{"slug": "human-vs-ai-vs-human-and-ai-who-does-better-work", "title": "Human vs. AI vs. Human and AI: Who Does Better Work?", "summary": "A review of eight controlled studies comparing human, AI, and human-plus-AI performance finds that AI's advantage is not universal: it is large and reproducible on some tasks but reverses or disappears on others that appear similar, challenging the common assumption that human-plus-AI always outperforms either alone. The studies span clinical diagnosis, management consulting, customer support, professional writing, and software engineering, and the pattern is corroborated by a large independent meta-analysis.", "body_md": "There's a comfortable assumption running through most corporate AI strategy right now: a skilled person plus an AI tool will outperform either one alone. It's intuitive, it's reassuring, and it lets organizations roll out AI everywhere without asking hard questions about where, specifically, it helps.\n\nIt's also not what the evidence shows.\n\nWe identified eight controlled empirical studies — through a structured literature search described below, not an arbitrary shortlist — that compare at least two of unassisted human performance, AI-alone performance, and human-plus-AI performance on a well-defined task, using real accuracy, speed, or quality measurements rather than survey sentiment. Only a subset of the eight contain all three conditions in a single design; the remainder provide controlled two-arm evidence (most often human-alone vs. human+AI, or human-alone vs. AI-alone) that helps test whether the broader pattern generalizes across tasks — which study includes which conditions is noted in each section below and in the summary table. They span clinical diagnosis, management consulting, customer support, professional writing, and software engineering. These are not eight versions of the same experiment. They differ in design, sample, and what they measure: most are randomized controlled trials, one is a staggered field rollout with a randomized pilot; some measure a single 20-minute task, one follows real production work across three companies over months. Read individually, each study answers a narrow question about its own task and population. Read together, a pattern emerges that is not \"AI helps.\" It's closer to a fault line: AI's advantage is real, large, and reproducible on some tasks, and reverses into a disadvantage — or simply disappears — on others that look, to a human, more or less the same. A large independent meta-analysis of the broader literature finds a closely aligned pattern, which is the strongest evidence this isn't an artifact of which eight studies we happened to pick.\n\n## Methodology: how these eight studies were selected\n\nTo avoid presenting a hand-picked set as more authoritative than it is, we ran a structured — not a formal systematic-review-protocol — literature search across Google Scholar, Semantic Scholar, PubMed, SSRN, NBER, arXiv, the ACM Digital Library, and IEEE Xplore for controlled studies published in 2022 or later that met all of the following: (1) compare at least two of unassisted human performance, AI-alone performance, and human+AI performance; (2) use a real or realistic work task, not a survey or self-reported opinion; (3) report a quantitative outcome — accuracy, speed, a graded quality score, or an error/completion rate, not satisfaction; (4) are peer-reviewed and published, or are a working paper from an established research institution or lab, not a vendor blog post or marketing study; and (5) cover clinical diagnosis, knowledge work, customer service, writing, or software engineering, with a broader scan across education, forecasting, hiring, translation, legal work, and classic human-AI \"centaur\" research to check for major work we might otherwise miss. This is a transparent, criteria-driven search, not a formal systematic review — it does not follow a pre-registered PRISMA-style protocol, log exact search strings or per-database result counts, or use a second independent screener, and readers who need that standard of evidence should treat it accordingly.\n\nThat search surfaced roughly two dozen candidates. Most were excluded for a specific, checkable reason: some measured self-reported time allocation or satisfaction rather than task performance (a large Microsoft 365 Copilot field study of over 7,000 workers, NBER Working Paper 33795, measured how workers reallocated their time, not whether their output improved); some were observational rather than randomized, which weakens causal claims (a study of 72,000+ GitHub pull requests found AI-assisted PRs merged faster, but without random assignment); and a few strong two-arm studies (AI-alone vs. human-alone, no combined condition) in medicine and forecasting were kept as corroborating context rather than counted as primary sources, to avoid overweighting any single domain. The eight studies below are the ones that survived that screen with the strongest designs available — eight primary studies in total, including one earlier preprint (Peng et al., 2023) that we retained specifically as a labeled historical predecessor to a later, stronger study on the same question, rather than as independent evidence in its own right.\n\nWe also checked our pattern against the closest thing to\nindependent confirmation available: Vaccaro, Almaatouq, and\nMalone's 2024 meta-analysis in *Nature Human Behaviour*,\nwhich pooled 106 prior experiments (370 effect sizes) comparing\nhuman-alone, AI-alone, and human-AI teams across many task\ntypes. 1 Its headline finding — that\nhuman-AI combinations underperform the better of either alone on\naverage, with the effect reversing between decision-making tasks\n(where combining hurts) and content-creation tasks (where combining\nhelps) — is a closely aligned, task-dependent, non-uniform pattern\nto the one this piece finds in its own eight studies. That\nconvergence is meaningful: it means the \"it depends on the task\"\nconclusion below isn't a quirk of which eight studies we\nselected, but matches what a systematic\nreview of the wider literature already shows.\n\n## The clearest three-way test: doctors, AI, and doctors using AI\n\nThe most direct answer to \"who does better work\" comes from a\nsingle-blind randomized clinical trial published in *JAMA\nNetwork Open* on October 28, 2024, run across Stanford, Beth\nIsrael Deaconess Medical Center, and the University of\nVirginia. 2 Fifty US-licensed\nphysicians (26 attendings, 24 residents; median 3 years in\npractice) in family medicine, internal medicine, and emergency\nmedicine were randomized into two groups and given up to six\nclinical vignettes — patient histories, exam findings, lab\nresults, adapted from a landmark 1994 diagnostic-systems study —\nto work through in a one-hour session. One group used conventional\nresources only (UpToDate, Google, and similar). The other used\nthose same conventional resources plus ChatGPT Plus (GPT-4). A\nthird, non-randomized condition ran the same vignettes through\nGPT-4 alone, three separate times, with no physician involved.\nParticipants completed 244 cases in total (125 in the LLM group,\n119 in the control group).\n\nThe results, as published 2:\n\n| Condition | Median diagnostic reasoning score |\n|---|---|\n| Physicians using conventional resources | 74% (IQR 63–84) |\n| Physicians using conventional resources + ChatGPT Plus | 76% (IQR 66–87) — adjusted difference +2 points, 95% CI −4 to 8; not statistically significant, p=0.60 |\n| GPT-4 working alone (median of 3 runs) | 92% (IQR 82–97) — 16 points higher than conventional resources, 95% CI 2 to 30; statistically significant, p=0.03 |\n\nOn the clinical cases tested, the LLM working alone scored a median of 92% — clearly ahead of both physician groups, including physicians who had access to the LLM. Adding ChatGPT to a physician's workflow did not produce a statistically significant accuracy gain over conventional resources alone. Physicians using AI were modestly faster on average (519 seconds per case vs. 565), a difference the paper's own results text reports as not statistically significant (95% CI −195 to 31 seconds, p=0.20).\n\nIt's worth being precise about what this study does and doesn't show, because the authors themselves are unusually direct about it. Their own words: \"Results of this study should not be interpreted to indicate that LLMs should be used for diagnosis autonomously without physician oversight.\" The clinical vignettes were curated and summarized by human clinicians rather than experienced firsthand, the study excluded patient interviewing and data collection, and the setting was acontextual — none of which capture the full scope of real clinical reasoning. What the study does show, within that bounded scope, is that a physician's real-time judgment about when to trust or override an AI's suggestion did not reliably improve on the AI's unaided output — and on these cases, the human addition did not clear the bar of statistical significance in either direction.\n\n**A second, larger trial by an overlapping author team\ncomplicates this picture in an important way.** A\nfollow-up randomized controlled trial published in *Nature\nMedicine* in 2025 — led by several of the same researchers,\nusing a different task design — found that physicians using GPT-4\non patient care tasks *did* score significantly higher than\nphysicians using conventional resources alone: +6.5 points (95%\nCI 2.7–10.2, p<0.001). 3 In that\ntrial, 92 physicians were split across three arms (conventional\nresources, GPT-4 plus conventional resources, and GPT-4 alone),\nand — as in the first study — GPT-4 alone was statistically\nindistinguishable from physicians using GPT-4 (difference −0.9%,\n95% CI −9.0 to 7.2, p=0.8). GPT-4-assisted physicians also took\nmeaningfully longer per case (+119 seconds, p=0.02).\n\nPut the two studies from the same research group side by side and the honest conclusion isn't \"human+AI helps\" or \"human+AI doesn't help\" — it's that the answer depends on task design in ways not yet fully understood, even to the researchers running near-identical experiments. Across both studies, GPT-4 alone was not significantly worse than the physician+GPT-4 condition; however, the studies differed materially in whether physician+AI improved on physicians using conventional resources — a significant gain in the 2025 trial, no significant gain in the 2024 trial.\n\n## Where \"human + AI\" earns its keep — and where it stops\n\nThe diagnosis studies aren't outliers so much as one end of a\nspectrum visible across the research base. The clearest\narticulation of why comes from a large field experiment run by\nresearchers at Harvard Business School with Boston Consulting\nGroup, since published in *Organization Science* as\n\"Navigating the Jagged Technological Frontier.\" 4\nThe paper's central idea — the\n\n**\"jagged frontier\"**— is that AI's competence doesn't fall off gradually as tasks get harder, the way a human's does. It's extraordinary on some tasks and unreliable on adjacent ones that look, to the person assigning the work, about the same.\n\nIn that study, 758 BCG consultants completed 18 realistic\nconsulting tasks, some with GPT-4 access and some without. On\ntasks the researchers classified as inside the model's competence\nzone, AI-assisted consultants completed\n**12.2% more tasks**, **25.1% faster**,\nat measurably higher quality than the unassisted group. On a task\ndesigned to sit *outside* that zone, AI-assisted\nconsultants were **19% less likely to produce a correct\nsolution** than the unassisted control group — a\nstatistically significant effect (p<0.01).\n\nThe finding that should complicate any tidy \"just train people\nbetter\" narrative: consultants who additionally received explicit\ntraining in how to prompt and collaborate with the model did not\ndo better on the out-of-frontier task — they did *worse*.\nThe trained group's correctness dropped by about 24.5 percentage\npoints relative to control, compared to a 13.9-point drop for\nconsultants using GPT-4 without training (both significant; the\ngap between the two AI conditions was marginally significant,\np≈0.08–0.10). Training did modestly improve output\n*quality* on inside-frontier tasks. But on the one task\nbuilt to sit beyond the model's competence, teaching people to use\nthe tool more fluently made them more confidently wrong, not less.\n\nThat's the uncomfortable version of the \"jagged frontier\" finding: the fix for AI failing outside its competence zone isn't more prompting skill. It's knowing, task by task, whether you're inside the zone at all — something the paper doesn't show people learn to detect just by getting better at using the tool.\n\n## The real, repeatable win: AI narrows the gap between your best and worst performers\n\nOne of the most consistent effects in this body of research is this: AI's largest gains tend to go to less experienced workers, not the strongest ones.\n\nA study by Brynjolfsson, Li, and Raymond, published in the\n*Quarterly Journal of Economics* after circulating as an\nNBER working paper, followed roughly 5,179 customer support\nagents at a Fortune 500 software company given staggered access\nto a generative AI assistant that suggested real-time responses\nduring chats. 5 The rollout was not\na full randomized trial — access expanded gradually across teams\n— but it was preceded by a seven-week randomized pilot with 50\nagents, and the main analysis uses three million real chat\ntranscripts. Average productivity, measured as issues resolved\nper hour, rose 14%. That average obscures the more interesting\nresult: agents in their first months on the job improved by\nroughly\n\n**34%**, while the most experienced, highest-performing agents saw little to no measurable benefit — and some evidence of modestly lower resolution quality. The authors' explanation is that the AI assistant functioned less like a universal productivity multiplier and more like a way of distributing the tacit knowledge of the company's best agents to everyone else — useful precisely to the degree someone didn't already have that knowledge.\n\nA separate 2023 study in *Science*, by Noy and Zhang, found\na similar shape in a different domain: professional\nwriting. 6 In a randomized trial,\n453 college-educated professionals (in roles like marketing, HR,\nconsulting, and data analysis) were asked to complete short\noccupation-specific writing tasks — press releases, reports,\nanalysis plans — with about half given access to ChatGPT. Blinded\nevaluators graded the output. ChatGPT access cut time spent by\nabout\n\n**40%** and raised graded quality by about\n\n**18%**, and — as the authors report explicitly — the gain was concentrated among participants who scored below average on the ungraded first task, narrowing the performance gap between weaker and stronger writers. The authors are direct about scope: this measured a specific, short, well-defined writing task with no factual-accuracy requirement and no need for firm-specific context — not general professional writing productivity, and not a test of output accuracy.\n\nA comparable pattern shows up in software engineering, with the\nstrongest evidence here coming from a large, peer-reviewed field\nstudy rather than the small lab experiment that first suggested\nit. Peng, Kalliamvakou, Cihon, and Demirer's 2023 arXiv preprint —\na randomized trial of 95 developers building the same scoped\nHTTP-server task, where the Copilot group finished 55.8% faster\non average (71 vs. 161 minutes) — was the first evidence of this\neffect, but it remained an unpublished preprint testing one\nnarrow, artificial task. A follow-up study by an overlapping\nteam, published in *Management Science* in 2026, replaces\nit as the primary evidence here: three separate randomized\ncontrolled field experiments at Microsoft, Accenture, and an\nanonymous Fortune 100 company, covering\n**4,867 developers doing real production coding\nwork**, not a lab task. 7\nDevelopers with AI coding-assistant access completed\n\n**26.1% more tasks**(SE 10.3%) — and, consistent with the pattern seen in customer support and writing, gains were largest for less-experienced developers. This is a materially stronger piece of evidence than the original preprint: peer-reviewed, fifty times the sample size, and measuring real work rather than a single standardized exercise.\n\nThe original Peng et al. HTTP-server experiment is still worth\nknowing about as the earliest controlled evidence in this\nspace, and its finding wasn't wrong — just\nthin. 8 Its completion-rate\ndifference (about 7 percentage points, not statistically\nsignificant, 95% CI −11 to 25) is not restated with more\nconfidence than the paper itself supports.\n\n## The exception that should worry every engineering leader\n\nWhich makes the sharpest counter-example land harder. In a study\npublished by the independent research nonprofit METR in July\n2025, 16 *experienced* open-source developers — each\naveraging roughly five years working in the specific, large\nrepositories (22,000+ GitHub stars, 1M+ lines of code) they were\nassigned to — worked through 246 real GitHub\nissues. 9 Each issue was randomly\nassigned to a condition allowing or disallowing the use of\nthen-current AI coding tools (Cursor Pro with Claude 3.5/3.7\nSonnet).\n\nDevelopers using AI took **19% longer** to complete\ntheir issues than developers who didn't.\n\nThe gap between belief and reality is the part worth sitting\nwith. Before the study, participants expected AI to speed them up\nby about 24%. Outside forecasters predicted even larger gains —\nan economics/superforecasting panel projected ~39% and a\nmachine-learning-expert panel projected ~38%. After the study,\nhaving just measurably taken longer with AI's help, developers\nstill believed AI had made them roughly **20%\nfaster**. METR tested about twenty candidate explanations\nfor the slowdown (task-selection effects, unfamiliarity with the\ntools, and others) and found the effect robust rather than an\nobvious artifact of the setup.\n\nThis is the most experienced, best-context case in our review — developers working in codebases they already knew well, on tasks squarely inside their own specialty — and it's the one place in this body of research where AI assistance measurably subtracted value while its users were most confident it was adding it.\n\n**A necessary 2026 update.** In February 2026, METR\npublished a follow-up describing serious selection-bias problems\nin its attempts to continue this line of research with new\nparticipants: developers increasingly declined to take on tasks\nwithout AI assistance, or avoided submitting tasks where AI\nhappened to help, both of which bias later samples toward showing\nAI in a better light for reasons unrelated to its actual\neffect. 10 Among the\n\n*original*2025 study's developers, re-measured later, METR's own wording is precise and worth quoting directly: \"we now estimate a speedup of -18%\" — that is, a smaller, still-negative slowdown, not a positive result — with a confidence interval wide enough to span roughly −38% to +9%; a newly recruited group showed an estimated speedup of −4% (CI roughly −15% to +9%). METR describes this as \"very weak evidence\" for how large AI's true effect might be, and notes the selection effects likely mask a larger benefit than the raw, still-negative estimate shows — not evidence that the slowdown has reversed into a speedup — and is redesigning the study to correct for the selection problems before drawing firmer conclusions. The original 19% slowdown remains the cleanest randomized estimate within METR's own research program to date; it should be read as a finding about specific tools and developers in early 2025, not a current or universal estimate of AI's effect on experienced software engineers in 2026, and not necessarily the strongest evidence in the wider literature on this question.\n\n## What the evidence actually supports\n\nPutting the eight studies side by side:\n\n| Study | Domain | Sample | Comparison | Result |\n|---|---|---|---|---|\nGoh et al., 2024 (JAMA Network Open) | Clinical diagnosis | 50 physicians, 244 cases | Physicians vs. physicians + GPT-4 vs. GPT-4 alone | GPT-4 alone: 92%, 16 points higher than conventional (p=0.03); physician + GPT-4 (76%) vs. conventional (74%) not significant (p=0.60) |\nGoh et al., 2025 (Nature Medicine) | Clinical diagnosis | 92 physicians | Physicians vs. physicians + GPT-4 vs. GPT-4 alone | Physician + GPT-4: +6.5 pts over conventional (p<0.001, significant); GPT-4 alone ≈ physician + GPT-4 (p=0.8) |\nDell'Acqua et al., 2023/2026 (Organization Science) | Management consulting | 758 consultants, 18 tasks | No AI vs. GPT-4 vs. GPT-4 + training | +12.2% tasks, +25.1% speed in-frontier; −19% to −24.5% accuracy out-of-frontier |\nBrynjolfsson, Li & Raymond, 2023/2025 (QJE) | Customer support | ~5,179 agents, 3M chats | AI-assisted vs. unassisted, by tenure | +14% average; +34% novices, ~0% experts |\nNoy & Zhang, 2023 (Science) | Professional writing | 453 participants | Unassisted vs. ChatGPT-assisted | −40% time, +18% quality; gains concentrated in below-average writers |\nCui et al., 2026 (Management Science) | Software engineering | 4,867 developers, 3 companies | AI-assisted vs. unassisted, real production work | +26.1% completed tasks; gains largest for less-experienced developers |\n| Peng et al., 2023 (arXiv preprint) | Software engineering (single task) | 95 developers | Copilot vs. no Copilot | 55.8% faster mean completion time; completion-rate difference not significant |\n| METR, 2025 | Software engineering (experienced) | 16 developers, 246 real issues | AI-permitted vs. AI-prohibited | 19% slower with AI; devs predicted +24%, believed +20% afterward |\n\nFive conclusions survive contact with all eight studies at once:\n\n**AI's advantage is task-specific, not domain-specific.**\n\"AI is good at coding\" and \"AI is good at writing\" are both true and\nboth misleading — inside the same general domain, the same tool\nswings from a 55.8–26.1% speed or throughput advantage to a 19%\ndisadvantage depending on the specific task and the specific\nperson doing it. Even within the single domain of clinical\ndiagnosis, two trials from an overlapping research team reached\ndifferent conclusions about whether combining a physician with AI\nbeats the physician alone — task design, not the domain itself,\ndecided the outcome.\n\n**Human involvement is not automatically a safety net —\nbut it isn't automatically useless either.** Across the\ntwo diagnosis trials, adding a physician to the AI's process\nsometimes produced a statistically significant improvement\n(Nature Medicine, 2025: +6.5 points, p<0.001) and sometimes\ndidn't (JAMA Network Open, 2024: +2 points, not significant) —\nand in the consulting study, adding training in how to use the AI\nmade an out-of-competence task *worse*, not better. The\none pattern that held in both diagnosis trials: GPT-4 alone was\nnever significantly worse than the physician-plus-GPT-4\ncondition. Human oversight is not a reliable corrective by\ndefault in either direction — it needs to be built and tested\ndeliberately for the specific task at hand, not assumed to help\nor assumed to hurt.\n\n**One of the most consistent effects across these studies\nis compression, not elevation.** In the studies that\nstratified results by prior skill or experience — customer\nsupport, writing, and now software engineering at production\nscale — AI's most reliable contribution was narrowing the gap\nbetween newer and more experienced workers, not making the\nstrongest performers meaningfully better.\n\n**Self-reported productivity beliefs cannot be trusted on\ntheir own.** The one study that measured both belief and\nmeasured outcome found them pointing in opposite directions, and\nthe belief did not correct itself even after participants had\nseen their own recorded work.\n\n**This pattern isn't specific to our eight studies — a\nsystematic review of the wider literature finds the same\nthing.** Vaccaro, Almaatouq, and Malone's meta-analysis of\n106 prior experiments found that human-AI combinations\nunderperform the better of either alone on average, with the\ndirection of the effect flipping between decision-making tasks\nand content-creation tasks. That's independent confirmation that\n\"it depends on the task\" isn't a hedge invented to fit eight\ncherry-picked results — it's what the broader field has already\nfound at scale.\n\nNone of this is an argument against using AI. It's an argument against using it uniformly, without measurement, and without training people on where the line between \"inside the frontier\" and \"outside it\" actually sits for their specific work. The organizations that benefit most are likely to be the ones willing to test at the task level, measure outcomes instead of trusting sentiment, and treat \"when to override the model\" as a skill worth teaching — not the ones that assume the upside is automatic.\n\n## Limitations\n\nThis is a selected evidence review of eight primary studies,\nchosen through the documented search and screening process\ndescribed above, corroborated against one independent\nmeta-analysis of the wider literature — it is still not a full\nsystematic review or meta-analysis in its own right, and does not\nrepresent the complete body of research on AI and worker\nproductivity. The eight studies use different populations, tasks,\nAI models and versions, outcome measures, and experimental\ndesigns, which limits direct comparison between them; the\nsummary table above should be read as eight separate,\ntask-specific findings rather than eight data points on a single\nscale. Most of these studies measure performance on a short,\nbounded task in a controlled setting, not sustained real-world\nworkplace productivity over time — the clearest exceptions being\nthe *Management Science* study of 4,867 developers' real\nproduction work, and the Brynjolfsson et al. field data — and\nseveral authors explicitly caution against extrapolating beyond\ntheir own task design. AI models and tools are also changing\nquickly — several of these studies used GPT-3.5/GPT-4-era models,\nand even the most recent study here (METR, 2025) is already the\nsubject of a 2026 follow-up suggesting results may shift as tools\nimprove. Taken together, the appropriate conclusion from this\nevidence base is not a single number for \"AI's productivity\neffect,\" but that the effect is heterogeneous and depends\nheavily on the specific task, the worker's prior experience, and\nhow the human-AI interaction is designed and trained for — a\nconclusion this piece's own eight studies and an independent\n106-study meta-analysis both reach.\n\n## Sources\n\n- Vaccaro, M., Almaatouq, A., & Malone, T. W. (2024).\n[\"When Combinations of Humans and AI Are Useful: A Systematic Review and Meta-Analysis.\"](https://doi.org/10.1038/s41562-024-02024-1)*Nature Human Behaviour*, 8(12), 2293–2303. Meta-analysis of 106 prior experiments; used as independent corroboration, not as a primary study. - Goh, E., Gallo, R., Hom, J., et al. (2024).\n[\"Large Language Model Influence on Diagnostic Reasoning: A Randomized Clinical Trial.\"](https://doi.org/10.1001/jamanetworkopen.2024.40969)*JAMA Network Open*, 7(10), e2440969. Published October 28, 2024. - Goh, E., Gallo, R. J., Strong, E., et al. (2025).\n[\"GPT-4 Assistance for Improvement of Physician Performance on Patient Care Tasks: A Randomized Controlled Trial.\"](https://doi.org/10.1038/s41591-024-03456-y)*Nature Medicine*, 31(4). - Dell'Acqua, F., McFowland III, E., Mollick, E. R., Lifshitz-Assaf, H., Kellogg, K., Rajendran, S., Krayer, L., Candelon, F., & Lakhani, K. R. (2026).\n[\"Navigating the Jagged Technological Frontier: Field Experimental Evidence of the Effects of Artificial Intelligence on Knowledge Worker Productivity and Quality.\"](https://doi.org/10.1287/orsc.2025.21838)*Organization Science*, 37(2), 403–423. Originally circulated as Harvard Business School Working Paper No. 24-013, 2023. - Brynjolfsson, E., Li, D., & Raymond, L. R. (2025).\n[\"Generative AI at Work.\"](https://doi.org/10.1093/qje/qjae044)*The Quarterly Journal of Economics*, 140(2), 889–942. Originally circulated as NBER Working Paper 31161, 2023. - Noy, S., & Zhang, W. (2023).\n[\"Experimental Evidence on the Productivity Effects of Generative Artificial Intelligence.\"](https://doi.org/10.1126/science.adh2586)*Science*, 381(6654), 187–192. - Cui, K. Z., Demirer, M., Jaffe, S., Musolff, L., Peng, S., & Salz, T. (2026).\n[\"The Effects of Generative AI on High-Skilled Work: Evidence from Three Field Experiments with Software Developers.\"](https://doi.org/10.1287/mnsc.2025.00535)*Management Science*, published online February 27, 2026. - Peng, S., Kalliamvakou, E., Cihon, P., & Demirer, M. (2023).\n[\"The Impact of AI on Developer Productivity: Evidence from GitHub Copilot.\"](https://arxiv.org/abs/2302.06590)arXiv:2302.06590. Earlier, smaller predecessor to source 7 above; retained for historical context, not as primary evidence. - Becker, J., Rush, N., Barnes, E., & Rein, D. (2025).\n[\"Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity.\"](https://metr.org/blog/2025-07-10-early-2025-ai-experienced-os-dev-study/)METR. Full paper:[arXiv:2507.09089](https://arxiv.org/abs/2507.09089). - METR (2026).\n[\"We Are Changing our Developer Productivity Experiment Design.\"](https://metr.org/blog/2026-02-24-uplift-update/) - AHRQ PSNet,\n[\"Large Language Model Influence on Diagnostic Reasoning: A Randomized Clinical Trial\"](https://psnet.ahrq.gov/issue/large-language-model-influence-diagnostic-reasoning-randomized-clinical-trial)— clinical-safety digest corroborating source 2's trial design and results.", "url": "https://wpnews.pro/news/human-vs-ai-vs-human-and-ai-who-does-better-work", "canonical_source": "https://rafaelresearch.com/research/human-vs-ai-vs-human-plus-ai", "published_at": "2026-08-11 22:02:10+00:00", "updated_at": "2026-08-11 22:11:51.255249+00:00", "lang": "en", "topics": ["artificial-intelligence", "ai-research"], "entities": ["Microsoft 365 Copilot", "NBER", "Google Scholar", "Semantic Scholar", "PubMed", "SSRN", "arXiv", "ACM Digital Library"], "alternates": {"html": "https://wpnews.pro/news/human-vs-ai-vs-human-and-ai-who-does-better-work", "markdown": "https://wpnews.pro/news/human-vs-ai-vs-human-and-ai-who-does-better-work.md", "text": "https://wpnews.pro/news/human-vs-ai-vs-human-and-ai-who-does-better-work.txt", "jsonld": "https://wpnews.pro/news/human-vs-ai-vs-human-and-ai-who-does-better-work.jsonld"}}