1.3 Million Scientists Ask ChatGPT 8 Million Questions a Week: 6 Strategies for Researchers to Stay in Control

If a researcher consults AI before a mentor, colleague, a paper, or even before their own notes, what does that make AI?

Recent estimates indicate that 1.3 million scientists generate over 8 million weekly queries on ChatGPT, many tied to core scientific reasoning and technical problem-solving. The scale is striking! But what is more notable is that AI is no longer peripheral to research workflows. It is becoming a part of the moment where knowledge is grasped, refined, and formed.

That shift–from assistance to active participation, is what turns a tool into infrastructure.

The Rise of Invisible Infrastructure

Infrastructure is most powerful when it is silently embedded. Electricity, cloud computing, and the internet have all followed this trajectory, moving from visible systems to invisible backbones. Historically, research tools like statistical software, reference managers, or search engines were clearly bounded in function and supported discrete steps in the research lifecycle. AI systems, particularly large language models (LLMs), are fundamentally different, as they can perform a range of functions.

This breadth has transformed AI into what many researchers now treat as a collaborator. The Researcher of the Future report by Elsevier indicates that a significant proportion of researchers are already using AI tools across tasks like literature review (51%), writing (38%), and data analysis (38%). This indicates how strongly (but quietly) AI is being embedded into routine tasks and processes.

Reporting from MIT Technology Review highlights how newer AI systems are collapsing multiple stages of the research process into a single interface. OpenAI’s newly introduced research workspace, designed to integrate conversational AI directly into writing, coding, and collaboration environments is a strong example.

What emerges here is not just a more efficient workflow, but one where AI is continuously present for shaping decisions at multiple stages, rather than assisting at isolated points. This model, described as enabling scientists to “vibe code” their work, seems to be a transition toward AI systems that are not just assisting tasks, but orchestrating workflows, reducing friction between ideation and execution.

This raises a critical question: When AI becomes always available, seamlessly integrated, and cognitively interactive; does it cease to feel like a tool and begin to function as an infrastructure for thought.
This invisibility, however, is precisely what introduces and/or amplifies the associated risks.

Hidden Risks of Invisible AI Dependence
 

1. Unseen Knowledge Mediation 

As a researcher, you may find yourself relying on AI-generated interpretations without fully interrogating underlying sources. For instance, researchers frequently publish papers with erroneous citations without checking, as seen in analyses of 176 citations where 20% were wholly invented. Furthermore, a separate study identified a “verification gap” as 41.5% of researchers copy LLM-generated citations without checking and 76.7% of reviewers do not thoroughly verify references. 

Another particularly telling case comes from the NeurIPS 2025 conference, where over 100 fabricated citations were identified across accepted papers, despite expert peer review. The fact that such errors passed through one rigorous review systems makes it deeply concerning.

This indicates when AI mediates knowledge early in the process, errors are not corrected but inherited. Over time, this creates the risk of citation laundering, where fabricated or distorted references enter the scholarly record and are unknowingly reused.

2. Silent Propagation of Bias and Hallucination

Bias and hallucination in AI systems are well documented. What changes in an infrastructure context is not their existence, but their scale, persistence, and invisibility. 

A study examining the validity of ChatGPT in identifying literature found that 68% of the links in the references were incorrect and 31% were fabricated. Even advanced models continue to fabricate references, particularly in specialized domains where training data is sparse. Furthermore, LLM accuracy is highly reliant on the how frequently the papers are cited; which implies that hallucinations increase when papers are less cited and domains are niche. 

Notably, these errors are not always random. They follow patterns like combining real author names with incorrect titles, fabricating DOIs that appear structurally valid, and blending multiple sources into a single “plausible” reference. This makes them exceptionally difficult to detect, especially at scale.

When AI errors are embedded into workflows, they do not remain isolated, they propagate across documents, disciplines, and decision systems.

3. Shifting Authority from Human Expertise to AI Reasoning

As AI becomes more integrated into research workflows, it begins to influence how you validate, prioritize, and trust information.

This shift manifests in subtle but significant ways, especially when you: 

• use AI outputs as starting points for reasoning
• treat AI-generated summaries as proxies for literature reviews
• seek confirmation from AI rather than from primary sources or peers

Over time, this can shift how you evaluate and trust information; subtly moving authority from primary sources toward AI-generated reasoning. Traditionally, authority in scholarly publishing was grounded in peer-reviewed literature, domain expertise, and methodological rigor. Perhaps in the future, it may increasingly be influenced by AI-generated reasoning pathways, which are opaque, non-verifiable, and not always reproducible. 

Why These Risks Matter More as AI Becomes “Invisible”

Individually, none of these risks are new. What is new is how they interact under conditions of scale and integration, particularly when: 

• Errors are introduced earlier in the workflow
• Verification is inconsistent or absent
• Outputs are highly persuasive
• Systems are used repeatedly and routinely

This creates a form of invisible dependency, and that is the core challenge. The more seamlessly AI integrates into research, the harder it becomes to distinguish between what we know, what we infer, and what the model has suggested.

Why Researchers Still Continue to Rely on AI

Despite the risks, AI adoption in research continues to accelerate for understandable reasons. At a practical level, AI addresses several long-standing pressures in academia like growing literature volume, interdisciplinary complexity, and tight timelines. AI/LLMs offer immediate and tangible outcomes. This prompts them to rely on AI for learning unfamiliar concepts/domains, creating starting points for thinking, and reducing the friction of early-stage ideation. The result is a fundamental trade-off between efficiency and epistemic control.

The challenge is not whether you adopt AI or not, it is how well you can calibrate its role. On one side, it supports faster workflows, broader information access, and lesser burden. On the other end, it blurs the reasoning pathways, passes on undetected errors, and erodes critical evaluation.

Practical Considerations for Researchers

If AI is increasingly functioning as an embedded layer within research workflows, the question is no longer whether it should be used, but how its role can be calibrated without compromising epistemic rigor. This requires a shift from passive adoption to deliberate use.

1. Maintain rigorous verification practices
AI-generated citations, summaries, and interpretations should be treated as provisional. Independent validation against primary sources remains essential, particularly because many errors are plausible and therefore difficult to detect.

2. Be intentional about when AI enters the workflow
The stage at which AI is introduced matters. Early reliance can shape problem framing, source selection, and reasoning pathways. Engaging with primary materials before consulting AI helps preserve intellectual independence.

3. Anchor analysis in primary literature
AI outputs should not replace direct engagement with full texts. While summaries accelerate access, they often compress nuance, omit methodological detail, and obscure uncertainty. Sustained interaction with original sources remains critical.

4. Treat AI as a support, not a substitute for reasoning
AI can assist with structuring ideas or exploring alternatives, but it should not become the primary driver of interpretation or judgment. Maintaining ownership of reasoning processes is essential to preserving scholarly rigor.

5. Periodically audit dependence on AI tools
As AI becomes habitual, its influence becomes less visible. Reflecting on when and why it is used can help distinguish between productive augmentation and unexamined reliance.

6. Ensure transparency and disclosure of AI use
Clearly documenting where and how AI has contributed supports reproducibility and accountability. As AI becomes more integrated, structured disclosure practices will be increasingly important in making its role visible.

Taken together, these strategies do not argue against the use of AI in research. Rather, they emphasize that as AI becomes more deeply embedded, maintaining rigor will depend less on limiting its use and more on making its role visible, bounded, and critically engaged.