“We will conduct a cutting-edge, comprehensive literature review in machine learning and energy consumption. (…)”
Using this prompt, we tested the new Deep Research feature, integrated into the Openai O3 inference model since the end of February, and conducted a state-of-the-art literature review within six minutes.
This function goes beyond normal web search (for example, using ChatGPT 4o): the survey query is broken down, the Internet searches for information, then evaluated, and finally creates a structured, comprehensive report.
Let’s take a closer look at this.
Content table
1. What is Openai’s deep research? So what can you do?
2. How does deep research work?
3. How can you use deep research? – Practical examples
4. Issues and risks of deep research functions
Final thoughts
Where can I continue my studies?
1. What is Openai’s deep research? So what can you do?
If you have an Openai Plus account (a plan of $20 per month), you will have access to in-depth research. This allows you to access 10 queries per month. Using a PRO subscription ($200 per month) extended access to deep research and access to GPT-4.5 research previews with 120 queries per month.
Openai promises that it can perform multi-step research using data from the public web.
Duration: 5-30 minutes depending on complexity.
Previously, such studies usually took hours.
It targets complex tasks that require deep search and thoroughness.
What do specific use cases look like?
Conduct a literature review: Conduct a literature review on cutting-edge machine learning and energy consumption. Market Analysis: Based on current market trends and ratings, we will create a comparative report on the best marketing automation platforms for businesses in 2025. Technology and Software Development: Investigating programming languages and frameworks for AI application development with performance and use case analysis. Legal Investigation: We implement an overview of European data protection laws compared to the US, including related rulings and recent changes.
2. How does deep research work?
Deep Research uses a variety of deep learning methods to perform a systematic and in-depth analysis of information. The entire process can be divided into four main phases.
1. Decomposition and structuring of research questions
In the first step, the tool uses natural language processing (NLP) methods to process research questions. Identify the most important terms, concepts, and sub-questions.
This step ensures that AI understands the question not only literally but also from a content perspective.
2. Get related information
When the tool constructs a research question, it searches for information in a special way. Deep Research uses a mixture of internal databases, scientific publications, APIs and web scraping. These are open access databases, for example, such as ARXIV, PubMed, and Semantic Scholar, and are also public websites and news sites such as Guardian, New York Times, and BBC. Ultimately, all the content you can access online is published.
3. Data analysis and interpretation
The next step is for the AI model to summarise a large amount of text into a compact and understandable answer. Transformers and attention mechanisms ensure that the most important information is prioritized. This means that it’s not just about creating an overview of all the content found. Additionally, the quality and reliability of the source is evaluated. Cross-validation methods are typically used to identify incorrect or inconsistent information. Here, the AI tool compares several sources with each other. However, we don’t know exactly how this is done in deep research or what criteria this has.
4. Generate the final report
Finally, the final report is generated and displayed. This is done using Natural Language Generation (NLG) to display easy-to-read text.
The AI system generates diagrams or tables when requested at the prompt and adapts responses to the user’s style. The main sources used are also listed at the end of the report.
3. How to use deep research: practical examples
In the first step, we recommend using one of the standard models to ask how to optimize the prompts to carry out deep research. I did this at the following prompt on chatgpt 4o.
“We will optimize this prompt and carry out deep research.
Perform a literature search: Perform a cutting-edge literature search on machine learning and energy consumption. ”
The 4O model proposed the following prompt for deep research features:
We asked whether the tool could clarify the scope and focus of the literature review. So I provided some additional specifications:
After that, ChatGpt returned an explanation and began his research.
In the meantime, I was able to see the progress and how more sources were gradually added.
Six minutes later, a state-of-the-art literature review was completed and a report containing all the sources was available to me.
An example of deep research. MP4
4. Issues and risks of deep research functions
Let’s look at two definitions of research.
“A detailed study of the subject, especially to discover new information and reach new understandings.”
Reference: Cambridge Dictionary
“Research is a creative and systematic task that has been done to increase knowledge inventory, including the collection, organization, and analysis of evidence to enhance understanding of topics characterized by specific attention to controlling bias and sources of error.”
Reference: Wikipedia Research
The two definitions indicate that research is a detailed and systematic investigation of the topic. It aims to discover new information and achieve deeper understanding.
Essentially, deep research functions meet these definitions to some extent. Collect, analyze and present existing information in a structured way.
But I think we also need to be aware of some challenges and risks.
Surface Hazards: Deep research is designed to efficiently search, summarise and provide existing information in a primarily structured form (at least at the current stage). Absolutely perfect for overview research. But how about digging deeper? Actual scientific research goes beyond mere breeding and takes a critical look at sources. Science has also managed to generate new knowledge. Enhanced existing biases in research and publications: If there are significant consequences, the paper is likely to be published already. On the other hand, “non-significant” or inconsistent results are less likely to be published. This is known to us as a publishing bias. This trend is reinforced when AI tools begin to evaluate mostly frequently cited papers. It is rare or so widespread, but perhaps important findings are lost. A possible solution here is to implement a mechanism for weighted source evaluation that also takes into account uncited but related papers. If AI methods cite mostly frequently cited sources, they are less widely available, but they can lead to loss of important findings. Perhaps this effect applies to us humans too. Research Paper Quality: It is clear that a bachelor’s, master’s, or doctoral thesis cannot be based solely on AI-generated research, but the question I have is how universities and scientific institutions address this development. Students can get solid research reports at just one prompt. Perhaps the solution here must be to adapt the evaluation criteria to provide greater weight to detailed reflections and methodology.
Final thoughts
In addition to Openai, other companies and platforms have integrated similar capabilities (even before Openai). For example, Prperxity AI introduced deep research capabilities to conduct and analyze searches independently. Gemini by Google also integrates such deep research capabilities.
This function provides a very simple overview of the first research question. It is still unknown how reliable the results are. Currently (from March 2025), Openai itself writes as a limitation that its functionality is still in its early stages, sometimes hallucinating facts as answers, drawing false conclusions, and struggling to distinguish authoritative information from rumors. Furthermore, we are currently unable to accurately convey uncertainty.
However, we can assume that this functionality will be further expanded and will become a powerful tool for research. If you have a simple question, we recommend using the standard GPT-4O model (with or without search) to answer it immediately.
Where can I continue my studies?
Want more tips and tricks on technology, Python, data science, data engineering, machine learning, and AI? Then I receive regular summary of my most read articles in my subsack – curated, free of charge.
Click here to subscribe to my Sacakk!
