Physicists
19-2012.00Median wage $172,250/yr20,430 employed (US)Rank #299 of 923 scored · top 32% by substitution
Conduct research into physical phenomena, develop theories on the basis of observation and experiments, and devise methods to apply physical laws and theories.
Sub-scores
0–100 · band = confidence interval from rater disagreement
Substitution34
Exposure30
Augmentation75
Substitution — the headline: capability discounted by cost, barriers and adoption.
Exposure — technical capability alone, regardless of whether anyone deploys it.
Augmentation — how much AI assists without replacing. High here + moderate substitution = a changing job, not a disappearing one.
Tasks on the substitution scale
10 rated tasks, binned by substitution score.
Position among all scored occupations
Distribution of 923 occupation scores; the marker is this occupation.
Tasks with substitution ≥ 70
10%
Run 1.0.0-draft.1 · computed 2026-08-05 · rater panel: claude-sonnet-5, claude-haiku-4-5-20251001 · intervals span rater disagreement.
Why this score
The five weighted dimensions of the composite, averaged across this occupation's tasks (importance-weighted, panel mean). Exact weights and formulas: /api/v1/methodology.
Task automatabilityw 35%30
panel mean rating 2.2/5 → substitution pressure 30/100
Technical feasibility todayw 20%32
panel mean rating 2.3/5 → substitution pressure 32/100
Cost vs. human wagew 15%31
panel mean rating 2.2/5 → substitution pressure 31/100
Adoption barriersw 20%inverted — strong barriers lower the score44
panel mean rating 3.2/5 (barrier strength) → substitution pressure 44/100
Sector adoption velocityw 10%34
panel mean rating 2.4/5 → substitution pressure 34/100
Task breakdown (10 tasks)
Substitution pressure per task, weighted by O*NET importance in the composite. Expand a task for the full rater audit trail — every rating, every model, every rationale.
Perform complex calculations as part of the analysis and evaluation of data, using computers.
77CI 72–81 · exposure 75 · augmentation 100 · importance 4.3/5 · click for rater detail
Perform complex calculations as part of the analysis and evaluation of data, using computers.
77| Dimension | Model | Rating | Rationale |
|---|---|---|---|
| Sector adoption velocity | claude-haiku-4-5-20251001 | 5/5 | Physics, computational science, and research institutions have adopted automated calculation frameworks deeply and early. High-performance computing clusters, cloud platforms, and numerical software are standard in research groups and industry physics labs. |
| Sector adoption velocity | claude-sonnet-5 | 3/5 | Academic and research physics environments are moderate adopters—computational tools and AI-assisted coding are increasingly used, but full trust in end-to-end automated analysis pipelines remains cautious and pilot-stage in many labs. |
| Augmentation potential | claude-haiku-4-5-20251001 | 5/5 | AI systems (interactive Jupyter notebooks, live symbolic solvers, numerical optimization assistants, visualization aids) substantially augment physicists' productivity by enabling rapid iteration, exploration, and error-checking while the physicist remains in the loop for interpretation and method selection. |
| Augmentation potential | claude-sonnet-5 | 5/5 | AI substantially augments physicists by accelerating code writing, debugging, statistical analysis, and exploratory data interpretation while the physicist retains judgment over methodology and interpretation. |
| Task automatability | claude-haiku-4-5-20251001 | 4/5 | Current AI systems (symbolic engines, numerical solvers, ML-based approximators) can automate the majority of complex calculations used in physics data analysis, achieving substantial time savings. However, task setup, selection of methods, and validation of results often require physicist judgment, preventing full end-to-end automation in all contexts. |
| Task automatability | claude-sonnet-5 | 4/5 | Computational analysis and complex calculations are already heavily done via software/scripts, and AI can generate, debug, and run analysis code and interpret numeric outputs with significant time savings for many standard analyses.atable |
| Adoption barriers | claude-haiku-4-5-20251001 | 2/5 | Few regulatory or legal barriers exist to automating physics calculations; institutional practices favor computation. The main friction is organizational (physicists' preference for hands-on verification) and epistemic (confidence in novel AI methods), both surmountable through adoption. |
| Adoption barriers | claude-sonnet-5 | 2/5 | No licensing requirement blocks computational analysis; the main friction is scientific validity/peer review and the need for expert judgment to interpret results correctly, not legal or regulatory barriers. |
| Cost vs. human wage | claude-haiku-4-5-20251001 | 5/5 | Computational costs (cloud instances, software licenses) for performing calculations are orders of magnitude cheaper than physicist labor ($100K+ annual salary). A single GPU or CPU performing calculations costs pennies per task relative to human wage equivalents. |
| Cost vs. human wage | claude-sonnet-5 | 4/5 | Once physics-specific code/workflows are set up, AI-assisted computation is far cheaper per calculation than a physicist's time, though initial setup and domain-specific validation add cost. |
| Technical feasibility today | claude-haiku-4-5-20251001 | 4/5 | Mature tools like Mathematica, MATLAB, Python libraries (NumPy, SciPy), and physics-specialized AI systems reliably perform complex calculations in production. Cloud-based scientific computing platforms are widely deployed; the limitation is mostly in selecting the right approach rather than execution reliability. |
| Technical feasibility today | claude-sonnet-5 | 4/5 | Deployed tools (Python/R/Mathematica integrations, code-generation assistants, numerical libraries) reliably perform complex calculations and data analysis in production research settings today, though novel physics-specific pipelines still need expert oversight. |
Analyze data from research conducted to detect and measure physical phenomena.
42CI 30–55 · exposure 42 · augmentation 88 · importance 4.2/5 · click for rater detail
Analyze data from research conducted to detect and measure physical phenomena.
42| Dimension | Model | Rating | Rationale |
|---|---|---|---|
| Sector adoption velocity | claude-haiku-4-5-20251001 | 2/5 | While academic institutions are piloting AI for data processing, actual displacement in physics research remains minimal; most labs use bespoke, custom analysis pipelines and trust human expertise for interpretation, limiting mainstream AI adoption. |
| Sector adoption velocity | claude-sonnet-5 | 3/5 | Physics research groups increasingly use ML/AI for data analysis (e.g., LHC experiments, astrophysics surveys), representing steady but uneven adoption depending on subfield and funding for computational infrastructure. |
| Augmentation potential | claude-haiku-4-5-20251001 | 4/5 | AI significantly assists physicists through automated data reduction, anomaly highlighting, visualization, and hypothesis suggestion, allowing researchers to focus on theory and interpretation; many research groups now rely on AI-augmented workflows for efficiency. |
| Augmentation potential | claude-sonnet-5 | 5/5 | AI substantially accelerates data processing, pattern recognition, and statistical analysis in physics research, letting scientists focus on interpretation and hypothesis generation while remaining central to the work. |
| Task automatability | claude-haiku-4-5-20251001 | 2/5 | AI can automate routine signal processing, statistical analysis, and data visualization, but cannot independently design experiments, interpret anomalies, or validate physical claims against theory—tasks requiring expert judgment and domain knowledge that define the core of physics research. |
| Task automatability | claude-sonnet-5 | 3/5 | AI/ML tools can perform substantial portions of data analysis (fitting, anomaly detection, statistical inference) but require domain-specific setup, validation, and physicist judgment to interpret novel phenomena, so only partial end-to-end automation is achievable today. |
| Adoption barriers | claude-haiku-4-5-20251001 | 4/5 | Physics research has institutional and epistemic barriers: peer review and publication gatekeeping require credentialed human researchers to sign off; liability for incorrect physical claims falls on human experts; funding bodies expect human principal investigators responsible for results. |
| Adoption barriers | claude-sonnet-5 | 2/5 | No licensing requirement for data analysis itself, but peer review, scientific credibility, and reproducibility norms create moderate friction against fully AI-driven conclusions in published research. |
| Cost vs. human wage | claude-haiku-4-5-20251001 | 2/5 | AI inference for data processing is cheap, but the high-value interpretive and validation work remains human-dependent; the net cost reduction is marginal relative to physicist salaries when oversight and validation are included. |
| Cost vs. human wage | claude-sonnet-5 | 3/5 | Computational analysis can be cheap once pipelines are built, but the engineering, domain customization, and validation overhead for research-grade physics data keeps costs roughly comparable to skilled labor for novel research tasks. |
| Technical feasibility today | claude-haiku-4-5-20251001 | 3/5 | Production systems exist for standard data processing (curve fitting, spectral analysis, noise reduction), but interpretation of novel phenomena and detection of systematic errors still require human physicists; no end-to-end product handles the full analytical pipeline reliably. |
| Technical feasibility today | claude-sonnet-5 | 3/5 | Deployed tools (e.g., ML pipelines in particle physics, astronomy pipelines, general-purpose data analysis copilots) reliably assist with specific analysis subtasks, but no product handles the full breadth of physics data analysis across novel experimental setups reliably. |
Write research proposals to receive funding.
34CI 25–43 · exposure 33 · augmentation 63 · importance 4.0/5 · click for rater detail
Write research proposals to receive funding.
34| Dimension | Model | Rating | Rationale |
|---|---|---|---|
| Sector adoption velocity | claude-haiku-4-5-20251001 | 2/5 | Academic and research institutions adopt AI writing assistants slowly; proposal writing remains a high-stakes, heavily scrutinized process with strong human-in-the-loop expectations. Uptake lags information-sector automation patterns. |
| Sector adoption velocity | claude-sonnet-5 | 2/5 | Academic and research science sectors have been slower and more cautious in adopting AI writing tools for high-stakes funding documents due to norms around originality, plagiarism policies, and funder scrutiny. |
| Augmentation potential | claude-haiku-4-5-20251001 | 3/5 | Current AI assists meaningfully in drafting boilerplate sections, literature synthesis, and editorial polish, raising physicist productivity in proposal mechanics. However, the core novelty and strategic positioning remain human-driven, limiting transformative augmentation. |
| Augmentation potential | claude-sonnet-5 | 4/5 | AI tools are already widely used to help physicists draft, edit, and polish proposal text, generate boilerplate sections, and improve clarity, meaningfully speeding up the writing process while the scientist retains control over content and strategy. |
| Task automatability | claude-haiku-4-5-20251001 | 2/5 | AI can draft sections of proposals (background, methodology outlines) but cannot independently conceive novel research ideas, assess scientific merit, or navigate funder-specific requirements and strategic positioning that expert physicists evaluate. The creative and judgment-heavy core remains human. |
| Task automatability | claude-sonnet-5 | 3/5 | AI can draft substantial portions of a research proposal (background, literature review, budget narrative boilerplate) but the core scientific contribution, novelty framing, and strategic alignment with funder priorities require expert judgment that current tools cannot fully replicate. |
| Adoption barriers | claude-haiku-4-5-20251001 | 4/5 | Funding agencies expect human authorship and accountability; many require PI signatures and review; institutional gatekeeping (grant office oversight) and disciplinary norms around original intellectual contribution create meaningful friction against full AI automation, even where technical capability might increase. |
| Adoption barriers | claude-sonnet-5 | 3/5 | There's no formal licensing requirement, but funding agencies expect the named PI's original scientific reasoning and accountability, and reviewers/institutions scrutinize authorship and originality, creating moderate friction against full automation. |
| Cost vs. human wage | claude-haiku-4-5-20251001 | 2/5 | A physicist's opportunity cost for proposal writing is high; AI tools (subscription + oversight) may reduce labor hours modestly but do not approach order-of-magnitude savings when factoring in quality assurance and expert review needed to produce competitive proposals. |
| Cost vs. human wage | claude-sonnet-5 | 3/5 | Using AI to draft sections is cheap relative to a physicist's time, but the human must still invest significant time reviewing, correcting technical content, and ensuring scientific rigor, so net savings are moderate rather than order-of-magnitude. |
| Technical feasibility today | claude-haiku-4-5-20251001 | 2/5 | No deployed system reliably generates fundable research proposals end-to-end; AI writing tools exist but require heavy expert revision and frequently miss disciplinary nuance, originality expectations, and funder priorities. Products can assist but cannot replace the human physicist's decision-making. |
| Technical feasibility today | claude-sonnet-5 | 2/5 | General LLM writing assistants are used informally by researchers to draft sections, but no deployed product reliably produces fundable, technically sound physics research proposals without heavy expert revision. |
Report experimental results by writing papers for scientific journals or by presenting information at scientific conferences.
33CI 16–50 · exposure 30 · augmentation 75 · importance 3.9/5 · click for rater detail
Report experimental results by writing papers for scientific journals or by presenting information at scientific conferences.
33| Dimension | Model | Rating | Rationale |
|---|---|---|---|
| Sector adoption velocity | claude-haiku-4-5-20251001 | 2/5 | Physics and academia lag in AI adoption for core research tasks; while some researchers use AI for drafting assistance, institutional policies on AI in authorship remain restrictive and evolving, slowing deep production adoption. |
| Sector adoption velocity | claude-sonnet-5 | 3/5 | Academic and research fields show growing but uneven adoption of AI writing tools, with many researchers experimenting but institutional and publisher norms still evolving with caution. |
| Augmentation potential | claude-haiku-4-5-20251001 | 3/5 | AI can assist with outlining, drafting initial text, language refinement, figure annotation, and presentation slide generation, moderately improving physicist productivity without replacing their interpretive and novelty-judgment roles. |
| Augmentation potential | claude-sonnet-5 | 5/5 | AI substantially boosts productivity in drafting, editing, formatting, literature summarization, and even slide preparation, while the physicist retains responsibility for scientific content and interpretation. |
| Task automatability | claude-haiku-4-5-20251001 | 1/5 | Writing scientific papers and presenting at conferences requires human interpretation of experimental results, novel insights, creative framing, and judgment about significance—tasks where AI cannot yet achieve 50% time savings at equal quality when the human standard is rigorous scientific communication. |
| Task automatability | claude-sonnet-5 | 3/5 | AI can draft substantial portions of scientific papers (methods, results summaries, literature framing) from data and notes, but synthesizing novel interpretation, ensuring scientific rigor, and final authorship judgment still require significant human involvement.' |
| Adoption barriers | claude-haiku-4-5-20251001 | 4/5 | Strong barriers exist: journal editors and peer reviewers expect human authorship and accountability; author liability and reputation stakes are high; institutional norms and funding bodies require verified human intellectual contribution; conferences expect live human presenters. |
| Adoption barriers | claude-sonnet-5 | 3/5 | Journals and conferences increasingly require disclosure of AI use and hold authors fully responsible for accuracy, creating moderate friction, though no formal licensure blocks AI-assisted drafting. |
| Cost vs. human wage | claude-haiku-4-5-20251001 | 2/5 | Current AI systems can reduce writing time on some components (drafting, editing), making them cheaper per word, but the bottleneck is scientific judgment and validation, which remain human-intensive; overall cost advantage is modest. |
| Cost vs. human wage | claude-sonnet-5 | 3/5 | AI drafting tools are cheap per use, but the physicist's time for data validation, figure creation, and scientific judgment remains the dominant cost, keeping overall cost roughly comparable to unaided human effort when quality-checking is included. |
| Technical feasibility today | claude-haiku-4-5-20251001 | 2/5 | While AI can draft sections (methods, basic results summaries) and assist with language polishing, no deployed product reliably produces publication-ready physics papers or effective conference presentations end-to-end; human scientists remain gatekeepers of correctness and novelty. |
| Technical feasibility today | claude-sonnet-5 | 3/5 | LLM-based writing assistants (e.g., ChatGPT, Overleaf/Writefull integrations) are widely used by researchers for drafting and editing papers, but no product reliably handles full paper authorship or conference presentation creation without heavy human revision. |
Design computer simulations to model physical data so that it can be better understood.
31CI 25–38 · exposure 25 · augmentation 75 · importance 4.0/5 · click for rater detail
Design computer simulations to model physical data so that it can be better understood.
31| Dimension | Model | Rating | Rationale |
|---|---|---|---|
| Sector adoption velocity | claude-haiku-4-5-20251001 | 2/5 | Physics and academic research are digitized but conservative in adopting automation for core intellectual work. Adoption of AI coding assistants is growing, but full automation of simulation design remains rare; most usage is augmentation in specialized research groups. |
| Sector adoption velocity | claude-sonnet-5 | 3/5 | Research and academic physics settings are adopting AI coding tools and ML-based modeling at a moderate pace, with pilots more common than fully embedded production workflows. |
| Augmentation potential | claude-haiku-4-5-20251001 | 4/5 | AI coding assistants and large language models meaningfully accelerate simulation code writing, parameter exploration, and documentation tasks, allowing physicists to focus on model selection and validation. This is a mature augmentation pattern in research institutions. |
| Augmentation potential | claude-sonnet-5 | 4/5 | AI tools substantially assist by generating boilerplate code, suggesting numerical methods, debugging, and exploring parameter spaces, meaningfully speeding up simulation design work. |
| Task automatability | claude-haiku-4-5-20251001 | 2/5 | While AI can assist with code generation and parameter optimization for simulations, designing simulations requires domain expertise, problem formulation, and judgment about appropriate models—tasks that demand human direction. AI cannot autonomously decide what physics to model or validate that a simulation design is scientifically sound. |
| Task automatability | claude-sonnet-5 | 2/5 | Designing simulations requires deep domain expertise, model formulation, and validation against physical theory that current AI cannot reliably do end-to-end without expert direction.rating. |
| Adoption barriers | claude-haiku-4-5-20251001 | 4/5 | Physics simulations in research and industrial contexts often require peer review, publication credibility, and scientist accountability for correctness. Liability and accuracy concerns—particularly in domains like nuclear or climate modeling—create strong organizational and professional incentives to keep humans as decision-makers. |
| Adoption barriers | claude-sonnet-5 | 2/5 | No licensing requirement blocks AI-assisted simulation design, but scientific rigor, peer review, and reproducibility norms create moderate organizational and epistemic friction. |
| Cost vs. human wage | claude-haiku-4-5-20251001 | 2/5 | Inference costs for LLMs are low, but simulation design requires iterative refinement, validation, and expert oversight that cannot be offloaded to AI alone. The total cost of human expert time remains dominant, making AI a complementary expense rather than a cost-reducer. |
| Cost vs. human wage | claude-sonnet-5 | 2/5 | Physicist labor for simulation design is expensive, but AI still requires substantial expert oversight and iteration, limiting cost savings to modest levels rather than order-of-magnitude. |
| Technical feasibility today | claude-haiku-4-5-20251001 | 2/5 | Code-generation tools (GitHub Copilot, Claude) can produce simulation code snippets, but no deployed product reliably handles the full task of designing a physics simulation from raw experimental data to validated model. This remains largely a human-led workflow with AI as a coding assistant. |
| Technical feasibility today | claude-sonnet-5 | 2/5 | AI coding assistants can help write simulation code snippets, but no deployed product autonomously designs full physics simulations reliably in production. |
Describe and express observations and conclusions in mathematical terms.
30CI 25–35 · exposure 25 · augmentation 63 · importance 4.0/5 · click for rater detail
Describe and express observations and conclusions in mathematical terms.
30| Dimension | Model | Rating | Rationale |
|---|---|---|---|
| Sector adoption velocity | claude-haiku-4-5-20251001 | 2/5 | Physics is a knowledge-intensive, high-expertise sector moving cautiously on automation. Adoption of AI for formalization is limited to assistive drafting in labs; no significant displacement of this task has occurred. Physicists remain skeptical of outsourcing mathematical interpretation to unvalidated systems. |
| Sector adoption velocity | claude-sonnet-5 | 2/5 | Academic and research physics adopts AI tools cautiously and unevenly; use is largely limited to coding/analysis assistance rather than core theoretical formalization, reflecting slower adoption than fast-moving digital sectors. |
| Augmentation potential | claude-haiku-4-5-20251001 | 3/5 | AI can assist by suggesting mathematical notation, checking dimensional consistency, or proposing functional forms—useful productivity gains for routine formalization. However, the physicist must still validate and often redesign the representation, limiting transformative impact compared to purely assistive drafting tasks. |
| Augmentation potential | claude-sonnet-5 | 4/5 | AI tools (symbolic computation, LLM-assisted derivation, code generation for simulations) meaningfully speed up expressing observations mathematically, letting physicists iterate faster while retaining judgment over validity and interpretation. |
| Task automatability | claude-haiku-4-5-20251001 | 2/5 | While AI can help formalize observations mathematically, the task requires deep domain insight into *which* mathematical framework is appropriate and how to interpret physical phenomena meaningfully. Current AI struggles with the creative, context-dependent selection of mathematical representations and cannot reliably perform the full task end-to-end without substantial physicist guidance. |
| Task automatability | claude-sonnet-5 | 2/5 | AI can assist in drafting mathematical formalizations and notation from descriptions, but translating novel physical observations into correct, insightful mathematical conclusions requires deep domain judgment that current systems cannot reliably perform end-to-end.4 |
| Adoption barriers | claude-haiku-4-5-20251001 | 4/5 | Physics research and publication carry high error-cost asymmetry: incorrect mathematical descriptions of observations can waste years of downstream work and damage credibility. Peer review, institutional validation, and the researcher's professional liability create strong friction against automated formulation without expert sign-off. |
| Adoption barriers | claude-sonnet-5 | 2/5 | No licensing barrier exists, but physicists' work is subject to peer review, reproducibility norms, and institutional expectations of researcher authorship, creating moderate friction against pure automation. |
| Cost vs. human wage | claude-haiku-4-5-20251001 | 2/5 | AI inference is cheap, but the integration cost is high: a physicist must review and often substantially revise the mathematical formulation, negating cost savings. The human remains the rate-limiting step, and oversight burden means all-in cost remains comparable to or exceeds unassisted expert work. |
| Cost vs. human wage | claude-sonnet-5 | 2/5 | AI assistance is cheap per query, but the human oversight, verification, and correction needed for scientifically valid formalization keeps effective cost comparable to or only modestly less than a trained physicist's time. |
| Technical feasibility today | claude-haiku-4-5-20251001 | 2/5 | No deployed product reliably performs this task independently. Large language models can assist with symbolic manipulation and can format results mathematically, but they lack the physical intuition to correctly describe novel observations or validate whether a mathematical expression actually captures the underlying physics. Production systems do not do this reliably without expert oversight. |
| Technical feasibility today | claude-sonnet-5 | 2/5 | Tools like LLMs with symbolic math capabilities can help formalize known relationships, but no deployed product reliably performs original scientific formalization of novel observations at production quality. |
Teach physics to students.
28CI 25–30 · exposure 30 · augmentation 75 · importance 4.0/5 · click for rater detail
Teach physics to students.
28| Dimension | Model | Rating | Rationale |
|---|---|---|---|
| Sector adoption velocity | claude-haiku-4-5-20251001 | 2/5 | Physics education in schools and universities remains highly traditional and resistant to automation; while some universities pilot AI-assisted homework and tutoring systems, classroom teaching by physicists remains a core institutional role with slow technological displacement in practice. |
| Sector adoption velocity | claude-sonnet-5 | 2/5 | Education sector adoption of AI is real but slow and uneven, with pilots for tutoring more common than deep integration into full teaching duties. |
| Augmentation potential | claude-haiku-4-5-20251001 | 4/5 | AI can meaningfully assist physicists by auto-generating problem sets, explaining student questions in multiple ways, providing real-time visualization suggestions, and flagging common misconceptions—enhancing instructor productivity and student engagement while keeping the physicist in the pedagogical driver seat. |
| Augmentation potential | claude-sonnet-5 | 4/5 | AI substantially aids lecture prep, explanation generation, problem sets, and personalized practice, meaningfully boosting a physics teacher's productivity while they remain the primary instructor. |
| Task automatability | claude-haiku-4-5-20251001 | 2/5 | AI can draft lecture notes, generate practice problems, and provide explanations of concepts, but teaching requires real-time interaction, adaptive feedback based on individual student understanding, and the ability to respond to novel questions and confusion—capabilities that current systems handle only fragmentarily and without the pedagogical coherence humans provide. |
| Task automatability | claude-sonnet-5 | 2/5 | Teaching involves live interaction, adaptive pedagogy, motivation, and mentorship that current AI cannot fully replicate end-to-end despite being able to generate explanations or practice problems.4 It can support parts of instruction but not replace the full teaching role reliably. |
| Adoption barriers | claude-haiku-4-5-20251001 | 4/5 | Teaching is deeply embedded in institutional accreditation, student-instructor interaction requirements, and organizational culture; institutions have strong preference for credentialed humans to deliver and sign off on curriculum, and parent/student demand for live instruction creates significant friction against AI substitution. |
| Adoption barriers | claude-sonnet-5 | 4/5 | Teaching in accredited institutions typically requires certified/licensed instructors, institutional accreditation and legal responsibility for grading and credentialing, creating strong structural barriers to full replacement. |
| Cost vs. human wage | claude-haiku-4-5-20251001 | 2/5 | Deploying AI tutoring systems with adequate oversight, content curation, and integration into institutional curricula is moderately expensive, while paying a physicist's salary remains substantial; the cost difference is marginal enough that organizations often prefer the credibility and human connection of live instruction. |
| Cost vs. human wage | claude-sonnet-5 | 2/5 | AI tutoring tools are cheap per interaction, but a full teaching role including assessment, mentorship, and accreditation still requires paid human labor, so overall cost parity is not achieved for the whole task. |
| Technical feasibility today | claude-haiku-4-5-20251001 | 3/5 | AI tutoring products and lecture-support tools exist and can explain physics concepts with reasonable accuracy, but they lack reliability in handling complex conceptual misconceptions, adapting to diverse learning styles, and managing classroom dynamics; most deployments are narrow or supplementary rather than replacing a physicist's teaching role. |
| Technical feasibility today | claude-sonnet-5 | 2/5 | AI tutoring products (e.g., adaptive learning platforms) exist and are used in some settings, but classroom-level physics teaching by an AI system without a human instructor is not a deployed norm. |
Develop theories and laws on the basis of observation and experiments, and apply these theories and laws to problems in areas such as nuclear energy, optics, and aerospace technology.
21CI 7–35 · exposure 13 · augmentation 75 · importance 3.7/5 · click for rater detail
Develop theories and laws on the basis of observation and experiments, and apply these theories and laws to problems in areas such as nuclear energy, optics, and aerospace technology.
21| Dimension | Model | Rating | Rationale |
|---|---|---|---|
| Sector adoption velocity | claude-haiku-4-5-20251001 | 2/5 | While computational tools and AI-assisted simulation are increasingly standard in physics labs, adoption of AI for autonomous theory development remains minimal and largely experimental. Most physics departments and research institutions continue to rely on human-led investigation; AI is a supporting tool rather than a replacement agent in production research workflows. |
| Sector adoption velocity | claude-sonnet-5 | 2/5 | Physics research is a highly specialized, low-volume field where AI adoption is mostly limited to computational modeling and data analysis pilots rather than theory generation itself. |
| Augmentation potential | claude-haiku-4-5-20251001 | 4/5 | AI substantially augments physicist productivity today through simulation acceleration, data analysis, literature mining, mathematical proof-checking, and hypothesis generation support. These tools can meaningfully reduce time spent on routine computations and pattern-finding, allowing physicists to focus on creative theorizing and experimental design while remaining firmly in the decision-making loop. |
| Augmentation potential | claude-sonnet-5 | 4/5 | AI significantly aids literature review, hypothesis generation, simulation, and mathematical derivation, meaningfully boosting physicist productivity while humans retain core theoretical judgment. |
| Task automatability | claude-haiku-4-5-20251001 | 2/5 | While AI can assist with data analysis, pattern recognition in experimental results, and mathematical formulation, the creative conceptualization of novel physical theories and laws fundamentally requires human insight, intuition, and judgment that current systems cannot replace end-to-end. AI cannot autonomously design experiments, interpret anomalies in ways that lead to paradigm shifts, or propose genuinely novel theoretical frameworks at the level required for significant scientific advancement. |
| Task automatability | claude-sonnet-5 | 1/5 | Original theory development requires deep creative insight, novel conceptual leaps, and integration of experimental context that current AI cannot perform end-to-end at equal quality with substantial time savings. |
| Adoption barriers | claude-haiku-4-5-20251001 | 4/5 | Strong institutional and professional barriers protect this task: peer review and publication requirements demand human authorship and accountability; funding agencies and academic institutions require human scientists as principal investigators; and liability for erroneous theories in applied domains (aerospace, nuclear) creates legal responsibility that falls on licensed physicists, not algorithms. |
| Adoption barriers | claude-sonnet-5 | 2/5 | No licensing requirement exists for theoretical work itself, but publication, peer review, and scientific credibility norms create moderate friction against AI-generated theories being accepted. |
| Cost vs. human wage | claude-haiku-4-5-20251001 | 2/5 | The computational and infrastructure costs for high-fidelity simulations, plus the human expertise required for oversight and validation, remain substantial. Physicist salaries are high, but the full end-to-end cost of AI-driven theory development (including failed hypotheses, retraining, validation overhead) does not yet achieve cost parity, let alone significant savings. |
| Cost vs. human wage | claude-sonnet-5 | 1/5 | Since AI cannot substitute for the core creative/scientific task, there is no meaningful cost-equivalent output to compare, making AI effectively unable to replace the human labor at any cost. |
| Technical feasibility today | claude-haiku-4-5-20251001 | 2/5 | Deployed AI products excel at narrow tasks like literature synthesis, simulation, or equation solving, but no production system reliably performs the full cycle of theory development from observation through experimental design to novel law formulation. Current tools are research-stage for the core creative task; they function as assistants rather than autonomous theory developers. |
| Technical feasibility today | claude-sonnet-5 | 1/5 | No deployed product autonomously formulates new physical theories and laws; existing AI tools are research-stage aids for narrow subtasks like literature search or symbolic manipulation. |
Observe the structure and properties of matter, and the transformation and propagation of energy, using equipment such as masers, lasers, and telescopes, to explore and identify the basic principles governing these phenomena.
17CI 7–26 · exposure 13 · augmentation 75 · importance 3.8/5 · click for rater detail
Observe the structure and properties of matter, and the transformation and propagation of energy, using equipment such as masers, lasers, and telescopes, to explore and identify the basic principles governing these phenomena.
17| Dimension | Model | Rating | Rationale |
|---|---|---|---|
| Sector adoption velocity | claude-haiku-4-5-20251001 | 2/5 | Physics research sectors are moderately digitized and beginning to adopt AI for data analysis, but adoption remains limited to assistive analytics rather than autonomous equipment operation. Most physics labs continue traditional human-led experimental design and execution; AI adoption is slower than in finance or information sectors. |
| Sector adoption velocity | claude-sonnet-5 | 2/5 | Academic and research physics labs adopt AI tools for data analysis but adoption of AI for hands-on experimental observation and equipment operation is slow and limited to niche automation pilots. |
| Augmentation potential | claude-haiku-4-5-20251001 | 4/5 | AI significantly augments physicists by automating data processing, pattern detection in large datasets, literature analysis, and hypothesis generation from observational data—allowing physicists to focus on experimental design, interpretation, and novel theoretical insights while remaining central to the discovery process. |
| Augmentation potential | claude-sonnet-5 | 4/5 | AI substantially aids physicists in data analysis, simulation, anomaly detection, and literature synthesis, meaningfully boosting productivity even though the physical observation itself remains human-led. |
| Task automatability | claude-haiku-4-5-20251001 | 2/5 | While AI can assist with data analysis and pattern recognition in observational data, the task requires skilled operation of complex scientific equipment (masers, lasers, telescopes) and real-time experimental judgment that demands human expertise. Current AI cannot autonomously design, calibrate, or troubleshoot experimental apparatus or make the nuanced decisions about what phenomena to observe next. |
| Task automatability | claude-sonnet-5 | 1/5 | This is core experimental physics work requiring hands-on operation of specialized equipment, hypothesis-driven observation, and real-time scientific judgment that current AI cannot perform end-to-end. |
| Adoption barriers | claude-haiku-4-5-20251001 | 4/5 | Significant barriers exist: Ph.D.-level scientific expertise and professional licensure are typically required; liability for equipment damage and experimental errors falls on responsible scientists; regulatory requirements govern certain types of research; and the fundamental nature of scientific discovery requires human judgment and intuition that organizations are reluctant to delegate. |
| Adoption barriers | claude-sonnet-5 | 3/5 | No licensing barrier per se, but experimental physics requires physical presence, equipment access, and institutional oversight (safety, grant accountability) that create moderate friction against automation. |
| Cost vs. human wage | claude-haiku-4-5-20251001 | 1/5 | The infrastructure, equipment maintenance, and specialized AI development required to automate observational physics would be prohibitively expensive compared to employing trained physicists who leverage AI as an analytical tool. The capital costs of scientific instruments far exceed typical AI inference costs. |
| Cost vs. human wage | claude-sonnet-5 | 1/5 | There is no AI system that substitutes for the physical apparatus operation and expert observation involved, so cost comparison favors the human scientist by default. |
| Technical feasibility today | claude-haiku-4-5-20251001 | 2/5 | Deployed AI systems can process and analyze observational data after collection, but no production system autonomously operates advanced scientific equipment or conducts end-to-end experimental exploration. Research tools exist for data interpretation, but the hands-on experimental and exploratory components remain beyond current deployable systems. |
| Technical feasibility today | claude-sonnet-5 | 1/5 | No deployed AI product operates lasers, masers, or telescopes autonomously to conduct novel physical experiments; this remains firmly research-stage and human-directed. |
Collaborate with other scientists in the design, development, and testing of experimental, industrial, or medical equipment, instrumentation, and procedures.
16CI 7–25 · exposure 13 · augmentation 63 · importance 3.5/5 · click for rater detail
Collaborate with other scientists in the design, development, and testing of experimental, industrial, or medical equipment, instrumentation, and procedures.
16| Dimension | Model | Rating | Rationale |
|---|---|---|---|
| Sector adoption velocity | claude-haiku-4-5-20251001 | 2/5 | Physics and experimental research move slowly on AI adoption. Most labs use traditional CAD, simulation software, and human expertise; AI-assisted design is still rare in production workflows. Digitization is partial, and conservatism around novel tools in high-stakes experimental environments slows adoption. |
| Sector adoption velocity | claude-sonnet-5 | 2/5 | Physical sciences and lab-based R&D adopt AI tools more slowly than purely digital fields, with pilots for simulation and data analysis but little deployment in physical equipment design/testing. |
| Augmentation potential | claude-haiku-4-5-20251001 | 3/5 | AI can meaningfully assist with literature synthesis, statistical analysis of experimental data, design optimization suggestions, and documentation. However, the core creative and experimental judgment remains human-driven, limiting augmentation to supporting tasks rather than transforming the full workflow. |
| Augmentation potential | claude-sonnet-5 | 4/5 | AI substantially aids simulation, design optimization, literature review, and data analysis supporting equipment development, meaningfully boosting productivity while humans still lead design and testing. |
| Task automatability | claude-haiku-4-5-20251001 | 2/5 | This task requires creative design decisions, hands-on experimentation, and judgment calls that depend on domain expertise and physical intuition. While AI can assist with literature review, computational modeling, or documentation, the core collaborative design and testing loop demands human scientists making trade-offs and iterating based on novel experimental outcomes. |
| Task automatability | claude-sonnet-5 | 1/5 | This is a highly collaborative, hands-on task requiring physical prototyping, creative scientific judgment, and interpersonal coordination that current AI cannot perform end-to-end.' |
| Adoption barriers | claude-haiku-4-5-20251001 | 4/5 | Institutional review boards, safety certifications for medical/industrial equipment, and legal liability for failures create hard barriers. Moreover, peer review, intellectual property ownership, and the need for a human scientist to sign off on designs and validate experimental procedures are strong structural protections. |
| Adoption barriers | claude-sonnet-5 | 4/5 | Medical and industrial equipment design often requires regulatory compliance, safety certification, and professional accountability, creating strong barriers to full automation. |
| Cost vs. human wage | claude-haiku-4-5-20251001 | 2/5 | AI inference is cheap, but the overhead of integration into a collaborative research environment, validation of design suggestions, and oversight by domain experts makes the all-in cost substantial relative to the marginal value added per task instance. Human scientists' labor dominates the cost structure. |
| Cost vs. human wage | claude-sonnet-5 | 1/5 | AI cannot substitute for the physical and collaborative labor involved, so there is no meaningful cost comparison—human physicists remain necessary and cheaper than any AI-plus-lab-equivalent alternative. |
| Technical feasibility today | claude-haiku-4-5-20251001 | 2/5 | No deployed product autonomously designs, develops, and tests experimental equipment end-to-end. AI tools exist for component design (CAD assistance) and data analysis, but the integrated collaborative design-build-test cycle remains human-led; AI products are narrow assistants, not autonomous performers. |
| Technical feasibility today | claude-sonnet-5 | 1/5 | No deployed product autonomously designs, builds, and tests physical experimental or medical equipment alongside human scientists; this remains research-stage at best. |
Related occupations — Life, Physical & Social Science
How to read this
A high substitution score does not mean this job disappears — it means a large share of its current tasks face replacement pressure, so the mix of tasks is likely to change. High augmentation alongside substitution typically means the occupation reorganizes around the protected tasks. Wide confidence intervals mean the rater panel disagreed: treat those scores as open questions, not verdicts.
What would change this score
New model capabilities (automatability, feasibility), falling inference costs (cost ratio), regulation and licensing shifts (barriers), and measured sector adoption (velocity) all re-enter at every index release. Each release is recomputed, versioned and kept queryable — scores are claims with a date on them, not permanent labels.