Ethan Mollick on AI Adoption, Education and the Uncertain Path to AGI

Ethan Mollick argues that AI labs can neglect product development because they are focused on building more capable models. He also sees a tension in funding narrow applications while predicting that general intelligence will arrive imminently. These are his interpretations of company priorities and hypothetical futures, not established outcomes.

Harry Stebbings interviews Mollick, a Wharton entrepreneurship professor and co-director of its Generative AI Lab. A former entrepreneur, Mollick says his work combines research on entrepreneurship, education and the practical use of AI.

Asked about the newly released Llama 3.1, Mollick says an openly available model approaching leading proprietary capabilities could enable widespread experimentation and fine-tuning. He also expects closed-model developers to release improvements, while emphasizing that even developers do not know all the consequences.

He thinks the next generation will be more capable based on what he has heard, but distinguishes model improvements from successful real-world applications. Social media's weekly competition for a leader can distract from the more important question: how far capabilities will progress and how they will be used.

Mollick lays out four possible trajectories: progress largely stops; capabilities improve roughly linearly; capabilities continue to grow exponentially; or systems reach AGI and superintelligence. He considers an immediate fizzle unlikely but treats all four as scenarios rather than predictions with known probabilities.

A plateau would still leave years of integration into organizations. Linear progress might allow a more gradual adjustment, while exponential growth would create a more disruptive transition. He says public debate concentrates on the extremes and underprepares for the middle possibilities.

Stebbings compares AI to the iPhone, whose early releases improved dramatically before changes became more incremental. Mollick agrees that technologies can reach plateaus, but notes that sustained progress may also involve replacing underlying technologies. Where AI's capabilities ultimately top out remains unknown.

He calls current AI capabilities jagged: a system can perform some tasks impressively and fail at others. That unevenness prevents simple substitution for an entire human job. Whether future systems can overcome the gaps is an open question.

Asked whether compute, data or algorithms are the bottleneck, Mollick says he does not know. He uses the idea of a reverse salient: technological progress exposes a lagging component, which attracts resources and research until another bottleneck appears. He also wants more attention to the human and organizational systems that must use the technology.

He questions the gold-rush picks-and-shovels analogy. In his steam-power analogy, the value emerged when skilled people adapted the engine to specific factory equipment. Similarly, he says, people who connect general AI capabilities to useful work inside organizations can create value.

Stebbings asks why providers offer so little practical instruction. Mollick thinks some labs prioritize scaling toward more capable systems over documentation and product work. He calls the resulting situation documentation by rumor: users experiment and share tips while ordinary customers face an unexplained chatbot.

He recalls a financial institution that built an expensive GPT-3 sales assistant which, in his account, became obsolete when ChatGPT appeared. The example illustrates the risk of building for a moving platform, though it does not establish that every application will be displaced.

On open versus closed models, Mollick favors technological progress and sees opportunities in education, healthcare and entrepreneurship. He also says openness brings genuine risks; the appropriate use of AI in healthcare depends on context, and he does not present it as a substitute for clinical care.

He expects safeguards on open models to be bypassed and worries about scalable spear-phishing and impersonation. He distinguishes those concerns from more speculative future biological risks, saying the systems he describes at the time were not sufficient for the latter scenario.

Mollick argues that corporate incentives shape the openness debate and that neither benefits nor harms should be assumed away. He endorses a model of observing a new technology closely and responding quickly to demonstrated problems, while recognizing that governments may struggle to react that fast.

He asks who will monitor the effects of open model releases, feed evidence back to developers and decide when changes are needed. His proposed response is active observation and timely policy, not a claim that releasing any model is automatically safe.

Stebbings worries that stringent European regulation could constrain development and adoption. Mollick agrees that premature restrictions can be a problem, but rejects a choice between no scrutiny and trying to prevent every hypothetical harm in advance. He says Europe's innovation challenges have several causes beyond regulation.

He also cites research suggesting geographic proximity and direct travel connections matter to venture funding. In his telling, local networks and the ability to monitor companies help explain Silicon Valley's advantages; those observations should not be read as a guarantee that every founder must relocate.

Mollick says AI labs do not always understand how large organizations work. He views OpenAI's Code Interpreter as a potentially important data-analysis product that received insufficient follow-through, and argues that technical documentation alone does not show managers how to incorporate AI into their jobs.

His explanation is that labs may allocate their best people and compute to future model capabilities instead of refining current products. He speculates that OpenAI had a roughly $3 billion annualized run rate, but the interview does not independently verify that estimate or establish actual annual revenue.

He says only a small fraction of people in audiences he addresses have spent substantial time with advanced models, based on informal show-of-hands observations. A blank chat interface and limited onboarding make it hard to discover use cases. He cites a Danish study in which respondents reported time savings on some tasks; the results are self-reported, not a general productivity guarantee.

Clear policy is essential, he argues. Regulated organizations need guidance on permitted uses, and employees need to understand what they may share, how their work will be evaluated and whether useful automation will be rewarded.

Mollick calls some employees secret cyborgs: they use AI for work but conceal it because they fear losing status, colleagues or their jobs, or simply receiving more work. He recounts a manager using ChatGPT to write a policy banning it, illustrating the contradiction he sees in vague rules.

A company that gains productivity can use it to cut costs or expand output and create new opportunities. Mollick says employees will have little incentive to reveal useful practices if the only reward is a smaller workforce. He wants organizations to make their choices and incentives explicit.

Stebbings worries that customer-service employees are already losing work and cites company examples. Mollick agrees that it is too easy to assume displaced workers will automatically find equally good jobs. He recalls telephone operators as an example of technological change whose effects differed by age and opportunity.

Even if a technological revolution eventually raises living standards, he says, the transition can involve serious hardship and unrest. Retraining and adjustment require specific policies and effort; neither an assumption that markets will fix everything nor a generalized disaster prediction resolves the practical problem.

Stebbings asks whether the benefits will concentrate among a small technical elite. Mollick cites an early Danish study in which users skewed male and wealthy, but notes that phone-accessible chat interfaces may lower some barriers. His optimism about wider access does not eliminate differences in resources or skills.

He says people accustomed to ordinary software can struggle when a model refuses a task or produces different answers. His wife, an education researcher and lab co-director, has used prompting effectively without coding; he argues that communication, management and subject expertise can be valuable AI skills.

For the consumer interface, Mollick expects voice and visual capabilities to make interaction more natural, with agents potentially able to take action. This is a forecast, not a description of a fully reliable assistant. He thinks the importance of specialized prompting may diminish if products become easier to use.

Stebbings asks why students share AI use faster than workers. Mollick points to the social environment of universities and the relatively direct fit between chatbots and many homework assignments. Workplace tasks such as marketing may require more context, expertise and adaptation.

Mollick argues that startup incentives favor incremental product-market fit when a fast-changing technology may require more imaginative bets. He does not dismiss customer discovery, but says a narrow product built for today's model could be overtaken if underlying capabilities change.

He urges founders and investors to articulate a view of the future: how capable might models become, where will gaps remain, and how will a product spread through an organization? A clear theory of adoption and domain-specific problems matters more than simply attaching a thin layer to a model.

He points to a tension between predicting AGI within five years and funding companies whose products assume today's limitations persist. He acknowledges that nobody knows what an AGI world would look like; his challenge is to make assumptions coherent, not proof that all startups will fail.

As an example of current experimentation, he describes asking a model to generate product ideas, assess them and create an interactive prototype that can be revised through conversation. The example illustrates a workflow he tried, not an assertion that today's models are AGI.

Stebbings suggests leaders' AGI timelines may reflect their funding needs. Mollick says incentives matter and confident statements deserve skepticism, while career commitments may be a signal of genuine belief. Historical forecasts have often been overly confident, so he advises holding a view with humility.

Stebbings cites a heuristic about whether a startup should welcome a model becoming 100 times better. Mollick objects that the phrase is too vague: better at reasoning, longer documents, factual reliability or something else? Different improvements threaten or enable different products.

He prefers concrete questions, such as whether a model can reliably review a legal document or complete a research workflow. Those tasks depend on context length, hallucination rates, handling of evidence and other specific capabilities. He says the gaps vary and that domain experts are needed to identify them.

Education is one area where he sees opportunity. Mollick says tutoring can be powerful, but disputes the idea that an AI tutor makes teachers unnecessary: students need motivation, practice and guidance, and schools also serve social, childcare and credentialing functions.

He cites an early randomized study of math tutoring in Turkey: students given an unstructured GPT-4 tool achieved higher homework scores but did worse on a later test, apparently because the tool did too much of the work. Better prompting removed that negative effect in his account, but did not show a learning gain. The finding should not be generalized to all AI tutors.

His longer-term hypothesis is a flipped classroom: students learn basics outside class with AI support, then practice and apply them with teachers and peers in class. He says this requires careful scaffolding and implementation rather than simply dropping a chatbot into schoolwork.

Learning requires effort and productive difficulty, he says. Active classroom work lets students practice, make mistakes and get help. Mollick describes a prototype his lab has built for an AI-assisted version of this model, while acknowledging that deployment still has a way to go.

Stebbings asks whether an AI tutor would really be dramatically better than high-quality videos. Mollick says educational outcomes are hard to express as a simple multiplier. He cites a classic human tutoring result of about two standard deviations but explicitly says it may not replicate and is not evidence that AI tutoring will achieve the same gain.

Instead of asking students whether they understand, a good tutor should probe their knowledge, expose gaps and support practice. Mollick highlights reflection, repeated practice and low-stakes testing, arguing that domain expertise is necessary to evaluate whether an apparently polished AI response is actually useful.

Asked whether homework remains worthwhile when AI can complete it, he says yes, but schools must define cheating and design assessments accordingly. One option is controlled in-class work; another is to redesign assignments around AI. He does not endorse intrusive monitoring as his preferred solution.

In his own classes, students use AI mentors, simulations and assignments that require reflection, including teaching an AI to do something. He says these approaches can make learning more active but will take time to develop and deploy well.

On energy, Mollick says the claim that compute becomes a future currency depends on a scenario in which highly capable intelligence is widely useful and demanded. In that hypothetical world, energy and compute could become binding constraints; neither near-term AGI nor a particular energy solution is established by the interview.

He notes uncertainty about energy comparisons and gives rough, time-specific estimates for data-center and AI electricity use. Those figures are his estimates, not independently verified measurements; he argues that the eventual impact depends on demand, efficiency and infrastructure.

Stebbings asks whether voters might one day choose algorithms rather than people. Mollick is skeptical of a rapid move to machine governance because political institutions and preferences are human and complex. He nevertheless expects persuasion tools and AI advice to affect politics.

He cites a persuasion experiment reporting an 81.7% relative difference in a particular comparison with humans. The interview does not provide enough study detail to generalize that figure to all AI persuasion. He also notes that misleading claims often spread even when people do not watch the underlying video.

Stebbings worries that abundant AI-generated content will reduce value and make discovery harder. Mollick says attention was already fragmented and wonders whether AI might also improve connections between creators and audiences. He discusses music generation and book bestseller lists as examples of a changing attention economy.

In the quick-fire round, he says a central concern is losing human agency to organizations and systems that use AI, not only to AI itself. Beneficial outcomes require deliberate choices about work and how people use productivity gains.

Mollick says he has become more confident that further model scaling has room to run, based on accumulating signals from people working in labs. He does not know when new models will appear; this is his assessment in 2024, not a verified forecast.

He expects companies with heavy chip expenses to explore more control over their supply chains. He is also surprised by the apparent cleverness of current models, while emphasizing earlier in the discussion that their strengths and weaknesses are uneven.

One question he thinks receives too little attention is why many people try AI and then stop. He says the answer may involve the way people relate to the tools, not just whether a particular task succeeded.

Another overlooked issue is meaning at work. Mollick worries that even partial automation can make people question the value of their roles, especially if colleagues or managers rely on AI-generated responses. He argues that employment and human purpose both deserve attention.