Privacy-friendly analytics help us see which jobs people explore. They stay off unless you agree.Analytics off unless you agree. Details

Read the record
Back to the river
Skip to sources
Time Machine

Computer research scientists

Official title Computer and Information Research Scientists (SOC 15-1221)

Also called Computer and information research scientist, Research scientist, ML research scientist, AI research scientist, Computer science research scientist

In brief

In May 2025, BLS counted 37,200 US computer research scientists, with median pay of $140,300 a year. Of the BLS counts since 2021, the highest was 38,480, in May 2024, close to the May 2025 count. The earliest figure here is an estimate of 500 from 1950, not measured the same way as today's count. The last full technology era for computer research scientists: Deep learning frameworks and GPU computing, 2010 to 2020. The deep learning era drove a sharp increase in demand for research scientists.

What the sources say

Official outlook

BLS projects employment to grow 21.8% from 2025 to 2035.

BLS Employment Projections 2025-35

AI exposure studies

Data as of BLS OEWS May 2025; BLS Employment Projections 2025-35. BLS publishes these counts as yearly survey snapshots and cautions against reading them as a precise trend.

Last updated .

Cite this page
APA

Future History. (2026, October 11). Computer research scientists: jobs and pay since 1950 [Data set; data as of BLS OEWS May 2025; BLS Employment Projections 2025-35]. https://futurehistory.earth/timeline/15-1221

Chicago

Future History. "Computer research scientists: jobs and pay since 1950." Data as of BLS OEWS May 2025; BLS Employment Projections 2025-35. Last modified October 11, 2026. https://futurehistory.earth/timeline/15-1221.

BibTeX

@misc{futurehistory_15_1221, author = {{Future History}}, title = {Computer research scientists: jobs and pay since 1950}, year = {2026}, month = oct, howpublished = {\url{https://futurehistory.earth/timeline/15-1221}}, note = {Last updated 2026-10-11. Data as of BLS OEWS May 2025; BLS Employment Projections 2025-35}, }

The access date is today's date on your device.

drag to travel through time
195019752000now
2026
Known then as Computing Research Mathematician / Computing Machine Operator-Scientist (pre-discipline era)
Known then as Computer Scientist / Computer Specialist (first academic departments era)
Known then as Computer Scientist / Computer Research Scientist (professionalization era)
Known then as Computer Scientist / Research Scientist (internet and workstation era)
Known then as Computer and Information Research Scientists (BLS SOC 15-1221)
Known today as Computer and Information Research Scientists (BLS SOC 15-1221)
Projected to still be called Computer and Information Research Scientists (BLS SOC 15-1221)
Data from the May 2025 survey
US Employment
37K
Est.

Stated in source: BLS OEWS

Official statistic
Report an error
Median Annual Wage
$140,300

≈ $136,703 in 2024 prices (derived by us, CPI-adjusted)

Official statistic
Report an error
Beat · 1946

March 1946: John von Neumann begins assembling the IAS computer project at Princeton, hiring Julian Bigelow as chief engineer and handing him Turing's 1936 computability paper as his first assignment. The same year, IBM opens the Watson Scientific Computing Laboratory at Columbia University under Wallace Eckert, the first corporate pure-science computing research facility in the United States. These two events define the twin poles of the emerging profession: academic theory and industrial application, already intertwined from the beginning.

Beat · 1957

RAND Corporation spins off the Systems Development Corporation (SDC) to build the SAGE air defense system, the world's first real-time computer network. SDC trained more than 10,000 employees in computer systems work, dispersing most to other companies and institutions across the country. This is the most consequential single act of computing workforce development in the 1950s: RAND and SDC effectively trained the first generation of computing professionals at scale, funded almost entirely by the US Air Force.

Beat · 1962

October 1962: Purdue University establishes the first Department of Computer Sciences in the United States. Stanford and Carnegie Mellon follow in 1965 (Carnegie Tech with a $5 million R.K. Mellon Foundation grant); MIT in 1968. DARPA's J.C.R. Licklider accelerates the process by calling his friends at MIT, Carnegie Tech, and Stanford and awarding them research contracts to "go do computer science." The transformation from scattered researchers at physics and math departments to an organized academic discipline with named departments, degree programs, and tenure lines happens in a span of about six years.

Beat · 1969

October 29, 1969: the first message is sent over ARPANET, connecting UCLA, Stanford Research Institute, UC Santa Barbara, and the University of Utah. The message was "LO", an attempt to send "LOGIN" that crashed the SRI system after two letters. ARPANET was funded by DARPA and built by BBN Technologies. The network was designed by computer scientists (most prominently Bob Kahn and Vint Cerf), for computer scientists, to enable collaboration between research teams. It became the internet. The researchers who built it were a direct product of the federal investment in university CS departments that had begun with DARPA's calls to Licklider's contacts in 1963.

Beat · 1991

Paul Ginsparg at Los Alamos National Laboratory launches arXiv as an email-based preprint server for high-energy physics papers. By 1998 it has expanded to cover mathematics, nonlinear science, and computer science (specifically including the cs.AI and cs.LG subject areas that would become the central venues for ML research). ArXiv fundamentally changes the pace of computer science research: findings circulate within days of completion rather than waiting 12-18 months for peer-reviewed conference proceedings. The deep learning era that began around 2012 is inseparable from the culture of immediate preprint sharing arXiv enabled.

Beat · 2012

September 2012: AlexNet, a deep convolutional neural network by Alex Krizhevsky, Ilya Sutskever, and Geoffrey Hinton at the University of Toronto, wins the ImageNet Large Scale Visual Recognition Challenge by a margin that stuns the computer vision community (15.3% top-5 error vs. 26.2% for the runner-up). The result was enabled by training on two NVIDIA GTX 580 GPUs. AlexNet is the starting gun for the modern deep learning era: within two years, virtually every computer vision research lab had switched from handcrafted features to neural networks. Google Brain (founded 2011), Facebook AI Research (founded 2013), and DeepMind (acquired by Google 2014) all expanded rapidly to hire the researchers who could advance this direction. BLS employment for 15-1221 begins accelerating.

Beat · 2024

July 2024: Google DeepMind's AlphaProof and AlphaGeometry 2 achieve silver-medal-level performance on International Mathematical Olympiad problems by combining large language model conjecture generation with Lean4 formal proof verification. This is the first time an AI system has matched elite human performance on original mathematical reasoning problems. For computer research scientists working in theoretical CS and mathematics, AlphaProof represents both a tool (LLM-guided proof search can now assist in verifying complex theorems) and a subject of study (the architecture raises fundamental questions about the nature of mathematical understanding and creativity). It is one of the few research results that simultaneously advances the state of the art and reshapes the methodology used to do so.

How to read this number

OEWS is a point-in-time survey snapshot, not a continuous time series; BLS advises against using it for year-over-year trend comparison.

Each dot is a cited figure over time; the dotted line only links them (values between aren't measured). Hollow dots are estimates.

Official statistic: Copied by a script from a statistics agency table or a published study's table, not retyped by hand. BLS figures and the forecast and exposure tables are also re-checked against that table on every data run. Researched context: Researched from the cited source rather than copied from it by a script, so it is not machine-checked yet. Spot a mistake? Each figure has a “Report an error” link.

Cited forecasts · present → 2035

Will AI replace computer research scientists? What the research says

Each figure is stated by the named source; we publish no forecasts of our own. Where sources disagree, we show the disagreement rather than an average.

Each figure is stated by the named source; we publish no forecasts of our own. Where sources disagree, the spread is the honest signal.

What published forecasts say
Forecast change in the number of people doing this work, as published by each named source. Bars share one scale on every role, centred on zero and reaching the end at ±30%. Grey marks a small change, 3% or less either way.
BLS Employment Projections 2025-35
→ 2035
+21.8%
BLS National Employment Matrix, 2025-35 cycle, line for SOC 15-1221 (Computer and information research scientists): 38.6 thousand jobs in 2025, projected 47.0 thousand in 2035, a change of +8.4 thousand (+21.8%). Matrix employment counts all jobs in the occupation, including self-employed workers. The average for all occupations in this cycle is +3.5%.
Official statisticReport an error
How exposed is this work to AI (research measures)
Scores as published by each study. Each row says what it measures. These measure what AI could technically do, not forecasts of jobs lost.
Eloundou et al. (2023), "GPTs are GPTs"
2023 study · Share of tasks exposed to LLMs (β, human-rated)
50%
of tasks
What this measures

Published occupation-level exposure from the paper's occ_level.csv for O*NET-SOC 15-1221.00 Computer and Information Research Scientists. Shown: human-rated β = 0.5, the share of the occupation's tasks an LLM could do in at least half the time at equal quality, counting tasks that need LLM-powered software at half weight (β = E1 + 0.5 × E2). Other published measures for this occupation: human-rated α (LLM alone) 0.231 and γ (all LLM-assisted) 0.769; GPT-4-rated β 0.635. This measures how much of the work an LLM could touch, not how many jobs disappear.

Official statisticReport an error
Today, in this role

What's changing day to day for computer research scientists

Three parts of your work where AI is already doing real lifting, and what stays yours.

AI is sitting alongside you hereConduct a comprehensive literature review across arXiv, NeurIPS, ICML, ICLR, and domain-specific venues using AI-assisted synthesis tools (Elicit, Semantic Scholar AI) to surface related work, identify citation gaps, map the theoretical landscape, and determine positioning of a new research contribution relative to the prior art.

Conduct a comprehensive literature review across arXiv, NeurIPS, ICML, ICLR, and domain-specific venues using AI-assisted synthesis tools (Elicit, Semantic Scholar AI) to surface related work, identify citation gaps, map the theoretical landscape, and determine positioning of a new research contribution relative to the prior art.[9],[10],[5]

Where your edge is

Use Elicit and Semantic Scholar AI to surface related papers quickly, but read the seminal works yourself: AI summaries lose the nuance, mathematical detail, and intellectual lineage that determines whether your contribution is actually novel or has a close predecessor. Develop the habit of building personal citation maps that connect ideas across subfields, not just within your narrow topic.

AI is sitting alongside you hereImplement novel model architectures, training algorithms, and evaluation frameworks in PyTorch or JAX, writing research-quality code that is reproducible, ablation-ready, and organized for collaboration, with GitHub Copilot or Cursor handling boilerplate while the researcher designs the non-obvious architectural choices that differentiate the contribution.

Implement novel model architectures, training algorithms, and evaluation frameworks in PyTorch or JAX, writing research-quality code that is reproducible, ablation-ready, and organized for collaboration, with GitHub Copilot or Cursor handling boilerplate while the researcher designs the non-obvious architectural choices that differentiate the contribution.[8],[6]

Where your edge is

AI coding tools dramatically accelerate boilerplate (data loaders, logging, standard training loops) but cannot replace the judgment required to design a novel architecture or identify why a training run diverged. Develop deep mental models of gradient flow, numerical stability, and distributed-training tradeoffs so you can diagnose issues that Copilot-generated code silently introduces at scale.

AI is sitting alongside you hereStay current with the ML and CS research frontier through systematic arXiv monitoring (Papers with Code, daily arXiv digest tools, Semantic Scholar alerts) and reproduce key results from influential new papers to build ground-truth understanding, distinguishing between papers with lasting theoretical impact and papers that overstate empirical results on benchmark leaderboards.

Stay current with the ML and CS research frontier through systematic arXiv monitoring (Papers with Code, daily arXiv digest tools, Semantic Scholar alerts) and reproduce key results from influential new papers to build ground-truth understanding, distinguishing between papers with lasting theoretical impact and papers that overstate empirical results on benchmark leaderboards.[11],[5]

Where your edge is

The skill of "knowing what is real and what is hype" in ML research is not automatable and is one of the most valuable competencies a research scientist can develop. AI tools (Elicit, Semantic Scholar AI) surface papers efficiently; the human judgment is in the reading, the reproduction attempt, and the slow accumulation of calibrated intuition about which results generalize. Reproduce three papers per year from scratch: the ones that do not reproduce teach more than the ones that do.

Where computer research scientists go next

Natural next steps for someone with your foundation: not exits, evolutions.

What each move takes

The research lab leadership path evolves a CS research scientist into a Research Director, VP of Research, or Head of AI: responsible for building research teams, defining multi-year research agendas, managing compute budgets, and representing research work to executive leadership and external stakeholders. This transition requires developing management skills (hiring, performance evaluation, conflict resolution) and organizational influence capabilities that differ fundamentally from individual research. The CRI delta is negative because management roles are more exposed to org restructuring. However, research leaders with strong scientific credentials and management acumen are genuinely rare and command exceptional compensation.

What you'd add
  • · Research team building: hiring Research Scientists and Research Engineers at all levels; calibrating research potential through paper reviews and technical interviews
  • · Research agenda setting: allocating team effort across explore/exploit tradeoffs; building a coherent lab identity
  • · Executive communication: translating long-horizon research bets into business value narratives for non-technical leadership
  • · Budget and compute resource management: negotiating GPU cluster allocation and headcount in competitive internal resource environments
  • · Mentorship and performance management: giving research feedback that improves junior researcher quality and career trajectory
What it takesA real upskill, but a natural one
Tools of the era

How has the work of computer research scientists changed since 1946?

The tools that defined the work, one line per era. Select an era to see how it reshaped the work.

  • Physical machines and mathematical notation (vacuum tube computers, punchcards, FORTRAN)

    The first computer scientists worked directly on the machines they studied: wiring circuits, writing binary and assembly code, and debugging via physical inspection of hardware state. Programming in 1950 meant setting toggle switches or feeding punched paper tape. FORTRAN (1957) was the first high-level language and a genuine leap: for the first time a scientist could express an algorithm in something closer to mathematical notation and have a compiler translate it to machine code. Research in this era was inseparable from engineering the machines. A "computer scientist" was often simultaneously the theorist, the builder, and the operator.

    Effect on the work

    Employment grew very slowly because the machines were expensive, scarce, and required specialized training. Federal defense contracts (SAGE, JOHNNIAC, ENIAC successors) and a handful of corporate labs (IBM Research, Bell Labs, Xerox) employed nearly all of the research workforce.

  • Time-sharing systems and interactive terminals (CTSS, Multics, ARPANET, Unix)

    MIT's Compatible Time-Sharing System (CTSS, 1961) and its successor Multics (1969) changed how researchers interacted with computers: instead of submitting a batch job and waiting hours for output, a scientist could sit at a terminal and iterate in real time. DARPA's ARPANET (first packet sent October 29, 1969) connected research computers across universities, allowing collaboration that previously required physical presence. Bell Labs researchers Ken Thompson and Dennis Ritchie created Unix in 1969 specifically because Multics was too complex, and Unix became the operating environment for computer science research for the next two decades. For computer scientists, these tools transformed the research process from asynchronous batch experimentation to interactive problem-solving.

    Effect on the work

    Time-sharing expanded the community of active researchers: a single machine could support many users simultaneously, lowering the per-researcher cost of access. DARPA's "Project MAC" at MIT drew heavily on NSF seed funding and trained a generation of researchers who later founded or populated university CS departments across the country.

  • Workstations and C/Lisp/Prolog (Sun, Apollo, Symbolics, LaTeX for papers)

    The Unix workstation era (Sun Microsystems, Apollo Domain, DEC VAX) gave individual researchers dedicated computing resources for the first time. A researcher no longer shared a mainframe with hundreds of others; a Sun-3 on a lab desk could run simulations, edit LaTeX documents, and compile code without waiting for batch queues. The AI and symbolic-computation communities developed Lisp machines (Symbolics, LMI) and Prolog interpreters for logic programming. LaTeX (1984) standardized scientific paper writing and is still universal in CS research today. The workstation era produced a major acceleration in empirical computer science: experiments that once required months of batch queue time could now be iterated in an afternoon.

    Effect on the work

    University CS department enrollment grew sharply through the 1980s as computing became economically visible. The 1983 "A Nation at Risk" and subsequent federal science policy reports identified computing education as a national priority. Graduate enrollment in CS PhD programs grew between 1975 and 1990 according to CRA data, producing the research workforce that would staff the internet era.

  • The web, open-source tools, and arXiv (Linux, Python, MATLAB, preprint culture)

    The World Wide Web (Tim Berners-Lee, CERN, 1991) changed scientific communication as fundamentally as the printing press changed scholarship. ArXiv, launched in 1991 by Paul Ginsparg at Los Alamos for physics preprints, expanded to computer science (cs.AI, cs.LG) in 1998 and transformed the field: research no longer needed to wait several months for peer review before circulating. Python (1991) and its scientific stack (NumPy 2006, SciPy 2001) gave researchers a more expressive alternative to C and Fortran for data analysis. Linux made research computing infrastructure essentially free. The open-source culture that emerged in this era also shaped how computer scientists think about their own tools: software as a public good, reproducibility as a norm, and citation of code repositories alongside papers.

    Effect on the work

    The internet era bifurcated the occupation: academic researchers benefited from lower barriers to collaboration and distribution, while industry research labs expanded dramatically as tech companies recognized the commercial value of CS research. By 2000 Microsoft Research (founded 1991), Google Research (founded 1999), and Yahoo Research had each hired hundreds of PhD research scientists.

  • Deep learning frameworks and GPU computing (Theano 2010, TensorFlow 2015, PyTorch 2016)

    The deep learning revolution began in earnest with Geoffrey Hinton's 2012 AlexNet result: a convolutional neural network trained on a GPU cluster won the ImageNet competition by a margin that stunned the computer vision community. GPU computing, pioneered by NVIDIA, compressed the research iteration cycle dramatically: experiments that required weeks on CPU clusters could be run overnight on a GPU. Theano (2010, Montreal), TensorFlow (2015, Google), and PyTorch (2016, Facebook AI Research) gave researchers an ecosystem of differentiable programming tools that let them prototype novel architectures in Python with automatic differentiation. For computer research scientists, these tools did not automate research but massively accelerated the empirical loop: hypothesis, implementation, experiment, result was now a cycle measured in days rather than months.

    Effect on the work

    The deep learning era drove a sharp increase in demand for research scientists. Google Brain was founded in 2011; DeepMind was acquired by Google in 2014; OpenAI founded in 2015; Uber AI Labs, FAIR (Meta AI Research), and dozens of other corporate labs expanded. BLS employment for 15-1221 reached 30,780 by 2019.

  • LLM-assisted research tools (Elicit, Semantic Scholar AI, GitHub Copilot, formal proof assistants)

    The large language model era produced a new class of research tools that began automating the most labor-intensive, least intellectually demanding parts of the research scientist's workflow. Elicit (launched 2022) uses LLMs to surface, summarize, and compare papers across literature, compressing a multi-week systematic review into hours. GitHub Copilot (2022) generates boilerplate code from natural-language descriptions. Weights and Biases AI-assisted insights automatically analyze experiment results. Google DeepMind's AlphaProof (2024) demonstrated LLM-guided formal proof search in Lean4 at silver-medal IMO level, opening a new frontier for AI-assisted theorem proving in theoretical CS. A 2025 Nature survey found that 68% of researchers use LLMs weekly for literature synthesis, hypothesis generation, and code writing. The paradox of this era is that the tools these researchers build are also accelerating their own work, creating a feedback loop: faster research iteration means the frontier moves faster, which means more demand for researchers who can keep up.

    Effect on the work

    The generative AI boom (ChatGPT launched November 2022; GPT-4 March 2023; Claude, Gemini, Llama) has not automated research scientists but has created massive new demand for them across both the AI labs that build these systems and the institutions that study their effects.

Part of Tech & Data · see all 28 roles →
Different role?

See the same long-arc view for your own profession.

Browse the directory by industry, or search by title or SOC code.

Browse all roles

Computer and Information Research Scientists employment and pay by year

On record since1946
Latest US employment37,200 (May 2025)Source: BLS OEWS May 2025Official statisticReport an error
Latest US median pay$140,300 (May 2025)Source: BLS OEWS May 2025Official statisticReport an error
Outlook+21.8% by 2035Source: BLS Employment Projections 2025-35Official statisticReport an error
View all 12 years of cited US data, each linked to its source
YearUS employmentAnnual pay (median unless marked)SourceReport
1950500n/aEst., National Academies Press Researched contextReport
196310,000n/aEst., National Academies Press Researched contextReport
197075,000n/aDecennial Census 1970 Researched contextReport
1990280,000n/aDecennial Census 1970 Researched contextReport
200025,800$70,590BLS OEWS 2000 Official statisticReport
201930,780$122,840BLS OEWS May 2019 Official statisticReport
202030,220$126,830BLS OEWS May 2020 Official statisticReport
202130,840$131,490BLS OEWS May 2021 Official statisticReport
202233,780$136,620BLS OEWS May 2022 Official statisticReport
202335,210$145,080BLS OEWS May 2023 (employment) Official statisticBLS OEWS May 2023 (pay) Official statisticReport
202438,480$140,910BLS OEWS May 2024 Official statisticReport
202537,200$140,300BLS OEWS May 2025 Official statisticReport
Embed this timeline on your site

Free for any site. Paste this where the timeline should appear; it stays interactive, every datapoint stays cited, and it sets no cookies on your page. How embedding works

<iframe src="https://futurehistory.earth/embed/15-1221"
  width="100%" height="430" style="border:0"
  title="Computer and Information Research Scientists, a Future History timeline"
  loading="lazy"></iframe>
<p style="font-size:13px;margin:4px 0 0"><a href="https://futurehistory.earth/timeline/15-1221">Computer and Information Research Scientists: jobs and pay over time</a>, from Future History</p>

See all roles in Tech & Data