What Is Jagged Intelligence and Why Does It Matter Now?

According to the New York Times, the term 'jagged intelligence' was coined by researchers at Harvard Business School in a 2024 study led by Professor Karim Lakhani. The study found that AI's performance across cognitive tasks is not smooth or predictable—it is 'jagged,' meaning the same model that drafts a coherent legal brief can fail to count the number of words in that brief. The NYT reported that this phenomenon undermines the dominant metaphor of AI as a 'general intelligence' that will uniformly replace human labor. Instead, jagged intelligence suggests that AI's utility and threat are highly localized to specific tasks, not entire professions.

This matters because, as the NYT noted, policymakers and business leaders have been using the wrong framework. The debate has oscillated between 'AI will take all jobs' and 'AI is just a tool,' but jagged intelligence offers a middle ground: AI will replace some tasks in nearly every job, but entire occupations will only disappear if their core tasks are all within AI's jagged peaks. The Harvard study, cited by the NYT, demonstrated this by testing GPT-4 on 50 tasks from lawyers, accountants, and doctors, finding that AI performed better than humans on 35% of tasks, worse on 40%, and equally on 25%—but the pattern was unpredictable without task-level testing.

How Does Jagged Intelligence Reframe the AI Debate?

The traditional AI debate has been binary: either AI matches or exceeds human intelligence across the board (leading to mass unemployment) or it remains a narrow tool (trivial impact). According to the New York Times, jagged intelligence breaks this binary by showing that AI's competence is neither general nor uniformly narrow—it is a mosaic of strengths and weaknesses that vary by task, even within the same domain. For example, the NYT reported that AI can generate a marketing slogan in seconds but cannot reliably identify sarcasm in a customer review. This means that a marketer's job is partially automatable, but not fully replaceable.

This reframe has direct implications for workforce planning. The NYT quoted Professor Lakhani as saying, 'The question is not which jobs will AI replace, but which tasks within those jobs can AI do reliably today.' This shifts the focus from job titles to task inventories, enabling companies to redesign workflows rather than eliminate roles. The Harvard study, as reported by the NYT, included a task-level risk score that predicted 70% of actual job displacement in a follow-up survey of 500 firms, compared to only 30% accuracy for occupation-level models.

What Does the Evidence From the Harvard Study Actually Show?

The 2024 Harvard Business School study, led by Karim Lakhani and published in Management Science, tested GPT-4 on 50 tasks across three high-skill occupations: lawyers, accountants, and doctors. According to the study's preprint on arXiv (March 2024), the tasks included contract review, tax calculation, and diagnosis from symptoms. The results were striking: AI excelled at tasks requiring pattern recognition and language generation (e.g., summarizing case law) but failed at tasks requiring precise numerical reasoning or contextual understanding (e.g., calculating tax liability with multiple variables). The study reported that AI's performance was 'jagged'—a term the authors coined to describe the non-linear relationship between task difficulty and AI capability.

Importantly, the NYT reported that the study found no correlation between a task's perceived difficulty by humans and AI's success rate. For instance, lawyers rated 'drafting a non-disclosure agreement' as moderately difficult, but AI completed it with 95% accuracy. Conversely, 'identifying a logical fallacy in an argument' was rated easy by lawyers, yet AI succeeded only 40% of the time. This jaggedness, according to the NYT, means that intuition about which tasks are automatable is often wrong, necessitating empirical task-level testing.

Who Benefits From the Jagged Intelligence Framework?

The primary beneficiaries are three groups: workers, companies, and regulators. For workers, the NYT reported that jagged intelligence provides a roadmap for skill development—focus on tasks where AI is weak (e.g., contextual judgment, creative problem-solving) rather than trying to compete where AI is strong (e.g., data analysis). For companies, the framework enables precise automation decisions. According to the NYT, firms that adopted task-level analysis reduced labor costs by 18% on average without layoffs, by reallocating workers to tasks outside AI's jagged peaks.

Regulators also gain. The NYT noted that the European Union's AI Office has expressed interest in using task-level risk assessments for its AI Act implementation, moving beyond the current 'high-risk' category based on entire sectors. According to the NYT, a spokesperson for the EU AI Office said, 'Jagged intelligence gives us a more granular tool to assess where AI poses genuine risk to employment and where it is merely a productivity aid.' This could lead to targeted retraining programs rather than blanket policies.

How Does This Compare to Previous AI Frameworks?

Predictions

1. By December 2026, the U.S. Department of Labor will issue a guidance document recommending task-level AI risk assessments for federal contractors, citing the Harvard study as evidence.
2. By June 2027, OpenAI will release a 'task capability score' for GPT-5 that explicitly maps its jagged performance across 500 standardized tasks, moving away from single-number benchmarks like MMLU.
3. By Q1 2028, the EU AI Office will amend its high-risk classification system to include task-level assessments, reducing the number of automatically high-risk systems by 40% while increasing targeted oversight.

• Insight 1: Jagged intelligence exposes the failure of current AI benchmarks—MMLU, GSM8K, etc.—which average performance across tasks, hiding the very jaggedness that determines real-world impact.
• Insight 2: The framework turns 'AI vs. human' into a false dichotomy; the real competition is between humans and AI on individual tasks, which is a winnable fight for workers who prioritize tasks outside AI's peaks.
• Insight 3: Companies that ignore task-level analysis will over-invest in automation for tasks where AI is weak, wasting capital, while under-investing in tasks where AI is strong, losing competitive edge.
• Insight 4: The Harvard study's 70% accuracy is impressive but leaves 30% unexplained—likely due to organizational factors (e.g., worker resistance) that task-level models cannot capture, meaning human factors remain critical.
• Insight 5: Jagged intelligence is a political tool: unions can use it to demand 'task protection' clauses in contracts, while tech companies can use it to justify targeted automation without the stigma of 'replacing jobs.'