Summary: OpenAI’s new Deep Research offers rapid data synthesis and a novel approach to user interaction with shared dialogue initiative. Requesting clarifications upfront reduces revisions and shapes customized outputs. Design teams investing in “Pro”-level AI will likely see significant boosts in knowledge-driven productivity.
I went for the $200/month OpenAI Pro subscription that includes the Deep Research product powered by the o3 model that’s otherwise not available outside OpenAI. (Not to be confused with Google’s product that’s also named “Deep Research,” but is currently not as powerful.) Other AI influencers have characterized it as being as good as a skilled researcher with a Ph.D., but much faster at conducting literature reviews.
The new “Deep Research” reasoning models from OpenAI and Google dive deep into published sources and synthesize the information into detailed and readable reports. (Leonardo)
Since I’m less interested in traditional science, I posed it the challenge of planning a short vacation in Hokkaido, Japan. Deep Research provided great overviews of the resorts and sightseeing options, but I was particularly impressed with its ability to ask follow-up questions before embarking on the research. For example, before selecting resorts, it asked how much value to place on the view from the room itself, as opposed to simply being in a location close to the sights. I do have a strong preference for views but had forgotten to specify this in the prompt. Similarly, when I asked for a sightseeing itinerary, Deep Research teased out my hiking preferences in terms of strenuousness, and also asked how tightly packed I wanted the days. (I used to be a hardcore tourist, but these days I take it more easy.)
Sample Research Task: AI Stigma
I did ask Deep Research to conduct an actual research project, with the following long prompt:
“I want an article about research investigating the extent to which humans prefer to interact with an AI instead of other people or when humans dislike information or interactions more when they know that they are dealing with an AI instead of a person. Include studies where research participants were deceptively told that they were dealing with an AI or with a human even when the opposite was actually the truth, to the extent that these studies reveal whether the study participants rate information or interactions differently depending on whether they think they are dealing with an AI or a human. Also include studies where humans were asked to perform tasks with or without the assistance of AI and where the AI also performed the task alone without humans. In cases where task performance can be measured, when were humans and AI together better than either alone, versus when the performance of AI on its own was better than when coupled with a human which would indicate that the human did not make full use of the AI's work. Also include qualitative research that investigates humans' attitudes toward working with AI and the extent to which those attitudes can lead users to perform poorly when they do not make full use of the AI's work. Conclude by estimating what criteria or circumstances lead users to work productivity with AI versus situations where distrust or negative attitudes lead users to underperform when they could have been helped more by AI.”
Here are Deep Research’s follow-up questions:
“That sounds like a fascinating topic! Before I begin the research, could you clarify a few details:
Are you looking for studies across all domains (e.g., workplace tasks, education, customer service, healthcare, etc.), or do you have specific fields of interest?
Do you prefer peer-reviewed academic studies, industry reports, or a mix of both?
Should I focus on recent studies (e.g., from the last five years), or are older studies also relevant?
Would you like a structured report with sections such as study summaries, key findings, and conclusions?”
After I clarified my request, Deep Research thought for 8 minutes and wrote a nice overview report of AI Stigma based on 22 sources. (I did edit it a bit and added some more content before publishing this, but the report was almost publication-ready.)
Follow-Up Questions = AI-Driven Conversation
AI-driven follow-up questions are an innovation in AI-UX and improve results substantially. Having the AI take some initiative in the dialogue (as opposed to being a passive respondent) enhances human–AI symbiosis by prodding the human to think about issues he or she had overlooked. Given the slow response time of current reasoning models (e.g., the 8 minutes for my report), it’s also best for usability to get the clarifying questions out of the way immediately, rather than having the user wait for the initial response and then have to issue a new prompt to modify the output.
8 minutes may not sound like a lot, but as a computer response time, it’s enough to make the user snooze off while the computer is working. Better to frontload a clarification dialog before setting the AI loose on a research project while reasoning models remain slow. (Ideogram)
Invest in AI for Design Teams
Deep Research is useful for many design team tasks, especially ones that involve assessing best practices or summarizing the available research. If you run a design team, go ahead and pay that $200 subscription fee for at least a few of your senior staff — this is cheaper than hiring a UX consultant and will be used more, because people will request many more overview reports when they are delivered in a few minutes. In this way, Deep Research becomes a great example of the third of my four metaphors for working with AI: treat AI as a teacher.
(Better: cover “Pro” subscriptions for everybody on your team, because they need to upskill with knowledge of the best AI, not yesterday’s AI. Design team leaders must invest in the best AI tools for their team to avoid having design fall even further behind their engineering colleagues who are sure to spend big on AI.)
Spending $200/month for each member of your design team on an AI “Pro” subscription sounds expensive, but Deep Research can save much more money than it costs. Plus you upskill your team to be ready for the even more powerful AI they’ll have next year. (Leonardo)
One-Minute Video
I made a short one-minute video about my initial experience with Deep Research (YouTube).
People often ask me why I make videos in so many different media formats. One answer is that it’s fun and that it’s useful to experiment with the many new AI models that constantly launch. For example, I used Qwen 2.5-Max (a new model from Alibaba) to make a B-roll clip for this video, even though I mainly use Kling for this purpose.
I strongly encourage you to embrace a similar approach. Not necessarily to make AI videos, but to experiment with new AI tools as they are released. Progress in AI is so fast that experience with last month’s AI will mark you as a dinosaur.
If you don’t regularly get personal hands-on experience with new AI tools, you’ll become an AI dino. (Midjourney)
The second, and more user-centered, answer is that different people prefer different media forms and also that different audiences can be reached on different platforms. For sure, a one-minute video appeals to many people who would not slog through a written article.
Different media forms have advantages and appeal to different audiences. AI excels at converting content between media. (Leonardo)