OpenAI’s ‘deep research‘ is an innovative artificial intelligence (AI) instrument garnering considerable attention for its ability to accomplish tasks in mere minutes that would typically require extensive hours of human expert analysis.

Integrated within ChatGPT Pro and advertised as a research assistant on par with skilled analysts, this tool autonomously navigates the web, aggregates information, and produces structured reports. It even achieved a remarkable score of 26.6 percent on Humanity’s Last Exam (HLE), a demanding benchmark for AI, surpassing many competing models.

However, the reality of deep research may not fully align with its marketing hype. While the system generates high-quality reports, it is not without significant shortcomings. Journalists who have engaged with the tool note its propensity to overlook critical details, exhibit difficulty with recent developments, and even fabricate information.

OpenAI acknowledges these limitations, indicating that the tool “can occasionally hallucinate facts or draw erroneous conclusions, albeit at a reduced rate compared to previous ChatGPT iterations, based on internal assessments.”

Unsurprisingly, inaccuracies can occasionally infiltrate its outputs, as AI systems do not “know” information in the same intricate manner as humans.

What is ‘deep research’ and who can benefit from it?

Targeted at professionals in finance, scientific fields, policy-making, legal sectors, and engineering, as well as academics and journalists, deep research is presented as the latest “agentic experience” offered through ChatGPT. It asserts the capability to streamline the burdensome aspects of research into mere minutes.

Currently, this functionality is exclusively available to ChatGPT Pro users within the United States at a monthly fee of $200. OpenAI plans to extend access to Plus, Team, and Enterprise tiers in the following months, with more economical options anticipated.

Distinct from standard chatbots that yield quick replies, deep research employs a systematic approach to generate comprehensive reports:

1. The user projects a query, which might range from market assessments to legal summaries.
2. The AI seeks clarification, posing additional questions to fine-tune the research parameters.
3. The agent independently scours the web, navigating numerous sources such as news articles, academic papers, and online databases.
4. It consolidates its findings, extracting salient points, organizing them into a structured report, and appropriately citing its references.
5. Within a span of five to thirty minutes, the user receives a detailed document summarizing the research outcomes, potentially equivalent to a PhD-level thesis.

At face value, this tool appears to be a boon for knowledge workers; however, a closer examination uncovers notable deficiencies.

Numerous initial evaluations have revealed its limitations:

• Lack of contextual awareness, as the AI can summarize but fails to grasp critical nuances.
• Neglecting recent developments, leading to the omission of significant legal or scientific advancements.
• Generating fabricated information, akin to other AI models.
• Inability to discern fact from fiction, failing to differentiate between authoritative and unreliable sources.

Despite OpenAI’s assertions that this tool rivals human analysis, it invariably lacks the discernment, rigorous evaluation, and specialized knowledge that elevate quality research.

What AI cannot supplant

ChatGPT is not the only AI capable of rapidly traversing the web and producing reports from minimal input. Notably, just 24 hours post-launch, Hugging Face released a free, open-source alternative that nearly parallels its effectiveness.

The principal risk associated with deep research and similar AI tools is the misconception that machines can adequately replace human cognition. While AI can effectively summarize information, it lacks the capacity to question its own premises, identify knowledge gaps, think innovatively, or appreciate diverse viewpoints.

Moreover, AI-generated summaries lack the depth and insight of a competent human researcher.

Any AI agent, no matter how expedient, remains a tool rather than a substitute for human intellect. It is paramount for knowledge workers to enhance skills that AI cannot emulate, such as critical thinking, evidence verification, in-depth specialization, and creativity.

For those inclined to utilize AI in research, responsible usage is crucial. Thoughtful integration of AI can elevate research efforts without compromising precision or depth. For instance, AI may be employed for efficiency—such as synthesizing documents—while human judgment remains the cornerstone of decision-making.

It is essential to validate sources, as AI-generated citations can be deceptive. Exercise critical thinking and verify information against reputable sources, particularly on high-stakes issues such as health, justice, and democracy; in such cases, complement AI findings with insights from human experts.

Despite extensive marketing that suggests otherwise, the limitations of generative AI persist. Individuals capable of creatively synthesizing information, questioning assumptions, and exercising critical thinking will continue to be indispensable—AI has yet to substitute their irreplaceable abilities.

Raffaele F Ciriello, Senior Lecturer in Business Information Systems, University of Sydney

This article is republished from The Conversation under a Creative Commons license. Read the original article.