The question both pieces are really trying to answer is: what is PhD training actually for, in a world where AI can do increasingly more of what we currently train scientists to do? That question has a better answer than either of us has given it so far.

Prachee’s argument and what I agree with

Prachee’s core claim is that the current structure of PhD training is a workaround for human limitations, not an expression of what produces the best scientists. We read slowly, hold limited information in working memory, and overfit to fashionable problems because careers depend on doing work that funding agencies and hiring committees can easily recognize and reward. Most of the elaborate infrastructure of PhD training exists because of these constraints, not in spite of them. What I find most refreshing about Prachee’s argument is that it’s not saying that PhD students can use AI to do more of the same things faster. She’s saying use AI to free time for training that the current system simply doesn’t provide at all, which is a fundamentally more interesting provocation.

Her central observation is that the first two years are dominated by content acquisition, building a mental map of what’s known and unknown in a field, and this takes so long not because deep immersion is intrinsically valuable, but because humans are slow. Understanding a field and finding its edges is now, she argues, functionally an AI task. The time this frees should go toward what PhD programs currently neglect: strategic thinking. Not “what is unknown?” but “which unknowns, if resolved, would reshape the most understanding?” Most PhD students exercise this judgment only a handful of times across five years. Qualifying exams are supposed to test it but almost no student fails them. Advisors substitute their own strategic judgment for the student’s. Grant structures reward safe incrementalism. The result is graduate programs that reliably produce content experts but not experts in scientific judgment.

Her prescription is to compress content acquisition into AI-assisted sessions and make strategic problem selection the core of the curriculum, which will provide students more reps on the exercises that currently happen only a handful of times across a five-year doctorate.

The strategic thinking deficit is absolutely real. The observation that PhD programs train people to find unknowns but not to identify which unknowns are worth resolving is a sharp diagnosis of graduate education that I agree with.

My framework and where it pushes back

The conceptual framework that I propose, described in my recent preprint, “Build expertise first: why PhD training must sequence AI use after foundational skill development”, starts from a different place. Its central claim is that generative AI is categorically different from previous technologies in what it automates.

Previous technologies automated mechanical execution: calculators ran arithmetic, search engines retrieved information humans then had to evaluate, statistical software ran analyses workflows designed by the scientist. These tools freed researchers for higher-order cognitive work. Generative AI, on the other hand, automates cognitive processes themselves: formulating analytical approaches, synthesizing arguments, generating interpretations, and producing complete analyses. It doesn’t execute your plan; it can produce the plan. This matters because PhD training is not about rapid task completion. It’s about developing independent scientific thought through sustained cognitive struggle, and generative AI offers to bypass exactly that struggle.

This new capability creates what the framework calls the verification paradox, where trainees cannot meaningfully verify AI outputs because verification requires the very expertise they are still developing. One cannot spot errors one hasn’t learned to recognize (and, as AI becomes increasingly better, the errors go from obvious to subtle, deep-seated ones that are harder still to recognize). So, using AI before developing this expertise means training yourself to produce unverified work. The polished appearance of AI output masks this gap dangerously well.

This is where my proposed framework pushes back on Prachee’s: arriving at the knowledge frontier quickly via AI synthesis is not the same as understanding it. Mental model construction is not content delivery. The Alpha School routine may surely work for stable, established content like multiplication tables and geography. But PhD-level scientific knowledge is contested, methodologically dependent, full of results that don’t replicate, and debates that turn on subtle technical judgments. Building a mental model of that landscape requires slow reading not because humans are inefficient but because encountering confusion and resolving it is how the map gets built. An AI-generated summary of the landscape is not the same as having walked it.

To be fair, this critique applies to AI content absorption as it currently exists in the form of passive summarization of a field’s literature. A more serious counter is that purpose-built pedagogical AI could be designed to do something genuinely different: posing increasingly complex synthesis problems, probing gaps in the student’s reasoning, and deftly surfacing confirmations, extensions, reinterpretations, and contradictions in the literature. This training-oriented-AI could adapt to the student’s background and learning style, and play the role of an infinitely patient, perfectly personalized scientific mentor that no human advisor has the time or bandwidth to be. If that version of AI-assisted training can be built and validated, some of my pushback on Prachee’s content-acceleration premise would need revision. That’s an empirical question worth taking seriously.

There is also a less obvious problem ahead. Students who used AI throughout undergrad (and soon throughout high school) will arrive at PhD programs in the next few years. These students will feel more prepared while being less equipped to catch errors. These incoming cohorts may present a harder version of this problem, not an easier one.

Where I think my framework falls short is that it calls for sequencing as a developmental principle while only partially answering why foundational expertise matters beyond the training period itself. What I left out turns out to be important, and it’s where the argument needs to go. To be clear, the sequencing framework I propose is a response to AI as it currently exists and is currently used in training contexts. It is not a claim about what purpose-built pedagogical AI might one day achieve. That distinction matters, because the harder challenge runs deeper than any particular tool and it unsettles my framework and Prachee’s simultaneously.

The regress problem

Prachee’s prescription is essentially a strategic retreat: concede content absorption to AI, concentrate human training on strategic thinking AI cannot yet do. This is a reasonable response to where AI capabilities stood a year ago or even where it stands now. It is going to be less stable in the very near future because, Prachee herself points out, AI co-scientists are no longer speculative. These systems are explicitly designed to navigate scientific literature, synthesize across fields, generate hypotheses, and prioritize experimental directions. These tools are targeting exactly the strategic layer Prachee wants to protect for human training. The pivot she recommends leads to a position that is itself being automated.

This creates a regress, i.e., a chain of retreating justifications with no stable endpoint. If the argument for human expertise is always “train for what AI can’t do yet,” we are in a permanent retreat toward a shrinking frontier. The framework that I propose doesn’t fully escape this trap either. Arguing that trainees need foundational expertise to verify AI outputs is correct but incomplete. It doesn’t explain what the deeply trained scientist is for in a world where AI co-scientists can do the strategic work too.

What the PhD is actually for: accountability, direction, and transparency

The durable answer is not about task performance at any level of the cognitive stack. It is about something that remains distinctively human regardless of how capable AI becomes: the capacity to bear responsibility for scientific claims. This capacity has three pillars.

• Accountability. Scientific claims must be attributable to agents who can defend their reasoning, acknowledge errors, and correct the record. This is not a bureaucratic formality. It is the mechanism through which science self-corrects. A scientist who cannot evaluate the work that carries their name cannot be genuinely accountable for it. They can sign it, but they cannot stand behind it. Foundational expertise is the prerequisite for accountability that means something rather than accountability that is merely performed.

• Direction. Even a maximally capable AI co-scientist requires a human who can direct it toward questions worth asking, not just strategically, in the sense of which unknowns reduce the most uncertainty, but ethically, in the sense of whose uncertainty, about what, and for whose benefit. Scientific inquiry embeds values that AI cannot determine. These values, including which populations get studied, which diseases get prioritized, and which results get amplified, are not algorithmic outputs. Developing the judgment to set those directions requires the deep domain immersion that foundational training builds.

• Transparency. Science is distinguished from oracle-consulting by the obligation to show your reasoning, not just your conclusions. This norm is also a diagnostic test: a researcher who built foundational expertise can document AI-assisted work in substantive detail. They will have the depth and clarity to report what the AI contributed, what they verified, and what they rejected and why. A researcher who depended on AI during training cannot produce that accounting in good faith, because they don’t fully understand what was done or why. Transparency is not just a scientific value. It is the mechanism that exposes AI dependence for what it is.

Together, these three pillars reframe what the PhD is for. It is not training to outperform AI on any task. It is training to be the human in an AI-augmented scientific enterprise who can bear these responsibilities genuinely, and whose foundational expertise makes them a capable, ethical partner rather than a passive passenger.

This reframe also resolves the regress problem. The argument for foundational expertise does not rest on “AI can’t do this yet.” It rests on the fact that science requires human agents who can be accountable, directive, and transparent, the capacities built through the sustained foundational period, not acquired by any other sped-up means. That argument holds regardless of what AI can generate.

The system problem

These arguments are sound on their own terms, but, as Shaun Mahony pointed out in a discussion, the institutional incentive structure1 actively works against them.

Academic reward systems measure ‘outputs’: number of (high-profile) papers, grant dollars, citation counts. A PhD student who uses AI dependently from day one will produce more in the short term and appear more productive. Their advisor, also evaluated on lab output, has at least a partially aligned incentive to permit this. The expertise deficit that early and pervasive AI adoption produces will be largely invisible during training. It will surface later at job interviews, when novel problems arise, and when collaborators or evaluators discover errors a trained scientist would have caught. The pessimistic view is that programs that mistake productivity for training will become the ‘norm’ before institutions move to prevent it. That view is probably correct at the macro level.

But pessimism about the macro does not require passivity at the micro. PhD programs already have the most important tool available: milestones that test independent capability in formats currently hard to fake. At their best, admissions interviews, preliminary exams, comprehensive exams, committee meetings, and dissertation defenses probe whether a student can reason in real time, defend choices, and acknowledge limits. So, the most tractable near-term action is to treat these milestones as the serious capability filters they were designed to be, and design them explicitly to probe verification ability and independent reasoning. This action does not require institutional transformation, just taking seriously what these milestones are actually for.

What both pieces are really arguing for

Prachee and I agree on way more than our differences suggest. We both believe the PhD’s protected time is precious and currently misused. We both believe that AI fundamentally and irreversibly changes what that time should be for. We both resist the extremes of “AI everything from day one” and “no AI during training.” And we both believe that the field is making consequential decisions right now, by default if not by design.

Where we differ is on the mechanism. Prachee sees content absorption as the bottleneck to remove. I think removing it without understanding what it builds risks producing scientists who can articulate a field’s frontier without being able to navigate it, and that AI-native incoming students make this risk more acute, not less.

But the deeper argument that neither of us has made fully until now is that the case for foundational expertise does not rest on what AI cannot yet do. It rests on what science requires humans to provide: accountability for claims, direction of inquiry, and transparency of reasoning. These capacities are built, slowly and effortfully, through the kind of training both of us are trying to protect.

Foundational expertise is not a byproduct of PhD training. It is what makes a scientist capable of bearing genuine responsibility for their work. That capacity is worth protecting deliberately, before institutional defaults make the choice for us. Achieving this reframe requires rethinking not just what we teach but what the PhD certifies, how advisors function, and what program milestones are actually measuring. I will return to this subject in a follow-up post. In the meantime, Prachee’s “Free the PhD” and my preprint “Build expertise first: why PhD training must sequence AI use after foundational skill development” and its practical implementation guide are the places to go for the fuller arguments.

Hi everyone, welcome!

I am a scientist working in the areas of computational biology and biomedical data science. That means I obsess about improving the ways in which all researchers can use the computer (algorithms & tools) to gain nuanced insights into health and disease from massive amounts of data.

What to expect in this space

My plan is to use this space to share and exchange thoughts about two broad areas:

1. I’ll write about topics and developments in my scientific fields, covering biomedical/computational problems, machine learning (ML)/AI methods & tools, publicly-available data, and open code & software.

2. And, from the perspective of a group leader, mentor, and educator, I will also write my reflections on academia & the scientific enterprise, research education & training, open science, and science communication.

[The most important part is done; feel free to skip to the end.]

Some personal context

… and this is the most personal it’s going to get!)

First, I belong to the Jurassic Park generation. This movie, which I saw when I was 9 years old and in a small theatre in a small South Indian suburb called Nagamalai, completely blew my mind. It filled me with fascination for genetics and molecular biology, and made me hope for an opportunity to do something in these areas when I grew up. Simultaneously, I have always had a deep love for mathematical and quantitative thinking. So, looking back, I consider myself lucky to have been able to follow both these streams throughout my education (without having to let go of one or the other) and make their confluence integral to my scientific career.

Next, having started my foray into computational biology in 2004–2006, I had the opportunity to be there at the early days of major advances in functional genomics and the application of ML to analyze, integrate, & gain insights from large numbers of high-dimensional datasets. These areas, along with the academic research & training enterprise itself, have transformed over these past 18 years, and I have had the good fortune to continuously work in this exciting interdisciplinary field building ML/AI methods & tools to advance biological and biomedical research.

Finally, though it feels like I had a straightforward journey to where I am now — K-12 ⭢ undergrad ⭢ grad school ⭢ postdoc ⭢ faculty position — my experiences along the way have been rich (with many highs & lows) and gained in the backdrop of two very different social, educational, & professional cultures (India & US). These experiences have, obviously, had a strong influence on my thinking about the education system, academic research, diversity-equity-inclusion in STEM, training & mentoring, and the role of science in society.

The exchange

In addition to writing a few things that may be interesting/useful to you, I mean for this to be a space for exchanging ideas with you. Folks from any background and at any career stage are welcome to chime in. And, to live up to this welcome, we all need to agree to keep our discourse constructive and respectful. From my end, I love to learn new ideas & perspectives, and am happy to change my mind about anything given compelling facts. So, keep your feedback coming. Cheers!

Leave a comment

Subscribe now