stage workflows, where data is explored and weaved until a conclusion (or other stopping criteria) are reached.

Some people are using current technologies like they are level 3 research assistants. ChatGPT is a perfect example of this. In reality, these technologies are level 2. This is similar to those self-driving cars where you’re supposed to have your hands on the steering wheel but people wedge an orange on the pressure sensor and read a book. In other words, a dangerous and misinformed use of the technology.

Outlook

We believe the components exist to reach level 4 today, but there remains substantial work to put the pieces together. We still have to instil the right training and guidance, as well as test and evaluate the systems to ensure they’re capable of what we want from them.

When it comes to information, quality and trust are king, and it’s of the utmost importance that we respect the limits of the technology we have while working towards the AI-assisted future we envision.