Yes, LLMs are useful for some of those activities. But I strongly disagree about writing emails. DON'T use LLMs for that, unless you want to torture the people you're sending it to. The emails will be longer than necessary and make more mistakes. They will also sound pompous.
Working on a meta-analysis currently, can confirm that LLMs are helpful but certainly can't do it themselves just yet, or at least not without some damning errors. Agree that AI will be able to do many of these things soon, but most publicly accessible models just aren't there yet. The overestimation of AI's current abilities embodied in this post (satirical or not) might help explain why some recent federal health documents have ended up with hallucinated references.
#6 - I don't see any particular logic to this assertion. How would AI comment about someone more knowingly than I would? References, personal letters are not all about gleanable data, at least in the arenas of medicine I have worked in. And the arenas of life.
You’re probably right. Mostly. Yet there is still room for inspired writing. Finding the right word. Using literary devices such as alliteration. Writing is basic. Literature elevates the conversation
Academia has in large part has been the human versions of AI, just disparate and with nuanced complexity.
Only the nuanced complexity remains its valid domain.
And for many purposes, questions in those domains are unlikely to be answered by clinical trials, because the questions pertain to individuals and not populations.
@Vinay Prasad MD MPH I honestly can’t tell where the satire ends and where your real view begins.
That makes the piece provocative, but also a little unsettling. If the reader can’t tell which parts you actually believe, does the satire start to blur the very thing that matters most in writing: your own voice?
I’d genuinely be curious how much of this reflects how you actually use AI day to day.
And maybe the hardest question is this: at what point does provocation stop sharpening your voice and start hiding it?
I took your advice and ran your article through co-pilot and asked it to critique. Even it disagrees with parts of your statement: My short take: Prasad is directionally right about workflow substitution, but dramatically overstates epistemic substitution. He's good at identifying tasks whose production value is collapsing; he's much less persuasive when he suggests AI can replace the intellectual process behind them.
Where I think he's mostly right
1. First drafts are dead.
Grant sections, letters, emails, CV updates, meeting summaries, reviewer responses, teaching materials, and many administrative documents can often be drafted by AI faster than humans can produce a decent first version.
I increasingly see academic writing divided into:
defining the question,
vetting the evidence,
refining the argument,
rather than producing prose.
The sentence-level writing itself is becoming the least valuable part.
2. Documentation is being transformed.
I agree strongly regarding note-taking and ambient documentation. If AI can reliably produce documentation while clinicians maintain eye contact and listen, that's genuinely valuable.
The real challenge is not generation. It is:
accuracy,
medicolegal accountability,
integration into workflow,
preservation of clinical reasoning.
3. Routine descriptive analyses are becoming commoditized.
This is probably the most provocative point, and there's some truth to it.
A paper that merely reports:
"Here are all FDA oncology approvals from 2011-2025"
without a compelling conceptual question is going to have less value.
AI lowers the cost of:
extracting data,
generating figures,
performing standard analyses,
summarizing trends.
The "minimum publishable descriptive unit" is becoming less defensible.
Where I think he is wrong
1. Review articles are not primarily information products.
This is the biggest flaw in the post.
A good review is not valuable because it aggregates papers.
It's valuable because it:
frames a field,
resolves tensions,
identifies conceptual gaps,
advances theory,
provides interpretation.
AI can already generate the equivalent of a mediocre review article.
What it cannot reliably do is produce the equivalent of:
a landmark conceptual review,
a theoretical synthesis,
a paradigm-shifting interpretation.
The value was never the references. The value was the thinking.
In fact, as AI floods the world with summaries, genuine synthesis may become more valuable.
2. Meta-analyses are not just statistical aggregation.
If you've ever done one, you know that the difficult work is:
deciding inclusion criteria,
identifying hidden biases,
assessing methodological heterogeneity,
interpreting contradictory evidence.
AI may automate much of the mechanics.
But the judgments remain.
A fully automated meta-analysis can produce precise nonsense if the underlying studies are conceptually incomparable.
3. Medical guidelines are not purely evidence exercises.
This is where I disagree most strongly.
Guidelines involve:
value judgments,
risk tolerance,
stakeholder negotiation,
resource allocation,
ethics.
The recommendation is not simply:
"What does the evidence show?"
It is:
"Given imperfect evidence, what should society recommend?"
Those are fundamentally normative decisions.
AI can support guideline development.
I don't think it can legitimately replace accountable human governance.
As a medical education researcher
This post highlights something many academics are reluctant to admit:
For decades we've conflated:
the production of scholarly artifacts;
the generation of scholarly insight.
AI mostly attacks #1.
Many academic reward systems were built around visible products:
reviews,
narrative summaries,
descriptive studies,
conference abstracts,
committee documents.
If AI can generate those cheaply, then our evaluation systems become unstable.
The important question becomes:
What intellectual contributions remain difficult to automate?
My suspicion:
asking important questions,
building conceptual frameworks,
interpreting ambiguity,
understanding social context,
exercising judgment,
establishing trust,
remain comparatively resistant.
Interestingly, these are precisely the things medical education has often struggled to measure.
The irony
Prasad compares resistance to AI with resistance to the printing press.
I think the better analogy is calculators.
Calculators eliminated huge amounts of arithmetic labor.
But they did not eliminate mathematics.
In fact, they increased the premium on conceptual understanding because computation became cheap.
Similarly, AI may eliminate much of the labor involved in:
writing,
summarizing,
formatting,
extraction,
basic synthesis.
That doesn't mean it replaces scholarship.
It means scholarship shifts upward.
The key error in the post is treating information processing and intellectual contribution as the same thing. For much of academia they have been entangled, which is why AI feels so disruptive. But they are not identical, and the distinction becomes clearer every month.
For those that had an army of pre-med, residents, and others writing papers for them, they now have AI is a great point. Similar with the summary overview of reading…I mean does anyone really want to read the things you should stop doing already because of AI? Honestly, I agree with the points in your post, what I fear is that those of us tat have gone deep, and wide…can use the technology to expedite and scale ourselves. There is an entirely new generation that does have the width of knowledge and doesn’t care about the depth. If technology is neither good or bad, but it is how people use it, then that becomes the underlying challenge that we as the elder statesmen need to remember as we lead.
Anyone who believes AI is able to replace humans at current is a fool. MAYBE one day they will get them advanced enough so they can do jobs done by humans but they are NOT there today even though that's how they are marketing AI to be now!
AI is quiet dumb. It is a little bit better than a generic google search but it is FAR from being able to replace humans. One need only chit chat with one of these AI bots for a little bit and you'll see them contradicting themselves and giving very much false information which when questioned about the AI may acknowledge or it may deny and argue with you instead.
We do not know how these tins work b/c they are grown and not built/created. These AI's are the digital equivalent of a gardener raising some vegetation. The gardener is NOT a botanist and thus doesn't know the inner workings if the plant only how to grow it by feeding and nurturing t, They do this with these AI bots! They have zero clue as to how these AI's work,
Therapist here, most of my colleagues are terrified of AI - What if a tool built around alllll the input you provide it with 24/7 mostly free availability better than a therapist? Maybe? So be better? Adapt? The clients I have using AI aren't using it instead of therapy - they aren't canceling sessions because of their bot. They Are bringing in their thoughts in more organized, self examined ways, and for some, it has felt transformative. My role is shifting slightly. Deep interpersonal work is happening more quickly. But it feels in tandem with our work in sessions, not instead of. And for my teammates who refuse to use a tool that seems inevitable to human society, they may need to put down that iPhone, something also changed the face of technology.
You are the "boss" of your AI. You can give your AI parameters, things to look at, things to ignore, things to analyze more deeply. You can push your AI in the direction you want to. Don't ever think you are supposed to simply accept the first answer AI gives you. Debate with it. You will be amazed at how quickly it catches on and comes over to your side with deeper analyses, using data that it initially ignored. Good luck! Have fun!
If you like slop...with a lot of inaccuracies...
Yes, LLMs are useful for some of those activities. But I strongly disagree about writing emails. DON'T use LLMs for that, unless you want to torture the people you're sending it to. The emails will be longer than necessary and make more mistakes. They will also sound pompous.
Working on a meta-analysis currently, can confirm that LLMs are helpful but certainly can't do it themselves just yet, or at least not without some damning errors. Agree that AI will be able to do many of these things soon, but most publicly accessible models just aren't there yet. The overestimation of AI's current abilities embodied in this post (satirical or not) might help explain why some recent federal health documents have ended up with hallucinated references.
#6 - I don't see any particular logic to this assertion. How would AI comment about someone more knowingly than I would? References, personal letters are not all about gleanable data, at least in the arenas of medicine I have worked in. And the arenas of life.
I disagree. AI can do a lot of things but it makes people lazy. Realistically, AI makes mistakes all the time and overlooks other values.
LOL, Did AI write this post? I am guessing...yes.
You’re probably right. Mostly. Yet there is still room for inspired writing. Finding the right word. Using literary devices such as alliteration. Writing is basic. Literature elevates the conversation
Academia has in large part has been the human versions of AI, just disparate and with nuanced complexity.
Only the nuanced complexity remains its valid domain.
And for many purposes, questions in those domains are unlikely to be answered by clinical trials, because the questions pertain to individuals and not populations.
@Vinay Prasad MD MPH I honestly can’t tell where the satire ends and where your real view begins.
That makes the piece provocative, but also a little unsettling. If the reader can’t tell which parts you actually believe, does the satire start to blur the very thing that matters most in writing: your own voice?
I’d genuinely be curious how much of this reflects how you actually use AI day to day.
And maybe the hardest question is this: at what point does provocation stop sharpening your voice and start hiding it?
Swift meant everything he said (even if said in metaphor) AND it was satire, AND it was funny. But also always true.
I think Vinay is doing the same thing.
I suspect I also detect some sadness that it is so. Also, he feels at a bit of a loss himself.
But more, he is frustrated with colleagues who are putting their heads in the sand and pretending AI is not here, now.
I took your advice and ran your article through co-pilot and asked it to critique. Even it disagrees with parts of your statement: My short take: Prasad is directionally right about workflow substitution, but dramatically overstates epistemic substitution. He's good at identifying tasks whose production value is collapsing; he's much less persuasive when he suggests AI can replace the intellectual process behind them.
Where I think he's mostly right
1. First drafts are dead.
Grant sections, letters, emails, CV updates, meeting summaries, reviewer responses, teaching materials, and many administrative documents can often be drafted by AI faster than humans can produce a decent first version.
I increasingly see academic writing divided into:
defining the question,
vetting the evidence,
refining the argument,
rather than producing prose.
The sentence-level writing itself is becoming the least valuable part.
2. Documentation is being transformed.
I agree strongly regarding note-taking and ambient documentation. If AI can reliably produce documentation while clinicians maintain eye contact and listen, that's genuinely valuable.
The real challenge is not generation. It is:
accuracy,
medicolegal accountability,
integration into workflow,
preservation of clinical reasoning.
3. Routine descriptive analyses are becoming commoditized.
This is probably the most provocative point, and there's some truth to it.
A paper that merely reports:
"Here are all FDA oncology approvals from 2011-2025"
without a compelling conceptual question is going to have less value.
AI lowers the cost of:
extracting data,
generating figures,
performing standard analyses,
summarizing trends.
The "minimum publishable descriptive unit" is becoming less defensible.
Where I think he is wrong
1. Review articles are not primarily information products.
This is the biggest flaw in the post.
A good review is not valuable because it aggregates papers.
It's valuable because it:
frames a field,
resolves tensions,
identifies conceptual gaps,
advances theory,
provides interpretation.
AI can already generate the equivalent of a mediocre review article.
What it cannot reliably do is produce the equivalent of:
a landmark conceptual review,
a theoretical synthesis,
a paradigm-shifting interpretation.
The value was never the references. The value was the thinking.
In fact, as AI floods the world with summaries, genuine synthesis may become more valuable.
2. Meta-analyses are not just statistical aggregation.
If you've ever done one, you know that the difficult work is:
deciding inclusion criteria,
identifying hidden biases,
assessing methodological heterogeneity,
interpreting contradictory evidence.
AI may automate much of the mechanics.
But the judgments remain.
A fully automated meta-analysis can produce precise nonsense if the underlying studies are conceptually incomparable.
3. Medical guidelines are not purely evidence exercises.
This is where I disagree most strongly.
Guidelines involve:
value judgments,
risk tolerance,
stakeholder negotiation,
resource allocation,
ethics.
The recommendation is not simply:
"What does the evidence show?"
It is:
"Given imperfect evidence, what should society recommend?"
Those are fundamentally normative decisions.
AI can support guideline development.
I don't think it can legitimately replace accountable human governance.
As a medical education researcher
This post highlights something many academics are reluctant to admit:
For decades we've conflated:
the production of scholarly artifacts;
the generation of scholarly insight.
AI mostly attacks #1.
Many academic reward systems were built around visible products:
reviews,
narrative summaries,
descriptive studies,
conference abstracts,
committee documents.
If AI can generate those cheaply, then our evaluation systems become unstable.
The important question becomes:
What intellectual contributions remain difficult to automate?
My suspicion:
asking important questions,
building conceptual frameworks,
interpreting ambiguity,
understanding social context,
exercising judgment,
establishing trust,
remain comparatively resistant.
Interestingly, these are precisely the things medical education has often struggled to measure.
The irony
Prasad compares resistance to AI with resistance to the printing press.
I think the better analogy is calculators.
Calculators eliminated huge amounts of arithmetic labor.
But they did not eliminate mathematics.
In fact, they increased the premium on conceptual understanding because computation became cheap.
Similarly, AI may eliminate much of the labor involved in:
writing,
summarizing,
formatting,
extraction,
basic synthesis.
That doesn't mean it replaces scholarship.
It means scholarship shifts upward.
The key error in the post is treating information processing and intellectual contribution as the same thing. For much of academia they have been entangled, which is why AI feels so disruptive. But they are not identical, and the distinction becomes clearer every month.
For those that had an army of pre-med, residents, and others writing papers for them, they now have AI is a great point. Similar with the summary overview of reading…I mean does anyone really want to read the things you should stop doing already because of AI? Honestly, I agree with the points in your post, what I fear is that those of us tat have gone deep, and wide…can use the technology to expedite and scale ourselves. There is an entirely new generation that does have the width of knowledge and doesn’t care about the depth. If technology is neither good or bad, but it is how people use it, then that becomes the underlying challenge that we as the elder statesmen need to remember as we lead.
Maybe we’re walking backwards to the future.
This is satire but very well made
Anyone who believes AI is able to replace humans at current is a fool. MAYBE one day they will get them advanced enough so they can do jobs done by humans but they are NOT there today even though that's how they are marketing AI to be now!
AI is quiet dumb. It is a little bit better than a generic google search but it is FAR from being able to replace humans. One need only chit chat with one of these AI bots for a little bit and you'll see them contradicting themselves and giving very much false information which when questioned about the AI may acknowledge or it may deny and argue with you instead.
We do not know how these tins work b/c they are grown and not built/created. These AI's are the digital equivalent of a gardener raising some vegetation. The gardener is NOT a botanist and thus doesn't know the inner workings if the plant only how to grow it by feeding and nurturing t, They do this with these AI bots! They have zero clue as to how these AI's work,
Therapist here, most of my colleagues are terrified of AI - What if a tool built around alllll the input you provide it with 24/7 mostly free availability better than a therapist? Maybe? So be better? Adapt? The clients I have using AI aren't using it instead of therapy - they aren't canceling sessions because of their bot. They Are bringing in their thoughts in more organized, self examined ways, and for some, it has felt transformative. My role is shifting slightly. Deep interpersonal work is happening more quickly. But it feels in tandem with our work in sessions, not instead of. And for my teammates who refuse to use a tool that seems inevitable to human society, they may need to put down that iPhone, something also changed the face of technology.
You are the "boss" of your AI. You can give your AI parameters, things to look at, things to ignore, things to analyze more deeply. You can push your AI in the direction you want to. Don't ever think you are supposed to simply accept the first answer AI gives you. Debate with it. You will be amazed at how quickly it catches on and comes over to your side with deeper analyses, using data that it initially ignored. Good luck! Have fun!
I ran this post through AI and it asked "Why do you still have humans in healthcare"?