Linjun Zhang concludes with the third in our invited series of articles on LLMs and AI (read part 1 and part 2), and their implications for the statistics profession. He turns to the practical consideration of how we can see AI as a collaborator, instead of as either a threat or a cure-all.
In my previous two articles, I explored the role of Large Language Models (LLMs) as a “new companion” for statisticians in education and outlined the “axioms” for our foundational role in research. To conclude this series, I want to focus on a more practical question: What does it mean to live with AI in our daily academic routines?
The metaphor that best captures our current moment is the “Warm Start.” In optimization, a warm start provides an initial guess close to the solution, drastically reducing the time to convergence. However, any researcher knows that a warm start guarantees nothing. Without a well-defined objective and carefully chosen updates, an algorithm may stall at a local optimum, or wander aimlessly.
LLMs now provide such a warm start for our field. New research problems at the intersection of statistics and AI are emerging; AI is accelerating our writing and coding and helping us analyze unstructured data in healthcare. They also reshape incentives by attracting funding, students, and institutional attention toward AI-related work. But warm starts require careful navigation. Momentum alone does not ensure meaningful progress, and we must be more cautious about how we proceed.
The Classroom Experiment: The Decay in the Middle-Game
Last semester, I conducted a small experiment in my graduate course on the statistical foundations of LLMs at Rutgers. I explicitly encouraged students to use LLMs only to write their final research papers and asked them to submit detailed reflections on their co-authorship process with AI.
Students quickly discovered that LLMs are exceptionally effective at getting started. They generated ideas, suggested problem formulations, proposed related literature, drafted outlines and plausible introductions within minutes, and presented the paper in a fun video. For many, the most difficult step in academic writing, the blank page, simply disappeared. The model provided what we might call a warm start: a structured initial direction that lowered the barrier to thinking and writing.
Yet the middle-game told a different story. As drafts evolved, students reported that models began to hallucinate references, introduce subtle logical gaps, and recycle earlier arguments in increasingly repetitive language. Rather than pushing ideas forward, the model often circled back to previously stated points. The resulting drafts appeared polished but stagnant: coherent on the surface, yet lacking genuine intellectual progression. Real progress required human intervention: decomposing questions into smaller parts, verifying each logical step, and exercising independent judgment.
The lesson was clear. Current LLMs excel at initiation but remain unreliable engines of sustained reasoning. For now, they only provide momentum, not direction.
A Warm Start for the Statistical Community
This classroom observation mirrors our broader professional landscape. The AI boom has given the statistical community its own institutional warm start.
We now see rapid advances in AI-assisted theorem proving and mathematical discovery. AI systems are beginning to solve research-level mathematical problems and assist in formal proof generation, though human verification remains essential. At the same time, “AI for science” initiatives and AI co-scientist frameworks are reshaping how research is conducted across disciplines. And in everyday academic life, generative tools dramatically improve productivity in coding, writing, exploratory analysis, and presentation.
This is our momentum. But warm starts can be deceptive. They create acceleration without guaranteeing direction. We are already witnessing a surge of low-effort AI-generated content across conferences, journals, and grant proposals. When productivity becomes frictionless, volume grows faster than depth. If we rely uncritically on the linguistic fluency of generative models, we risk drifting toward derivative research—work that appears sophisticated yet rarely moves beyond established ideas.
The appropriate response is neither rejection nor blind adoption of current AI. Instead, we should advocate for deliberate co-evolution between statisticians and AI systems: a partnership in which each shapes the development and responsible use of the other.
Dimension 1: Co-Evolution in Theoretical Research
AI systems are increasingly capable of assisting in mathematical reasoning. They can propose conjectures, sketch proof strategies, generate counterexamples, and search through combinatorial spaces of possibilities. In exploratory phases of theoretical research, such capabilities can be transformative.
Yet the warm-start dynamic persists. AI can suggest directions, but it does not yet reliably identify which paths are structurally meaningful. According to my own experience, human efforts — such as decomposing complex statements, identifying appropriate abstractions, and simplifying problems to their essential components — dramatically improve the quality of AI-generated proofs. Effective collaboration therefore requires structured interaction. Researchers must guide models with carefully designed prompts, iteratively refine questions, and critically evaluate outputs. When treated not as an oracle but as a heuristic generator, AI becomes a powerful exploratory partner in theoretical discovery.
Dimension 2: Co-Evolution in Applied Research–The AI Co-Scientist
In applied settings, the co-evolution between AI and statistics becomes even more visible. The idea of an “AI co-scientist” is no longer speculative. LLMs and AI agents can assist with hypothesis generation, literature review, code development, experimental design suggestions, and preliminary data analysis. Multi-agent systems can coordinate tasks across complex research pipelines.
Yet successful application requires careful human scaffolding. Benchmarks must be curated. Evaluation pipelines must be rigorously designed. Domain knowledge must guide interpretation. Without such structure, AI-generated analyses risk reinforcing spurious correlations or optimizing for superficial metrics.
Statisticians are uniquely positioned to provide this scaffolding. Our discipline emphasizes experimental design, uncertainty quantification, and causal inference. In areas such as biostatistics and public health, where decisions affect patient outcomes and policy, these principles are indispensable. The statistician’s role evolves from sole analyst to architect of reliable AI-assisted scientific workflows.
Dimension 3: Co-Evolution in Academic Practice
The influence of AI extends beyond research into the infrastructure of academia itself. Peer review, authorship, and scholarly communication are already being reshaped. Authors are increasingly utilizing LLMs for drafting and editing, while reviewers experiment with AI-assisted summaries and critiques. Simultaneously, editors confront urgent questions regarding disclosure, originality, and the delegation of responsibility.
Moving forward, transparency must be the cornerstone of our practice. The reproducibility of AI-assisted analyses and explicit human accountability for every claim must become standard requirements. Yet, this evolution is not purely defensive; properly integrated, AI can actually strengthen scholarly evaluation. These tools excel at detecting statistical errors, identifying incomplete reporting, or spotting internal inconsistencies across large manuscripts that might elude a tired human eye.
Statisticians, again, are uniquely positioned to lead this institutional shift. Our expertise in evaluation, rigorous testing, and the quantification of uncertainty equips us to distinguish genuine scientific insight from the “superficial fluency” of a generative model. By designing frameworks for responsible, AI-assisted scholarship, we ensure that while the tools of our trade change, the integrity of the record remains uncompromised.
Beyond the Warm Start
Warm starts reduce friction and open new directions, but they also create the illusion that progress is inevitable. A favorable initialization improves convergence only when guided by well-defined objectives and disciplined updates. Otherwise, the process may stall or drift toward suboptimal outcomes.
The AI era offers statistics a powerful initialization: unprecedented attention, resources, and opportunity. Realizing this potential requires deliberate navigation. We must resist viewing AI as either a threat or a cure-all and instead treat it as a collaborator whose strengths and limitations must be continuously calibrated.
LLMs help us begin. They do not complete the journey.
AI gives us a warm start. What follows depends on how we play the middle-game—carefully, critically, and collaboratively—toward a future in which statistical thinking and artificial intelligence evolve together.