AI 抬高下限；在大多数研究里，它几乎为所有人抬高下限，对最弱的用户尤其如此。同样的模型对任何人都可以买到。上限则是另一回事，而且它是个人性的：AI 会变成用户能够指定出来的样子。当目标只存在于用户那里时，这一点最明显——那是一套模型看不见、任何提示词库也给不了的标准。它发生两次：一次在工作被下达时，另一次在结果被采纳或搁置时。
Jahani, E., Manning, B.S., Zhang, J., TuYe, H.-Y., Alsobay, M., Nicolaides, C., Suri, S., and Holtz, D. “Prompt Adaptation as a Dynamic Complement in Generative AI Systems.” arXiv:2407.14333, v7, 2026 年 1 月 7 日修订（早期标题为 “As Generative Models Improve, We must Adapt Our Prompts”）。https://arxiv.org/abs/2407.14333。两项预注册、随机、盲法在线实验（复现任务 1,893 名参与者；logo 任务 1,857 名；合计 3,750 名）。v7 摘要把复现任务描述为“有固定评价标准和明确目标的任务”；发表版摘要称其为“有边界且目标清楚的任务”。logo 任务中的数字（超过 90% 来自模型、约 7% 来自提示适应）以及复现任务的拆分（约 51% 来自模型、约 48% 来自提示适应）来自 v7。任务使用图像、DALL-E 2 与 3（2022–23 年模型）以及付费在线样本，由机器评分（余弦相似度；LLM 评审）。这个拆分分解的是平均收益；它不测量用户之间的分布差异。用户适应项是对更好模型的观察到的回应，不是随机分配。第三组实验中 GPT-4 改写了参与者提示词，降低了复现任务收益（arXiv 图中为 58%）；正文没有纳入它，因为它测量的是改写器，而不是用户。 ↩↩2↩3
Jahani, E., et al. Information Systems Research，2026 年 4 月 30 日在线发表，DOI 10.1287/isre.2025.2029。发表版摘要（通过 Crossref 记录阅读）写到，“大约一半……来自模型本身，另一半来自用户自然调整提示词”，并说自动提示词改写在复现任务中“显著损害了表现，而在这个任务中，精确的用户控制最重要”。 ↩
Conger, J.A., and Kanungo, R.N. “The Empowerment Process: Integrating Theory and Practice.” Academy of Management Review 13(3), 1988。此处依据赋能研究的二手综述来描述：一种是权力分享的关系意义，另一种是后来的自我效能感的动机意义；本文没有阅读该论文原文，也没有引用其措辞。 ↩
Turing, A.M. “Computing Machinery and Intelligence.” Mind 59(236), 1950, 第 6 节，引用 Lovelace。阅读的是一份课程托管全文：https://courses.cs.umbc.edu/471/papers/turing.html。Lovelace 自己的文本见她 1843 年翻译 Menabrea 时所写的 Note G；二手来源给出的措辞是 “no pretensions whatever to originate anything”，其中有 “whatever”，Turing 的引用省略了这个词。 ↩↩2
Brynjolfsson, E., Li, D., and Raymond, L. “Generative AI at Work.” NBER Working Paper 31161, 2023 年 4 月，2023 年 11 月修订。https://www.nber.org/papers/w31161。数字以及“提示性证据”这句话来自 NBER 摘要。后来发表的版本（arXiv:2304.11771 v2）报告了 5,172 名客服、平均 15% 的提升以及低技能员工约 30% 的提升；用顶尖员工文本训练模型的描述见该 arXiv v2 正文第 2.3 节。单家公司、单一任务、交错上线、观察性研究；传播最佳实践是作者的假设，不是被检验过的机制。 ↩↩2
Otis, N., Clarke, R., Delecourt, S., Holtz, D., and Koning, R. Management Science, 2026 年 7 月 10 日，DOI 10.1287/mnsc.2024.06909；同行评议版本见 https://escholarship.org/uc/item/0fb001r8（摘要、第 3–4 页、第 22 页）。低表现者效应比高表现者效应低超过 0.20 个标准差。工作论文（HBS WP 24-042，2023 年 12 月）给出了低表现者“约 8%”的说法；那个数字只属于工作论文。统计功效有限，机制来自日志和调查，而不是随机化本身。 ↩↩2
AI 抬高下限；在大多数研究里，它几乎为所有人抬高下限，对最弱的用户尤其如此。同样的模型对任何人都可以买到。上限则是另一回事，而且它是个人性的：AI 会变成用户能够指定出来的样子。当目标只存在于用户那里时，这一点最明显——那是一套模型看不见、任何提示词库也给不了的标准。它发生两次：一次在工作被下达时，另一次在结果被采纳或搁置时。
用户就是 AI 的上限。这是一个论点，不是一项测量。我能找到的研究里，还没有哪一项把真实工作中的用户差异与模型差异放在一起，比较二者分别造成多大的结果分布。1 下限是 AI 给出的；上限是用户赋予 AI 的。
Haase, J., Gonnermann-Müller, J., Hanel, P., Leins, N., Kosch, T., Mendling, J., and Pokutta, S. “Within-Model vs Between-Prompt Variability in Large Language Models for Creative Tasks.” arXiv:2601.21339；CHI 2026 Extended Abstracts, “It's Not Just the Prompt: Model Choice Dominates LLM Creative Output,” DOI 10.1145/3772363.3799284。最接近的模型与提示词方差分解：在输出数量（流畅性）上，模型解释 51.25%，提示词解释 4.22%；在原创性上，二者份额相近（40.94% 与 36.43%）。提示词由研究者固定，而不是由用户选择，因此它测试的是提示词，不是用户。
Noy, S., and Zhang, W. “Experimental Evidence on the Productivity Effects of Generative AI.” Science 381, 2023, pp. 187–192。另一项随机化的下限抬高研究；是本文“共享下限”部分的背景。
Dell'Acqua, F., et al. “Navigating the Jagged Technological Frontier.” Organization Science 37(2), 2026；HBS Working Paper 24-013。顾问实验；AI 在能力边界内拉平技能，在边界外误导用户。
Vaccaro, M., Almaatouq, A., and Malone, T. “When combinations of humans and AI are useful.” Nature Human Behaviour, 2024。关于人机组合何时输给更强一方的元分析；是棋类段落中边界问题的背景。
Shen, J., and Tamkin, A. “AI assistance and coding skills.” Anthropic, 2026 年 1 月 29 日。同一个助手，在不同互动模式下出现不同结果；样本格较小，属探索性。
Riedl, C., and Weidmann, B. “Quantifying Human-AI Synergy.” PsyArXiv preprint, 2025。预印本证据：协作能力是一种不同于单独能力的特质。
AI raises the floor, and in most studies it raises it for nearly everyone, the weakest users most. The same models are on sale to anyone. The ceiling is another matter, and it is personal: the AI becomes what its user can specify. That holds most strongly where the target lives only with the user — a standard the model cannot see, and no prompt library supplies. It operates twice: when the work is ordered, and again when the result is taken up or set aside.
The user is the AI's ceiling. That is an argument, not a measurement. No study I could find sets the spread of outcomes across users against the spread across models on realistic work.1 The floor is something AI gives. The ceiling is something the user gives the AI.
Whether that distinction survives depends on one prior question: where does the target live?
The cleanest test splits a single model upgrade between the model and its users. Jahani and colleagues gave thousands of paid participants either DALL-E 2 or DALL-E 3. Assignment was random, and participants were not told which model they had.2 Then came two tasks.
In the first, a task the authors describe as having fixed evaluation criteria and an unambiguous goal, participants saw a target image and tried to reproduce it, prompt by prompt, over ten attempts. The criteria were fixed and the model could not see them; only the participant could steer toward them. The better model produced closer copies. But the users who received it also rewrote their own prompts, and those rewrites carried about half of the upgrade's gain. The authors describe this task as the one "where precise user control mattered most."3
In the second, participants designed a logo to an open brief, where many answers would do. Here the model carried more than 90% of the gain.
Side by side, the two results differ in one variable. In the first task the target lived with the participant, out of the model's sight, so only the participant could steer toward it. In the second it lived nowhere in particular; any good logo would be accepted. When the target sits with the user, the AI becomes what the user can specify. When any good output will do, the model's improvements arrive without the user.
Something happened in that first task that deserves a better word than "prompting."
The word has two senses, and the argument turns on telling them apart.
The first is the vendor's. Capability flows from tool to person: AI empowers me. I grant it entirely: it is the floor, and it is real. As the empowerment literature summarises it, management writing has used the word both for a leader sharing authority with subordinates and, since the late 1980s, for raising people's sense of their own efficacy.4 The vendor sense borrows the second.
The other is older and plainer: "to give power or authority to," as the word's record from the 1650s has it.5 To empower someone is to give them the authority and the brief to act on one's behalf. In that sense the user empowers the model, by supplying the objective, the standard, the context and the permission to proceed. Execution is handed over; purpose is not.
Every principal delegates to every tool in some thin sense, so the sense needs a test: does what comes back depend on what the brief contains? A hammer takes nothing from its owner's brief. A model does, and in Jahani's replication task that dependence carried about half the gain.2 The inversion therefore claims capability, not just permission: what the brief carries is something the model lacks. In vendor slogans the word almost always points from the platform to the person. The inversion points it back.
Ada Lovelace said as much in 1843, in words Turing quotes in 1950: the Analytical Engine "can do whatever we know how to order it to perform."6 The weight falls on know how to order. Turing's reply is the honest limit, and the first clause of its second sentence carries the point: "Machines take me by surprise with great frequency. This is largely because I do not do sufficient calculation to decide what to expect them to do."6 Even his surprise came from an orderer who had not worked out all that an order contained. The ceiling sits where the orderer's foresight does.
Two arrows, then: one from tool to user, one from user to tool, at different heights. The floor is the first. The ceiling is the second. The first is the place to start.
The first arrow is the best-measured of the two. Brynjolfsson, Li and Raymond studied a generative-AI assistant rolled out to 5,179 customer-support agents. Productivity rose 14% on average and 34% for novice and low-skilled agents, with "minimal impact" on the most experienced.7 That is the vendor sense, measured, and I agree with it.
The assistant, though, had been trained on text from the firm's top performers. The authors offer "suggestive evidence that the AI model disseminates the best practices of more able workers."7 The arrow runs expert, then model, then novice. The novice did not empower the AI. But what the floor gives was first specified by someone. The floor is shared, and it was built from somebody's ceiling.
Above the floor, the tool stops treating its users alike, and the floor itself is not guaranteed.
In a five-month randomized trial of 640 Kenyan entrepreneurs, half received a GPT-4 business mentor on WhatsApp. The average effect was indistinguishable from zero. That average hid two opposite signs: entrepreneurs who had been performing poorly at the start did "nearly 10% worse," while high performers "may have benefited by over 15%."8
The published paper finds no evidence that the gap came from the questions people asked or the advice the AI gave. It appears to lie in which advice each entrepreneur chose to act on. High performers worked with the AI on tailored, context-specific changes; low performers disproportionately adopted generic moves such as cutting prices and advertising more.
The reading needs care. Starting performance bundles capital, market knowledge and room to act, and it is not the same as being a better user. A mentor with no local context hands out generic advice, and generic advice hurts most those least able to filter it. The authors "cannot fully rule out alternative explanations."8 Where the ceiling sits is one question and who is at fault is another, and this trial speaks, at most, to the first.
Within those limits, the choice is where the target shows. What the high performers could bring to it was knowledge of their own markets: which change fits this shop, this street, these customers. That is a target the AI could not see, and it bore on which of the same advice was worth acting on. This is the delegation sense at the receiving end. The AI said much the same to everyone — and the brief the owner held decided what became of it.
Anthropic's data measure something narrower: the sophistication of the prompt text against the sophistication of Claude's response, across 117 countries (r = 0.925). Anthropic writes that Claude tends to respond in a highly sophisticated manner "only when users input sophisticated prompts."9 That is a correlation across countries, not individuals, and harder tasks demand both.
Everything above holds the model fixed. The strongest objection is that the model is what does not stay fixed.
Give it full weight: models differ, the model term is causal, and it moves on a clock no user can match. METR's January 2026 update reads: "The post-2023 doubling-time is 131 days," for the length of task that frontier agents can complete.10 The older estimate, covering 2019 to 2025, was about seven months.11
The randomized evidence agrees. In one experiment, Ali Merali assigned 300 professional translators across 13 language models, and each tenfold increase in model compute made them 12.3% faster. "Gains were four times larger for lower-skilled workers."12 That is the floor again, now with the causal arrow running from the model.
All of this compares models. None holds the model fixed and asks what a user adds on a target only the user holds. At any one moment the same models are on sale to anyone, so the model term is the floor rising, the vendor sense at speed. What a user gets is set by whichever binds first: the model's capability or the user's specification. Model progress lifts the first for everyone at once. On a target the model cannot read, the second binds sooner, and it has to be re-supplied at each upgrade.
Two places have looked at that second term more directly. The first is Jahani's own result. The half that users carried was a response to the better model, unlocked by it and paid again at the upgrade.2 A faster model term makes the user's part come due more often. It does not retire it.
The second is a later preprint by Merali: a randomized experiment with more than 500 consultants, analysts and managers, in which each year of model progress was associated with about an 8% fall in task time. On the non-agentic tasks, the quality of work that people produced with the AI stayed flat across model generations, while the quality of what the models produced alone kept rising. Merali writes that "human users effectively cap the realized capabilities of frontier models, satisfying to a fixed quality threshold rather than maximizing the tool's potential."13 The result covers non-agentic tasks only, and expert peers, not the users, graded the work. For the strongest models, the paper reports, participants degraded the raw output. Where the user aims only at good enough and holds no target of their own, the user is still the ceiling — only a low one, set beneath what the model could reach alone. Chess is where that boundary has been walked.
In 2005 a freestyle chess tournament was won, Kasparov wrote, by "a pair of amateur American chess players using three computers at the same time. "His formula: "Weak human + machine + better process was superior to a strong computer alone and, more remarkably, superior to a strong human + machine + inferior process."14
After AlphaZero in 2017, the common claim became that the human "now adds absolutely nothing" to such teams. That line is Tyler Cowen's commentary, widely repeated and not measured.15
Two kinds of user contribution sit inside that history. The first is compensating work: it makes up for what a weak model lacks, such as which engine to trust and when. The amateurs' better process was of this kind, and it expires as the model improves; by 2017, on Cowen's account, it had. The second is target-holding work: supplying a target the model cannot read. Chess never needed it, because winning is written in the rules and checked by the machine. Model progress retires the first kind. The ceiling argued for here is the second kind only — and it lasts as long as the target stays out of the model's sight. My interpretation is that much of the work people care about (a house standard, the right answer for this client, an essay that says what its writer means) has a target of that kind, living in a head and not in a rulebook.
That second kind has an exit. Writing a target down turns one person's ceiling into the next person's floor, which is how the support agents' floor came to be. The ceiling is the target not yet written, and only its holder can write it.
The claim can be tested at small scale. When the next model ships, keep two lines in a log: one task whose target only its owner can see, and one where any good output will do. Run each with the old prompt, then with a re-specified one, and note how much of the gain needed the re-specifying. Run each more than once and score blind, since a single run can be swamped by noise. The gap between the two lines is a personal version of Jahani's number. Writing the target down is the practical step, and Why Serious LLM Use Needs a North Star Prompt describes one kind of brief that does it.
The floor will keep rising, and it will arrive for everyone at once. That is its virtue. What it cannot bring is the knowing of what this user wants ordered. The engines have changed beyond recognition since Lovelace; her sentence has not. The floor is what the machine gives, and every target written down joins it. The ceiling is whatever the user knows how to ask of it, and has not yet said.
From a search of arXiv, institutional corpora and the general web: no study found that holds a realistic non-creative task fixed, varies both users and models, and partitions outcome variance between them. "None found," not "none exists. "Nearest studies: Jahani et al. (users against a model upgrade, one task family); Haase et al. (researcher-fixed prompts, not users; see Further Reading); Otis et al. (one model, divergent users, no model comparison). ↩
Jahani, E., Manning, B. S., Zhang, J., TuYe, H.-Y., Alsobay, M., Nicolaides, C., Suri, S., and Holtz, D. "Prompt Adaptation as a Dynamic Complement in Generative AI Systems." arXiv:2407.14333, v7, revised 7 Jan 2026 (earlier title: "As Generative Models Improve, We must Adapt Our Prompts"). https://arxiv.org/abs/2407.14333. Two preregistered, randomized, blind online experiments (replication task, 1,893 participants; logo task, 1,857; 3,750 in all). The v7 abstract describes the replication task as "a task with fixed evaluation criteria and an unambiguous goal"; the published abstract calls it "a bounded task with a clear objective. "Figures for the logo task (more than 90% model, about 7% prompt adaptation) and the replication split (about 51% model, about 48% prompt adaptation) are from v7. The tasks used images, DALL-E 2 and 3 (2022–23 models) and a paid online sample, scored by machine (cosine similarity; an LLM judge). The split decomposes a mean gain; it does not measure spread between users. The user-adaptation term is an observed response to the better model, not a randomized assignment. A third arm, in which GPT-4 rewrote participants' prompts, reduced the replication-task gain (58% in the arXiv figures); it is left out of the text because it measures the rewriter, not the user. ↩↩↩
Jahani, E., et al. Information Systems Research, online 30 Apr 2026, DOI 10.1287/isre.2025.2029. The published abstract (read via the Crossref record) gives "roughly half ... from the model itself, and half from users naturally adjusting their prompts," and says automatic prompt rewriting "substantially hurt performance in the replication task, where precise user control mattered most." ↩
Conger, J. A., and Kanungo, R. N. "The Empowerment Process: Integrating Theory and Practice."Academy of Management Review 13(3), 1988. Described here as summarised in secondary accounts of the empowerment literature (the relational sense of power-sharing, and the later motivational sense of self-efficacy); the paper itself was not read, and no wording is quoted from it. ↩
Online Etymology Dictionary, "empower." https://www.etymonline.com/word/empower. First sense, from the 1650s: "to give power or authority to, authorize as by law." ↩
Turing, A. M. "Computing Machinery and Intelligence."Mind 59(236), 1950, section 6, quoting Lovelace. Read in a course-hosted full-text copy, https://courses.cs.umbc.edu/471/papers/turing.html. Lovelace's own text is in Note G of her 1843 translation of Menabrea; secondary sources give her wording as "no pretensions whatever to originate anything," with "whatever," which Turing's quotation omits. ↩↩
Brynjolfsson, E., Li, D., and Raymond, L. "Generative AI at Work. "NBER Working Paper 31161, Apr 2023, rev. Nov 2023. https://www.nber.org/papers/w31161. Figures and the "suggestive evidence" sentence are from the NBER abstract. The later published version (arXiv:2304.11771 v2) reports 5,172 agents, 15% on average and about 30% for lower-skilled workers; the training on top performers' text is described in that arXiv v2 text (§2.3). One firm, one task, staggered rollout, observational; the dissemination account is the authors' hypothesis, not a tested mechanism. ↩↩
Otis, N., Clarke, R., Delecourt, S., Holtz, D., and Koning, R.Management Science, 10 Jul 2026, DOI 10.1287/mnsc.2024.06909; peer-reviewed copy at https://escholarship.org/uc/item/0fb001r8 (abstract; pp. 3–4; p. 22). The low-performer effect is more than 0.20 standard deviations lower than the high-performer effect. The working paper (HBS WP 24-042, Dec 2023) gave "about 8%" for low performers; that figure belongs to the working paper only. Power is limited, and the mechanism comes from logs and surveys, not from randomization. ↩↩
Anthropic. "Anthropic Economic Index: January 2026 report." https://www.anthropic.com/research/anthropic-economic-index-january-2026-report. Sophistication is measured as the years of education needed to understand the prompt and the response. For US states, r = 0.928 (N = 50). Prompt sophistication also tracks task difficulty, so the correlation does not isolate user input from the task; the quoted sentence is Anthropic's interpretation, from a vendor. ↩
METR. "Time Horizon 1.1," 29 Jan 2026. https://metr.org/blog/2026-1-29-time-horizon-1-1. Task length is measured at 50% reliability on METR's task suite; the interval is 131 days [107, 161], and the doubling time since 2024 is 89 days. Part of the speed-up comes from re-estimating older models downward, and few of the longest tasks have human baselines. ↩
Merali, A. "Scaling Laws for Economic Productivity: Experimental Evidence in LLM-Assisted Translation." arXiv:2409.02391, v2, 7 Dec 2024. https://arxiv.org/abs/2409.02391. Preprint, preregistered, single author; tasks of about ten minutes. "Lower-skilled" is a median split on completion time in a baseline task done without AI. ↩
Merali, A. "Scaling Laws for Economic Productivity: Experimental Evidence in LLM-Assisted Consulting, Data Analyst, and Management Tasks." arXiv:2512.21316, v1, 24 Dec 2025. https://arxiv.org/abs/2512.21316. Single-author preprint; tasks averaged about 26 minutes in standard chatbot interfaces. The flat human-assisted line is for non-agentic tasks (p ≈ 0.85); the 8% annual figure is "associated with" model release date. Sec. 4.4 reports that for the strongest models "participants actively degrade" the raw output; outputs were graded on a seven-point scale. ↩
Cowen, T. "The age of the centaur is over."Marginal Revolution, 7 Dec 2017. https://marginalrevolution.com/marginalrevolution/2017/12/the-age-of-the-centaur-is-over.html. A 2024 modelling paper (Shoresh and Loewenstein, arXiv:2412.18593) finds high potential for human-machine synergy in chess, and Kasparov (Science, 7 Dec 2018) continues to argue for partnership; the empirical state of centaur chess after 2017 is not independently measured. ↩
Haase, J., Gonnermann-Müller, J., Hanel, P., Leins, N., Kosch, T., Mendling, J., and Pokutta, S. "Within-Model vs Between-Prompt Variability in Large Language Models for Creative Tasks." arXiv:2601.21339; CHI 2026 Extended Abstracts, "It's Not Just the Prompt: Model Choice Dominates LLM Creative Output," DOI 10.1145/3772363.3799284. — The nearest variance partition of model against prompt: on output quantity (fluency) the model explains 51.25% and the prompt 4.22%; on originality the shares are comparable (40.94% and 36.43%). The prompts were fixed by the researchers, not chosen by users, so it tests the prompt, not the user.
Noy, S., and Zhang, W. "Experimental Evidence on the Productivity Effects of Generative AI."Science 381, 2023, pp. 187–192. — A second randomized floor-raising study; background for the shared-floor half of the argument.
Dell'Acqua, F., et al. "Navigating the Jagged Technological Frontier."Organization Science 37(2), 2026; HBS Working Paper 24-013. — Field experiment with consultants; AI levels skill inside the frontier and misleads outside it.
Vaccaro, M., Almaatouq, A., and Malone, T. "When combinations of humans and AI are useful."Nature Human Behaviour, 2024. — Meta-analysis of where human-AI combinations lose to the better party alone; background for the boundary in the chess section.
Shen, J., and Tamkin, A. "AI assistance and coding skills. "Anthropic, 29 Jan 2026. — Same assistant, divergent outcomes by interaction pattern; small cells, exploratory.
Riedl, C., and Weidmann, B. "Quantifying Human-AI Synergy. "PsyArXiv preprint, 2025. — Preprint evidence that collaboration ability is a trait separate from solo ability.
