Since the advent of ChatGPT, I have been a fan of Wharton Professor Ethan Mollick’s One Useful Thing Substack blog. Not only am I a subscriber, but I have also recommended it to many friends.
Last week, Professor Mollick posted an article titled The Overhang. After writing about a few fun projects that he used AI tools to create, Mollick wrote about four specific advantages that humans have if they “want to use AI in unique and enhancing ways: deep knowledge, wide knowledge, taste, and agency.” The discussion about the four advantages is from his upcoming book, Co-Existence, a sequel to his previous book, Co-Intelligence.
I copied the section of the post that was about the four advantages and pasted it into a Word document. I then uploaded the document to ChatGPT Pro with the following prompt.
I am uploading a section of a recent blog post by Wharton professor Ethan Mollick. Dr. Mollick states that there are four human advantages over AI tools that we should consider as we evaluate how to work in a world with continually improving AI tools. Do you agree with his premise? Are there other advantages that should be mentioned in a 1,500 to 2,500-word blog article that you could write for me to critique his article?
I had intended to paste ChatGPT Pro’s response below. However, that’s not the end of my story. I decided that I would upload the same document to several other AI tools. I wanted to see how similar the responses were and evaluate them for their differences, if any. The document that I uploaded was the same, and I did not change the prompt. The responses were different. I have an overview of each response as well as a link to the individual articles to keep the length of this post to 2,000 words or less, instead of 10,000 plus.
ChatGPT Pro
ChatGPT Pro is my most frequently used AI tool. It knows that I am an educator and a business person. Its response referred to education frequently. It cited six articles and research papers. While I found those sources to be relevant to the analysis and conclusion, I have no idea if those six citations are the best sources possible.
The title of the ChatGPT Pro article is:
What Ethan Mollick’s Four Human Advantages Get Right—and What I Would Add
ChatGPT Pro mentions four additional contributions that it says deserve attention, even though they may fall within the definitions of knowledge, taste, or agency (Mollick’s four advantages, with deep and wide knowledge consolidated into one). These are:
- Purpose and problem framing
- Accountability and moral responsibility
- Reciprocal relationships, trust, and collective action
- Firsthand experience and local context (knowledge acquired through direct involvement)
ChatGPT’s concluding paragraph is an excellent summary of its response. Titles and bolded words come from the response and are not mine.
A broader definition of human advantage
Mollick gives us a useful place to begin. Deep knowledge, wide knowledge, taste, and agency are all qualities worth developing.
I would not discard his framework. I would broaden its purpose.
We should prepare people not only to direct AI toward better outputs, but also to identify worthwhile goals, evaluate consequences, sustain reciprocal relationships, investigate what the available information leaves out, and accept responsibility for decisions.
None of this requires predicting a permanent technical ceiling for AI. It requires deciding what kind of people and institutions we want to develop as the technology improves.
Human dignity should not depend on continually finding another task at which we can outperform a machine.
The goal should be more than better AI outputs. It should be more capable people, more trustworthy institutions, and decisions we are prepared to stand behind.
Gemini Pro
I use Gemini frequently enough that I have a subscription to the Pro version through Google. One thing I noted was that the Gemini response was the fastest among all the tools that I used. Unlike ChatGPT Pro, Gemini Pro did not provide any sources for its points, other than the Word document I uploaded containing the partial article from Mollick’s Overhang.
The title of the Gemini Pro article is:
The Executive Dilemma in the Age of AI: Why Curation and Prompting Are Not a Durable Strategy
Gemini Pro suggests that the four advantages are thin-walled defensively. What is missing are four “institutional” imperatives. These are:
- Moral liability and fiduciary “skin in the game.”
- Tacit, embodied, and politically entangled context
- Relational trust, formation, and pastoral care
- Radical intent and non-consensual risk.
Gemini Pro argues that humans possess the capacity for radical conviction, which it defines as the willingness to pursue an unproven path because their vision demanded it. AI cannot originate conviction.
The concluding section of Gemini Pro is on target.
Beyond the Curation Trap
Ethan Mollick’s contribution to our understanding of the emerging AI landscape is valuable. He correctly urges us to abandon the futile ambition of competing with machine intelligence on raw processing capacity. But his four personal advantages—deep knowledge, wide knowledge, taste, and agency—remain primarily a defense of the individual knowledge worker operating within digital workflows.
For institutional leaders charged with shaping the future of corporations and universities, curation is an incomplete framework. True human advantage does not rest on our ability to serve as high-end prompts or tasteful editors for probabilistic engines.
Our durable advantage rests on our willingness to carry liability when stakes are existential, our capacity to read the unwritten currents of the human heart, our commitment to build trust through shared vulnerability, and our courage to take risks that defy statistical precedent. As the generation of content, code, and analysis becomes commoditized, the value of the non-computable aspects of humanity will only grow. The mandate for higher education and corporate leadership is clear: stop training people to think like machines, and start running institutions that honor what only humans can do.
Claude Opus 5
I rarely use Claude, but fortunately, it gave me three free opportunities to use its powerful Opus 5 version when I signed in.
The title of the Claude Opus 5 article is:
The Four Advantages Are Right. The List Is Incomplete.
Claude’s response did not provide any citations. Since it’s not my normal “go-to” AI tool, perhaps the weakness was that I did not ask for any in the prompt. However, I wanted to use the same prompt that I used with ChatGPT Pro.
Claude states that the four advantages are all about being good at working with the AI model. However, there is an axis that Mollick doesn’t discuss (in this essay that Claude ingested), which has advantages that have nothing to do with capability comparisons at all. These are:
- Accountability
- Access to unrecorded reality
- Legitimacy in relationships
- Choosing what problem to work on
- Verification as a discipline
Claude’s conclusion, appended below, provides an excellent summary, but the entire article is worth reading, particularly if you like reading a biting critique.
The reorganization I’d propose
Sort the list by why the advantage holds, and two groups fall out.
Advantages that are contingent on current model limitations: taste, verification, some of deep knowledge. These are real today. They’re also the ones a better model erodes, and anyone planning a career on them should hold them loosely.
Advantages that are structural: accountability, access to unrecorded reality, relational legitimacy, problem selection, and the parts of wide knowledge that are about framing. These don’t depend on a capability gap. They’d survive a model that was better than any human at every task with a text interface, because they’re about position in the world rather than performance on a task.
Mollick’s four are mostly drawn from the first group, which is why the essay reads as advice for the next two years rather than the next twenty. Given that he’s writing about a frontier he describes as unmapped and fast-moving, that’s a defensible choice. But the sorting is the thing I’d want a reader to take away, and it isn’t in the piece.
His conclusion, that inevitable change does not mean the type of change is inevitable, is the right note. I’d just add that the type of change is determined far more by how institutions allocate accountability and permission than by how many individuals cultivate good taste.
Perplexity
I rarely use Perplexity, even though I have friends who swear by it, because it gives you the option of using many different AI tools. I deliberately did not choose any of the available options. Instead, I uploaded the file and used the same prompt I used for the other tools.
The title of the Perplexity article is:
Beyond Prompting: The Human Advantages That Matter in an AI World
Perplexity’s response provided me with 16 sources, but only two were specifically referenced next to the text where they were used.
Perplexity writes that the list of four advantages needs to be expanded because the central challenge is to develop people who can determine what ought to be done, assess whether it is being done well, and accept responsibility for the consequences. It recommends five additional advantages to consider:
- Judgment: deciding under uncertainty and trade-offs
- Moral responsibility: someone must own the consequences
- Relationships: trust is not a byproduct of fluent communication
- Embodied and local knowledge: reality resists clean abstraction
- Purpose: humans decide what is worth optimizing
The conclusion appended below provides an excellent summary.
A practical agenda for leaders
Organizations should resist two unhelpful extremes. One is technological denial: refusing to experiment because AI is imperfect. The other is technological surrender: assuming that a fluent system should be trusted with decisions simply because it can produce them quickly.
A more productive agenda has five parts:
- Build AI literacy across the organization. Employees need practical experience with the tools, their strengths, and their failure modes. They should understand that confidence and correctness are not the same thing.
- Protect and develop expertise. Do not treat AI as a substitute for learning the profession. Use it as a tutor, research assistant, simulator, and productivity partner—but retain the human capacity to verify and challenge its output.
- Create safe, structured experimentation. Establish approved use cases, privacy rules, review requirements, and ways to share lessons. Reward employees who identify both opportunities and risks.
- Make accountability explicit. For consequential use cases, specify who owns the decision, how outputs will be checked, how errors will be documented, and how affected people can seek correction.
- Lead with mission, not novelty. Begin with the organization’s purpose and the human problem it is trying to solve. Then determine whether AI helps advance that purpose without compromising trust, fairness, privacy, or dignity.
Mollick is correct that people should not compete with AI simply by trying to produce more output. The real opportunity is to become better at the work that surrounds output: framing the problem, asking the question, recognizing quality, interpreting context, choosing among trade-offs, building trust, and accepting responsibility.
Deep knowledge, wide knowledge, taste, and agency are indispensable. But the future of human work will depend equally on judgment, moral responsibility, relationships, contextual awareness, and purpose.
AI may help us do more. Humans must decide what “more” is for.
At the end of the Perplexity article, I was offered a choice to “use the computer” for a stronger article. Since I am not a regular Perplexity user, I clicked that button, and a slightly different article and response were generated (that I am calling Perplexity 2).
Perplexity 2
The title of the Perplexity 2 article is:
Four Advantages, and a Few Mollick Left Out
Perplexity 2 recommends adding several advantages, specifically:
- Accountability and responsibility
- Relationships and trust
- Judgment about ends
- Lived context
- Epistemic vigilance
The conclusion to the “computer” response is appended below. It might be the most creative response of the group.
The View From the Overhang
Mollick ends where I’d like to end: the question is not whether change is coming — it is, even if the labs all stopped tomorrow — but what kind of change it will be. His four advantages are an excellent answer to the individual version of that question: how do I get more out of these tools than the next person does?
But there is an institutional version too, and it’s the one his list underserves. How do we build organizations, professions, and an economy in which the humans who bear responsibility, hold relationships, choose the ends, know the context, and check the answers remain at the center — not as a sentimental gesture, but because those are the functions we should not casually delegate away.
So: yes to deep knowledge, wide knowledge, taste, and agency. Read widely, go deep, know what you like, and try weird things. But if you’re planning your career, your team, or your school around the AI transition, add five more to the ledger. Take responsibility for what ships. Build relationships that the tool cannot hold. Keep asking whether the thing being optimized is the thing worth doing. Show up, physically, where the work actually happens. And stay just suspicious enough of anything that arrives fluent, confident, and cheap.
Those are the advantages that the overhang does not erase. These are not claims that AI will never simulate parts of these capacities; they are claims about where human authority, trust, and responsibility still matter — and are still worth building on.
A Few Final Thoughts
I know more than a few friends and colleagues who have admitted to using an AI tool to generate an analysis and a report about that article. I have tried to avoid doing that except in situations like this, where I found it useful to upload a specific source article and, using the same prompt, compare the responses generated by the AI tool.
All of them are getting better. Each of the tools that I used clearly ingested the article enclosed in the Word document, assessed Ethan Mollick’s four advantages, and, as I asked, added other components to those four advantages. Two of the tools went beyond my upload and located Mollick’s original article as published on Substack.
Some of the tools were better critics than others. In the way the articles were written, at least two of the tools appear to be empathetic to the human cause and tailored their conclusions to appeal to those who are hopeful that their job and the jobs of others aren’t going to be eliminated by AI.
From my perspective, most of the AI tools added additional advantages that were similar.
ChatGPT Pro added:
- Purpose and problem framing
- Accountability and moral responsibility
- Reciprocal relationships, trust, and collective action
- Firsthand experience and local context (knowledge acquired through direct involvement)
Gemini Pro added:
- Moral liability and fiduciary “skin in the game.”
- Tacit, embodied, and politically entangled context
- Relational trust, formation, and pastoral care
- Radical intent and non-consensual risk.
Claude Opus 5 added:
- Accountability
- Access to unrecorded reality
- Legitimacy in relationships
- Choosing what problem to work on
- Verification as a discipline
Perplexity 1 added:
- Judgment: deciding under uncertainty and trade-offs
- Moral responsibility: someone must own the consequences
- Relationships: trust is not a byproduct of fluent communication
- Embodied and local knowledge: reality resists clean abstraction
- Purpose: humans decide what is worth optimizing
Perplexity 2 added:
- Accountability and responsibility
- Relationships and trust
- Judgment about ends
- Lived context
- Epistemic vigilance
Lined up next to each other, four different tools added similar advantages to consider. Perhaps it was the probabilistic nature of the “thinking” built into the generative AI tools themselves. Nonetheless, if I were using these tools to frame a personally written critique of Mollick’s essay, I’d have a healthy weighting of additional advantages to propose.
I suggest that as the temptation to use and rely on AI increases, we remember that “AI may help us do more – humans must decide what ‘more’ is for.” I can’t wait for my copy of Co-Existence to arrive. Perhaps we’ll see if Mollick has included more than those four advantages expressed in The Overhang.
