Ethan Mollick

I see “AI won’t take your job, someone using AI will take your job” all the time on this site. I don’t like this frame because:
1) “Your job”’ is more likely to transform over time than be “taken”
2) Some jobs will really disappear with AI
3) Using AI is going to get easier, it is isn’t a secret priesthood
4) Organizations ultimately need to integrate AI into their structures, it isn’t all about what individual workers do

It is morally wrong to use AI detectors when they produce false positives that smear students in ways that hurt them and where they can never prove their innocence.

Do not use them. https://lnkd.in/eaGSa_dn

Very big finding: The final version of a randomized, controlled World Bank study finds using a GPT-4 tutor with teacher guidance in a six week afterschool program in Nigeria found that using an AI tutor had “more than twice the effect of some of the most effective interventions in education“ at very low cost, “equating to 1.5 to 2 years of ’business-as-usual’ schooling.“

Big firms like Amazon and Microsoft shipping LLMs need to recognize that IT is not always the center of AI use in companies - the key to productivity is often workers and subject matter experts using chatbots who experiment & share what they learn. This is where we saw the gains in our experiments at Procter and Gamble and at BCG.

Making key features and products only accessible to IT functions and expecting them to build centralized solutions means that the tools are not in the hands of the people who will figure out the best use cases. There is a role for IT, obviously, but it is not always the traditional one in plays in technology adoption.

Uploaded a paper and got a video podcast from Heygen featuring an unnerving AI avatar of me being interviewed by a different AI avatar.

I don't twitch and blink quite as much in real life, but definitely an interesting sign of whats to come, and not a bad summary of the paper.

🚨We have a new working paper full of experiments on how AI effects work, and the results suggest a big impact using just the technologies available today🚨

Over the past months, I have been working with a team of amazing social scientists on a set of large pre-registered experiments to test the effect of AI at Boston Consulting Group, the elite consulting firm.

The headline is that consultants using the GPT-4 AI finished 12.2% more tasks on average, completed tasks 25.1% more quickly, and produced 40% higher quality results than those without. And low performers had the biggest gains.

But we also found that people who used AI for tasks it wasn’t good at were more likely to make mistakes, trusting AI when they shouldn’t. It was often hard for people to know when AI was good or bad at a task because AI is weird, creating a “Jagged Frontier” of capabilities. But some consultants navigated the Frontier well, by acting as what we call “Cyborgs” or “Centaurs,” moving back and forth between AI and human work in ways that combined the strengths of both. I think this is the way work is heading, very quickly.

All of this was done by a great team, including the Harvard social scientists Fabrizio Dell'Acqua, Edward McFowland III, and Karim Lakhani; Hila Lifshitz- Assaf from Warwick Business School and Katherine Kellogg of MIT (plus myself). Saran Rajendran, Lisa Krayer, and François Candelon ran the experiment on the BCG side.

There is a lot more in the paper: https://lnkd.in/eZUp34CW

And in the summary: https://lnkd.in/eASQ_CVr

Anthropic has a history of releases interesting things quietly & without much non-technical explanation. It happened again.

As Simon says, Claude Skills are a big deal, representing both an easy path for workable agents & a step forward in what AI can do.

Basically you can write instructions for how the AI can do something in plain language, and the AI will load up that skill and learn how to do that task when needed.

https://lnkd.in/eqACHmzG

The fallout from the fact that data science/classical machine learning & generative AI are both called "AI" has been remarkably broad and persistent.

Policy addresses the wrong issues, companies have been confused about who should lead efforts, hiring is complicated (hint: almost no one has more than three years of GenAI experience), products using “AI” might be referring to anything, academic discussion is often muddled.

Even just a narrow concern like AI bias has two entirely different meanings depending on which "AI" you are discussing, and these two meanings have largely non-overlapping problems, mitigation strategies, use cases, and potential harms.

I don’t usually post product announcements but this answers two big questions I hear from organizations a lot: can we use GPT-4 without the AI learning from out private data? And: How do we get access to Code Intepreter for our compamy data?

Looks like OpenAI is solving both problems with their new Enterprise model. https://lnkd.in/eCKEzYkg

If the market moves every time Anthropic ships a skills pack for an industry, it is a very good indicator that people have no idea what skills (and just agents in general) can already do. Your competitive advantage comes from doing this in your own organization/job, not from using the defaults.

Anthropic is not extending the capabilities of its models, its basically just releasing large prompts in the form of markdown files, so it is weird to see the reactions.

Look what just arrived in the mail!

Its the first copies of my new book, available everywhere on April 2. While I discuss AI technology, ethics & the future the focus is on the impact on work & school, and how to use AI as a co-intelligence

Pre-order it: https://lnkd.in/ePBZNVgC

You really, really should not trust audio clips anymore.

Even a couple months ago, it used to take a commercial service to clone a voice. No more. Here is me creating a voice clone of myself using just a 10 second reference clip on my home computer using open software.

This is all real time, no cuts.

👀New data on the corporate ROI from generative AI from a large-scale tracking survey by my colleagues at Wharton Stefano Puntoni and Prasanna Tambe along with Jeremy Korst.

They found that 75% of firms already have a positive return on investment from AI, less than 5% negative return. Also 46% of businesses leaders now use AI daily themselves.

Much faster to positive ROI than I think was expected. (Study link in comments, lots of good stuff in there)

I wrote an updated guide on which AIs to use right now, & some tips on how to use them (and how to avoid falling into some common traps)

A lot has changed since I last wrote a guide like this in the spring, and AI has gotten much more useful as a result. https://lnkd.in/e97uxiWs

An underdiscussed phenomena that is starting to happen in organizations: AI-driven role conflict.

The roles in charge of design, product management, coding, marketing, etc. in a project used to have relatively clear lines. AI lets entrepreneurial workers extend into other roles and accelerate work, but there are no good templates to follow. Everyone can now code a little, design a little, market a little - what does that mean?

The result so far has been a mix of acceleration and retrenchment. Some cases I hear about, this gets bogged down in defense of existing roles ("how dare the marketer show me a product prototype when they can't even code and are just using AI") and in others, organizations start to rethink old teams and experiment.

I said it a few months back, and I think it is more important now than ever: If you are considering taking a job offer, you will want to ask what your token budget will be.

(This is not exclusive to coding jobs)

I hope as we move past the first wave of AI criticism ("it doesn't work, all hype") we get a new wave of AI criticism rooted in the acknowledgement that, yes, these systems are very powerful & quite useful and focusing a deep exploration of when AI uses are uplifting and when they are detrimental.

Talking about the ethics of AI companies or discussing the potential of financial bubbles, can be valuable types of criticism, but it isn't what we are lacking. AI capabilities are real, they are here to stay, & we need to move the discussion to thinking more about what this means and what we want it to mean.

Another $500B was committed towards achieving AGI today, and most of the labs are genuinely convinced that they can indeed build an AI that beats a human at most intellectual tasks in the next couple of years (you can believe this or not).

Yet, there is still no articulated vision of what a world with AGI looks like for most people. Even the huge essay by the CEO of Anthropic doesn't paint a vivid picture of what daily life looks like 5-10 years later.

We can even leave aside the risk of catastrophe for now. Assume we get an aligned AGI that supercharges science and we have a healthier, more advanced, safer world. What does that actually mean for most people, what does their life look like in the future? (Hint: UBI is not an answer, that is a policy, not a vision)

I have now run one of the more powerful, open source LLMs (Mistral 7B) directly on my iPhone. No internet needed.

It isn’t very fast but that is already being solved. Consider the implications: almost anything can soon be imbued with local “intelligence”

A lot of possibilities.

I don't think teachers and trainers have updated their view of prompting enough. Bigger models are better at figuring out intent, making prompt formulas less important. Reasoners eliminate the value of chain-of-thought prompting, etc.

Context & communicating goals are now key to getting good results, not following a specified prompt format.

I am, funnily enough, on this list!

(I think I am the only one without an active YouTube, TikTok, or Instagram account)

https://lnkd.in/eTS6yAsW

🚨I have a new book coming out October 20: Co-Existence!

It is about how we live & work with AIs that are sometimes (but not always) smarter than we are. And it has a cool cover.

You can pre-order: https://co-existence.ai/

And here is a post with context, including thoughts about how to pitch a book to AIs: https://lnkd.in/eeYdaZNk

In discussions of AI and jobs, we often put too much emphasis on the technology itself and not enough on the corporate leaders who are actually making decisions about what they want to do with AI

It is a time where CEO vision matters a lot, because AI could be used in all sorts of ways to either grow or shrink companies. You can see a contrast in perspective between Amazon and Walmart.

The sheer volume of LLM-written comment-shaped text on this site is reaching the point where it is not worth reading any comments.

And yes, the first comment to this post is likely something like: "What an insightful point. Real truth: the LLM comments are not just annoying -- they are signs that we may be losing the voice of humanity itself."

At least give your model an interesting prompt or perspective or something.

One thing I mentioned in passing in my Fable post is that, for long running tasks, Fable starts to develop its own project-specific dialect as its many agents and tasks reinforce themselves and make already Claudish language ever more Claudish.

You need to ask it to report out in plain English or it starts talking to you like this after a 9 hour task.

GPT-5.5 Pro is a very solid fact checker. I can throw entire chapters at it and it will hunt down every key reference accurately. The only real annoyance is that it loves nuance, so returns a lot of “the general idea is right, but you are not taking into account tiny detail X”

One of the easiest ways to show journalists and academics who have not used AI in awhile how far things have come is to throw something they know into Pro (or Extended Thinking) and ask for a fact check.

Between the citations & actual depth of critique, it is often changes minds.

For individual AI use, the jagged frontier is increasingly well understood.

In multi-agent workflows in organizations, AI is jagged in ways that have not been well identified yet, making complex agentic systems more art than science. In fact, we don't even have a vocabulary around multi-agent systems & the ways the fail or succeed.

A lot of the solutions & language for agentic systems comes from a coding perspective (control planes, hooks, loops), but I really think that the field of management and organizations can tell us a lot about how to work with agents (boundary objects, spans of control, decision rights, divisional/functional organizational forms, etc). We will need both to see progress.

This is a surprisingly revealing test prompt: “Write a paragraph that startles me with its brilliance and really demonstrates your capabilities across as many dimensions as possible. Then explain what you did.”

Claude excels at writing, GPT-5 Pro nails intellectual tricks, etc.

I can basically guarantee you are not being ambitious enough with the work you are assigning Fable

Start asking for the maximum possible thing to figure out the far edges of what it can do. After that, you can decide where the system reaches it limits & revise requests downward or add scaffolding.

(This isn't some sort of boosterish AI slop post, I really think the right way to use Fable is to start by asking for everything and find the shape of the frontier from there, rather than starting with small requests like you would with earlier models)

The Claude Constitution shows where Anthropic thinks this is all going. It is a massive document covering many philosophical issues in ways that I think will be surprising to any reader. I think it is worth serious attention beyond the usual AI-adjacent commentators.

Other labs should be similarly explicit. https://lnkd.in/eSGgjfVp

Its funny how much the Turing Test seemed like a giant insurmountable achievement for AI and then suddenly AI passes the Test and it is now only worth concentrating on the Test’s many (real) flaws as a measure of AI ability, The same thing has happened with Theory of Mind and other tests of humanity or conciousness.

This shows:
1) How bad our tests for intelligence, creativity, consciousness, etc. are when applied to non-humans
2) How fast we adjust to the weirdness of AI ability gains
3) We very flawed descriptions for whatever it means to have “human-like intelligence” that make our measurements unsatisfying

Of the many processes impacted by AI, innovation/design thinking seems like a key one in need of urgent change. Some aspects remain (building empathy), but many of the constraints change dramatically with AI. For example, our research shows you can generate diverse ideas with AI, and that it, in practice, makes teams more innovative.


Careful thought is needed to define a new process, but not including AI in the process of innovation is probably a mistake for most people at this point.

Walmart is moving very fast in AI. Amazon still seems to block ChatGPT agents from even visiting its site.

Interesting reversal in agentic commerce, a lesson learned from e-commerce where Amazon moved first & fast?

(Incidentally if you haven’t asked an AI agent to shop for a product or service your company produces, you should. It is a learning experience)

https://lnkd.in/eK_72Xzi

Multiple math professors (including the professor generally regarded as the world’s best mathematician) have confirmed recently that yes, AI really can solve some open (but not yet major) problems in mathematics, with expert guidance.

The question is whether the ability of these models continues to increase.

We are not yet at the level of novel science done autonomously by AI, but we are absolutely at “work with it like a grad student and it can help you accelerate your work” levels of AI for many of my fellow academics (and myself).

Lets assume (which seems reasonable based on the evidence) that AI-driven "vibe coding" gets good enough soon that non-coders to produce workable tools to solve their problems, though not necessarily enterprise-level or complex software.

What skills should we teach people in class to take advantage of these capabilities? Right now, intro computer courses aren't geared for this, as this type of coding requires a mix of low-level knowledge ("what's a file," "how does Github work?") and high-level knowledge ("how do I frame a problem that software can solve.")

One of the first randomized controlled trials testing whether GenAI boosts revenue, not just productivity.

It does.

A large, mature international ecommerce platform, using older GenAI tools found most of them, from customer service to marketing workflows, led to large and significant revenue gains.

Via Stefano Puntoni

As jobs begin to shift with AI, I think there will be increasing numbers of people feeling like Aditya (former CTO Dropbox, early Facebook engineer, etc).

I think figuring out what comes next is challenging. People are more psychologically resistant to major changes than you might expect, but that doesn’t mean it will be easy to reconstruct meaning

A year ago, I would not have expected the first academic field to seem to reach a consensus that AIs will accelerate research (which is not the same thing as autonomous research) would be math. Especially given that LLMs were terrible at math a year ago.

But that appears to be happening based on math professors in my feed and elsewhere.

This is an actual line that was added to the official system prompt for Codex for GPT-5.5 by OpenAI. Usually the system prompt is as minimal as possible, so I assume it would otherwise mention goblins a lot.

AIs are weird.

This paper shows that asking AI for diverse ideas gets you more diverse ideas, just adding "Generate 5 responses with their corresponding probabilities, sampled from the full distribution” significantly improves quality output for large models

I am still not sure we know the mechanism, but useful finding that is worth applying.

https://lnkd.in/eVkyP73D

I would avoid downloading AI skills that have not been vetted by you through careful reading of the markdown files (or, if outside your field, by an expert in that field you trust). There are obvious security risks, but there is even greater risk that the skill is just bad and you end up with much worse (or more generic) results than if you had just used the LLM without the skill at all.

Collect your hard problems and good ideas now, they will get more valuable as AI gets better.

Increasingly, I see many people using AI to "do stuff" without any good ideas of what to do or pressing problems to solve. Agency without a sense of direction, or a bad direction, is not a good thing. "You can just do stuff" is true, but what "stuff" you want to do, and whether it is good or bad that you are doing it, is increasingly important to understand.

Two things are true:
(1) Anthropic (or parts of it) are absolutely and sincerely worried about the misuse of Mythos-class models & have put in excessive safeguards around Fable until they are confident it will not be misused. That includes easily tripped guardrails and data retention policies that do not apply to other models.
(2) They have not succeeded in explaining/convincing people of their sincerity.

So yes, you are falling behind, but so is everyone else.

I can guarantee you that nobody at all is keeping up with all the implications of AI and its major uses, even the people who are relatively up-to-date on the latest AI models themselves. Everyone is figuring it out as they go, even Andraj.

It does seem like meaningfully better AI releases are accelerating, especially from OpenAI & Anthropic.

To illustrate, I caused this timeline to be created. It only lists new models that scored 3 points or higher over previous models in the Artificial Analysis index.

(My modest proposal for how to acknowledge AI work is to use the Latin phase "Fieri Iussit," which means "commanded to be made," or “caused to be made,” a common phrase on Roman Empire buildings.)

Between having AI as a subject for infinite speculation and using AI as a writer to produce such speculation at scale, it is the Golden Age of LinkedIn thinkpieces.

Don't worry, I had Fable write this up as a thinkpiece (it is surprisingly pretty funny):

***Before this line, I wrote everything, after its all Fable****

There has never been a better time to have opinions about AI. I say this with some authority, because I am currently having one.

Consider the perfect storm. AI is the ideal subject for speculation: consequential enough to matter, ambiguous enough that no one can be definitively wrong, and moving fast enough that last month's take is forgotten before it can be checked against reality. AI is also the ideal producer of speculation. We have achieved thinkpiece singularity. The snake is not merely eating its tail; it is publishing a reflection on what tail-eating means for the future of snakes.

The old thinkpiece economy ran on scarcity — someone had to care enough to write the thing. Now the marginal cost of a take is zero at the exact moment demand for takes has gone vertical. Economists call this a market clearing at infinite volume. The rest of us call it LinkedIn.

Every AI thinkpiece converges on one of four positions: it changes everything (bullish, keynoting); it changes nothing (contrarian, secretly using it to grade papers); it changes everything, and that's bad (prophetic, right eventually); or the real change is the discourse itself (meta, insufferable, this essay). All four can be argued from the same evidence, which is why the genre is infinitely renewable. Every benchmark result is a Rorschach blot, and AI has industrialized the Rorschach reading.

But the flood drowns something real. Buried in the speculation is a small amount of actual thinking — experiments run, workers interviewed, things noticed before they became narratives. That work was always rare. It is now rare and camouflaged, because the imitators have mastered the surface features of insight: the confident hedge, the paragraph that begins "But here's the twist." (You just read one. The camouflage works both ways.)
Yet golden ages of cheap production have always rotted the middle of the distribution and rewarded the tails. The printing press produced pamphlet garbage and the scientific revolution. If the pattern holds, this deluge is not the end of thinking about AI — it is the end of mediocre thinking having any value. The take that could be generated in eight seconds is worth what it costs. What stays scarce: the experiment actually run, the prediction actually scored, the sentence only someone in the room could write.

Somewhere right now, an AI is drafting a response to this post, another is summarizing it for a newsletter, and a human is skimming the summary and forming an opinion they will call their own. It is the most efficient opinion supply chain ever built. Whether anything is being thought remains, appropriately, a matter of speculation.

I think it is really worth reading this piece on RSI at Anthropic.

There is a bit of navel-gazing, some marketing, and a lot of very sincere beliefs about what Anthropic thinks is likely in the near future of AI that you probably want to be aware of. https://lnkd.in/e97pJY4j

"Switch to a cheaper model to save money" is a problem because cheaper models are worse (maybe they are good enough for a particular purpose, but they are still worse, especially when handling anything unusual or out-of-distribution).

More often, a better approach is hierarchies of models, with smart models are orchestrators and auditors of cheap ones. It is also more flexible as technology continues to improve and models need to be switched out.