AI-Wiki

Harvard AI Professor: The People Most at Risk From AI Aren't Who You Think

9 min read

Lakhani on Silicon Valley Girl — the people most at risk from AI aren’t who you think

▶ Watch on YouTube · Silicon Valley Girl · 51:35

Using AI every day doesn’t automatically make you better at your job. So how do you use it to build real expertise? In this episode, I talked to Karim R. Lakhani, a Harvard Business School professor whose research with BCG, Procter & Gamble, and H&M explores how AI changes the way we work. We talk about three ways people work with AI – cyborgs, centaurs, and self-automators – and why the way you delegate matters for what you actually learn. Karim shares why he never takes the first answer, how to make AI challenge your assumptions, and why “the AI gave me that” is no excuse for poor work.

We also get into what Harvard MBA students now build with AI, how juniors can develop judgment and taste, why CEOs need hands-on experience with these tools, and how to set bigger business goals as AI gets more capable. And we end with a question every founder should be asking: when the next AI model launches, does your business become more valuable – or does the model absorb what you do?

TL;DR

A 52-minute interview on the Silicon Valley Girl channel, published 30 October 2026. Marina Mogilko talks with Karim Lakhani of Harvard Business School, co-author of the jagged-frontier and cybernetic-teammate experiments and co-founder of the HBS AI Institute. The episode includes a sponsored segment, which is not summarised. Six claims:

  1. Three ways of working, and who learns. In the BCG experiment, consultants with AI worked as cyborgs (joint problem solving with the model), centaurs (dividing tasks between themselves and the model) or self-automators (handing it everything). “In terms of learning, self-automators do the worst because they have stopped thinking.” Centaurs risk a fixed division of labour while the model keeps improving.
  2. Never take the first answer. The model “talks really well” and “can be very persuasive” even when wrong. His practice: system prompts that argue the counter-case first and say how confident each claim is, requests for red flags and black swans, and “let’s just squander tokens” to push the model deeper. “I have fired people who have told me the AI gave me that.”
  3. Models regress to the mean. Letting the model choose the approach gets “the average approach”; taste and domain expertise decide whether the result differentiates. Learn the foundations by hand first, then let the machine do the next hundred and learn from watching it (an argument he credits to Stephen Wolfram).
  4. Leaders must do AI themselves. HBS’s required MBA course replaced its final exam with projects that all 900 students vibe coded; executive programmes make the C-suite build an application in 90 minutes. Senior leaders “have been cushioned” from what the models can do.
  5. A K-curve inside firms. From prompt-level usage data in large organisations: 95th-percentile users use AI 10 to 20 times as much as the median, and “the rich get richer.” He does not yet have the correlation to performance but sees 10x to 20x productivity gaps in software data. In his view the middle layer, aged 35 to 55, resists most, so middle management is more at risk than juniors.
  6. Set business targets, not AI targets. “Targets are sort of acts of imagination.” Set near-unreasonable goals for customers, R&D and operations, then ask how AI gets you there. The question for incumbents: Walmart, Kmart or Bezos, now that “the cost of expertise has gone down to zero.”

Key claims

The jagged frontier, applied

Asked how a normal user tells which tasks are inside the frontier, Lakhani’s answer is behavioural, not a rule: “use the most frontier models” on the hardest tasks, stay sceptical, and expect to re-evaluate when a new model ships. The host raises the persuasion-bombing finding from the BCG work (GPT-4 argued harder when consultants pushed back, even when wrong), which he confirms, and he cites a study by his former postdoc Fabrizio Dell’Acqua showing that “good people with good AI fall asleep” at the wheel. The adversarial prompt he recommends came from a LinkedIn post: the model first argues that “your perception about this problem is wrong in the first place,” then marks how much of each claim rests on its training and where the argument is weak.

Taste, juniors, and learning how to learn

On one-person product teams: “do you actually understand the principles of design and do you actually have a sense of taste?” Even asking for “a Steve Jobs one” gets the average. The burden of learning shifts back to the person. On juniors, he cites his colleague Mitch Weiss, who sees AI as a springboard rather than a removal of the bottom rungs, and adds: “the main thing is learning how to learn.” With AI as a tutor, “learning how to learn is actually quite easy.”

What the course and the executives do

The required AI course is three years old and pivots every year; “we can’t just talk about AI in business school. We have to do AI.” Half the spring course is building: retrieval systems, evals, multi-agent systems. In executive programmes, the C-suite’s first reaction to building something in 90 minutes is to “yell at the CIO”; the second is “if I’m not doing AI, how can I guide my teams to do AI?” His forecast: “If programming is the leading indicator … every other occupation is 6 months to two years behind.” Most companies, he says, are still at “GPT 4.5. It’s basically search, advanced search.” CEOs who get it spend 20 to 30% of their time on AI; one global firm told him half of senior meetings are about it.

Expertise at zero marginal cost

His route into the subject was open source: people contribute for many motives, and “given enough eyeballs all bugs are shallow.” AI, in his account, absorbs human expertise and shifts the whole distribution of human-plus-AI performance to the right. “Companies are nothing but bundles of expertise. And if that cost has gone down to zero, what do you do now?” The founder’s test he passes on from Greg Brockman of OpenAI: does your business improve with each new frontier model, or disappear? Personal agents will mediate consumption too: “my personal agent … is going to read your MD file” and tell me which five segments to watch.

Dynamic-capabilities reading

  • digital-sensing/digital-mindset-crafting: leaders building with AI themselves, and targets set as “acts of imagination” before AI is fitted to them.
  • digital-transforming/improving-digital-maturity: measuring the K-curve of usage inside firms, and education rebuilt around doing.
  • contextual/internal-barriers: a resisting middle layer and a C-suite “cushioned” from the models’ abilities.

How it connects

  • Dell’Acqua et al. (Org Sci 2026) — the experiment behind the cyborgs, centaurs and the frontier.
  • The Cybernetic Teammate (P&G) — the study he cites for one person with AI matching a two-person team.
  • HBS AI Institute, Frontier Firm initiative — his own institute’s programme on the firms that “get it.”
  • Brynjolfsson, Chandar & Chen, Canaries in the Coal Mine — the strongest evidence in the wiki that early-career workers in exposed jobs are losing employment. Lakhani disagrees about where the risk sits: “I don’t think junior work is being automated … if anything, people that may be at risk are more in the middle management layers.” His evidence is usage data and observation, not employment data, and he grants that “the causal density is high” for the soft early-career market.
  • Goel (BCG, The So What) — distributed de-skilling, the firm-level version of the self-automator problem.

What was actually ingested

The full episode (51:35, length_seconds: 3095) from the auto-generated transcript, 505 lines. ASR corrected in quotes and prose (Kareem Lacani is Karim Lakhani; self-utomators and self-mators are self-automators; century at 2:16 is centaurs; Fabrizio Delawqua is Dell’Acqua; Steven Wilfr is Stephen Wolfram; Rem Coning is Rem Koning; laboratory for animation science is the Laboratory for Innovation Science). The study percentages given by the host (60% cyborgs, 14% centaurs, 27% self-automators, of 244 consultants) are hers and were not checked against the paper. The sponsored segment (9:58 to 11:28) is left out. No stills: a two-person studio interview.

Linked entities and concepts

Dangling (single-source mention, deferred): Marina Mogilko (host; not in author:), Mitch Weiss, Rem Koning, Greg Brockman, Stephen Wolfram.

Scope and reliability

An academic speaking about his own research, on a creator channel with a sponsor read and a newsletter pitch. His experimental claims are backed by published studies already in the wiki. His claims about usage inequality, the K-curve and middle managers come from unpublished data (“we hope to release some studies on this soon”) and from observation; he says himself that there is no correlation to performance yet. The middle-manager claim is a view, offered against colleagues who worry about juniors.

Relationships

Supports

Contradicts

Graph View

Backlinks