The Chess Progression

Radiology is the clearest centaur field in medicine. Hundreds of FDA-cleared AI tools assist radiologists in image interpretation, workflow triage, and preliminary report drafting. AI frequently achieves diagnostic performance comparable to or surpassing human experts, including earlier detection of skin cancer and diabetic retinopathy. But 63% of radiologists worry about algorithmic bias, and liability when AI errs remains unresolved.

Pathology is entering the centaur phase more recently. Stanford's Nuclei.io platform learns from individual pathologists and adapts to their workflows. In initial testing, AI catches approximately 5% of cases that pathologists miss, while making fewer errors overall. The key innovation: pathologists can share models like a professional social network.

Drug Discovery represents a dramatic acceleration. Insilico Medicine's AI-designed drug candidate progressed from target identification to preclinical candidate in under 18 months -- a process traditionally taking 3-4 years. In June 2025, Insilico published the industry's first proof-of-concept clinical validation of AI-driven drug discovery in Nature Medicine. Zasocitinib, designed with AI, is positioning to become potentially the first FDA-approved AI-discovered drug.

Trajectory: Medicine is firmly in the centaur phase, but with an important caveat -- regulation, liability, and the requirement for "meaningful human oversight" under the EU AI Act may artificially extend this phase well beyond what technological capability alone would dictate. The centaur phase in medicine may last longer than in chess not because AI can't do better alone, but because society won't let it.

Finance is the field closest to chess's Phase 3. AI-driven algorithms now facilitate 89% of global trading volume, up from ~60% in the early 2020s. The progression mirrors chess almost exactly: human traders dominated, then quant-human hybrids like Renaissance Technologies' Medallion Fund emerged (returning 63.3% annually from 1988-2018), and now pure algorithmic systems handle the vast majority of market activity.

However, unlike chess, financial markets have a critical asymmetry: they are adversarial with non-stationary rules. Markets change because participants change their strategies. AI algorithms often have safety triggers that cause de-risking during high volatility, potentially creating destabilizing feedback loops. The IMF warns this could make markets "more efficient and more volatile" simultaneously.

Where humans still matter: "Unknown unknowns" -- unprecedented events, macro-economic interpretation, regulatory shifts, and ethical judgment. As one analyst noted: "Algorithms struggle with unprecedented events, but humans still have the edge in unknown unknowns."

Trajectory: High-frequency and quantitative trading are effectively in Phase 3. Portfolio strategy, risk management, and client relations remain in Phase 2. The split may persist because markets are reflexive -- unlike chess, the "opponent" changes its strategy based on what you do.

Law entered 2026 with AI moving from "interesting tool" to "operational infrastructure." Tools like Luminance, Harvey, and Kira Systems can review thousands of contract pages in the time it takes a paralegal to handle fifty. McKinsey estimates 22% of a lawyer's job can be automated today, with 44% technically automatable.

The centaur model is clear: AI handles document review, legal research, contract analytics, and e-discovery. Humans handle courtroom advocacy, client counseling, ethical judgment, and novel legal reasoning. Analysts predict zero-touch contracting for low-risk agreements and AI redlining tools achieving 95% accuracy.

The junior lawyer problem: Entry-level roles are being reshaped. The work that traditionally trained junior associates -- document review, basic research, memo drafting -- is increasingly AI-handled. Some firms report that the junior associate is becoming an "AI Supervisor," reviewing AI-generated work rather than creating it from scratch. A Stanford study found hiring for entry-level, AI-impacted roles like junior legal work fell 16% over two years.

Trajectory: Law's centaur phase may be extended by regulatory and ethical requirements -- courtroom advocacy, fiduciary duty, and attorney-client privilege create structural barriers to full automation. But back-office legal work (contract management, compliance review, due diligence) is accelerating toward Phase 3.

Software engineering is the field Dario Amodei explicitly called out as being in its centaur phase: "I think we're already in our centaur phase for software, and during that centaur phase, if anything, the demand for software engineers may go up. But the period may be very brief."

The market has evolved dramatically from autocomplete to autonomous agents. Tools like Claude Code, Codex, Cursor, and GitHub Copilot now understand entire repositories, make multi-file changes, run tests, and iterate with minimal human input. Nearly 50% of all AI agent activity in 2026 is concentrated in software engineering.

However, real-world gains are more modest than early hype suggested: 20-30% productivity improvements concentrated in specific workflows, not the "10x productivity miracle" promised in 2023. The gap between AI-generated code and production-ready code remains significant for complex systems.

The 2005 parallel: Today's best results come from experienced engineers effectively directing AI agents -- echoing the 2005 freestyle tournament where amateurs with "better process" beat grandmasters. Senior engineers who know how to decompose problems, specify requirements precisely, and validate outputs are the new centaurs.

Trajectory: Amodei warns this phase could be "very brief" and that entry-level disruption is "happening over low single-digit numbers of years." The progression from code completion to autonomous agents mirrors chess's progression from calculation assistance to full game play.

Creative fields represent the most contested application of the chess model. The core debate: does the chess progression apply at all when the "game" has no objectively measurable winner?

In chess, superiority is unambiguous -- Elo ratings, win/loss records, engine analysis. In art, music, and writing, "better" is subjective, culturally contingent, and emotionally mediated. This may fundamentally alter the progression pattern.

Evidence for the chess model applying: AI-generated music has attracted large audiences on streaming platforms. Spotify removed over 75 million AI-generated "spammy tracks" in 2025. Generative AI produces logos, marketing copy, and product descriptions at commercial quality. Google DeepMind demonstrated AI creativity through chess puzzle composition, suggesting creative capacity extends beyond rule-following.

Evidence against the chess model: Research consistently finds that when audiences learn content was AI-generated, they rate it as less sincere and morally credible -- even when the wording is identical to human output. "The human touches and lived human experiences that shaped traditional art forms are irreplaceable." Human creativity shows the ability to continually push artistic boundaries in ways AI -- limited by standardized training data -- cannot replicate.

Trajectory: Creative fields may follow a modified chess model -- AI handles production-level content (marketing copy, stock imagery, background music) while human creators are valued for authentic voice, cultural meaning, and emotional resonance. This could create a permanent bifurcation rather than a clean Phase 3 transition. The "game" is different because the audience cares who (or what) made it.

Science presents the most dramatic recent example of AI Phase 3 breakthrough: AlphaFold solved the 50-year-old protein structure prediction problem, earning a Nobel Prize and surpassing the best physics-based tools. AlphaFold 3 is 50% more accurate than the best traditional methods on the PoseBusters benchmark.

The impact has been transformative. Researchers using AlphaFold 2 see a 40%+ increase in novel experimental protein structure submissions. AlphaFold-linked research is twice as likely to be cited in clinical articles and significantly more likely to be cited by patents.

Drug discovery acceleration: AI compresses early discovery timelines by 30-40% and reduces preclinical candidate development to 13-18 months versus the traditional 3-4 years. As of early 2026, over 173 AI-discovered drug programs are in clinical development. AI-discovered drugs show 65-75% success rates compared to 30-45% for traditional approaches.

Materials science: AlphaFold 3's capability to predict small molecule structures paves the way for novel biomaterials, bioengineered enzymes, and agricultural applications. The underlying approach -- training AI on experimental data to predict physical properties -- is generalizing across scientific domains.

Trajectory: Specific prediction tasks (protein folding, molecular docking) have reached Phase 3. Broader scientific reasoning -- hypothesis generation, experimental design, interpretation of novel phenomena -- remains firmly in the centaur phase. The research paper "When the Centaur Fell" argues that AI conquered chess but cannot yet replace human intuition in fields like cosmology where the problem space is not well-defined.

The military domain provides one of the most visceral demonstrations of the chess progression. In DARPA's AlphaDogfight Trials (2020), an AI agent defeated a human F-16 pilot 5-0 in simulated dogfights, making decisions in nanoseconds versus the human's tenths-of-second reaction time.

By 2024, this moved from simulation to reality: an AI-controlled F-16 (the X-62A VISTA) engaged a manned F-16 in within-visual-range air combat, executing maneuvers at 550+ mph. Australia's autonomous Ghost Bat combat aircraft completed over 100 test flights by 2025.

The "Minotaur" concept has emerged as the military's version of centaur chess -- a hybrid where human and AI forces work together, with pilots able to adjust autonomy levels based on fatigue, mission complexity, and threat assessment. Collaborative Combat Aircraft (CCAs) are designed as AI "wingmen" to manned fighters, not replacements.

Wargaming transformation: NATO's Frontlines of Rasputitsa evaluates LLM capabilities in rule adjudication, player assistance, and autonomous faction management within structured military wargames. AI enables every military student to practice planning with coalition partners against adversarial AI replicating enemy doctrine.

Trajectory: Tactical execution (dogfighting, target identification, route optimization) is approaching Phase 3. Strategic decision-making -- with its moral dimensions, political constraints, and rules of engagement -- will likely remain in Phase 2 for decades due to both capability limitations and deliberate policy choices about autonomous lethal force.

Accounting shows the chess progression clearly. Basis demonstrated the first AI agent to complete an end-to-end 1065 tax return autonomously, while CPA firms using the tool report 30-50% efficiency gains. Tax preparation has become one of the clearest centaur fields: AI ingests source documents, applies prior-year context, and prepares draft returns ready for professional review.

The entry-level impact is already measurable: a Stanford study found hiring for junior accounting roles fell 16% in approximately two years. Today's junior accountant is becoming an "AI Supervisor" -- reviewing AI-generated drafts, handling exceptions, and managing data security rather than performing the routine work directly.

Trajectory: Routine tax preparation and bookkeeping are approaching Phase 3. Advisory services, complex tax strategy, and audit judgment remain firmly in Phase 2. The bifurcation mirrors what happened in chess: tactical calculation was automated first, while strategic evaluation lasted longer.

Education exemplifies the centaur ideal. Khan Academy's Khanmigo, available for $4/month, provides Socratic AI tutoring that guides learners to find answers themselves rather than simply providing them. This design choice -- AI as coach, not oracle -- mirrors the centaur chess philosophy.

Microsoft partnered with Khan Academy in 2024 to scale Khanmigo globally, and by 2025 the platform launched in multiple languages across North America, Europe, Asia, and Latin America. Custom math AI models further improve tutoring quality.

The chess parallel is striking: just as amateur chess players with good process beat grandmasters, students with effective AI tutoring tools may outperform those with access only to expert human instruction. The democratization potential is enormous.

Trajectory: Education may represent a field where the centaur phase is genuinely durable. Learning requires motivation, emotional support, social development, and mentorship -- capabilities where human presence remains essential. AI tutoring will likely augment rather than replace human teachers for the foreseeable future, making education one of the strongest counter-examples to the chess inevitability thesis.

MIT Sloan researchers identified five human capability groups that show employment growth even as AI advances -- suggesting they represent genuinely durable advantages, not temporary ones:

The key insight: chess has zero EPOCH requirements. It is a closed, deterministic system with perfect information and an objective scoring function. Fields with high EPOCH demands -- therapy, leadership, crisis management, social work -- operate in fundamentally different problem spaces.

Several fields have structural features that chess lacks, which may permanently arrest the progression at Phase 2:

Regulatory lock-in: The EU AI Act classifies medical AI as "high-risk," requiring documented training data curation, bias checks, and mandatory human oversight. Even if AI diagnostics outperform humans, regulations may require human sign-off indefinitely.

Liability gaps: 63% of radiologists worry about who holds legal liability when AI errs. In law, fiduciary duty and attorney-client privilege create structural barriers. In medicine, "the doctor did it" provides a clear liability chain; "the algorithm did it" does not.

Moral agency requirements: Military doctrine, medical ethics, and legal practice all require a morally responsible agent who can be held accountable. AI systems lack moral reasoning, self-reflection, and the capacity for what researchers call "moral participation." You can punish a negligent doctor; you cannot punish an algorithm.

Social choice: Society may simply choose to keep humans in the loop for certain decisions regardless of AI capability -- just as we could automate judicial sentencing but choose not to. This is a values question, not a capabilities question.

1. Closed vs. Open Systems. Chess has fixed rules, a finite board, and perfect information. Real-world domains operate with incomplete information, shifting rules, and unbounded possibility spaces. Markets change because participants adapt. Diseases mutate. Geopolitics is reflexive.

2. Objective vs. Subjective Evaluation. Chess has an unambiguous scoring function. Creative work, leadership, counseling, and education do not. "Better" depends on context, values, and stakeholder perspective.

3. Single-Agent vs. Multi-Stakeholder. Chess involves two players. Real-world decisions involve patients, families, communities, shareholders, regulators, and the public. The social and political dimensions of many decisions are not reducible to optimization.

4. Deterministic vs. Embodied. Moravec's Paradox reminds us that sensorimotor skills -- walking, grasping, navigating physical spaces -- are exponentially harder for AI than intellectual tasks like chess. Fields requiring physical presence (surgery, emergency response, skilled trades) face fundamentally different automation timelines.