MIT randomized professional-writing study
Verified evidenceAutonomous + backstopQueue eliminationCreator to judgeThe redesign is a shorter drafting loop with human revision and accountability, not autonomous publication.
Engineering design · Online experiment
Collections: Queue eliminated · Embodied work
Executive brief
Writing acceleration is defensible only when the accountable author owns verification and submission.
AI value · Task time and evaluator-rated quality
VerifiedAverage time decreased 40%; output quality increased 18%
Short incentivized tasks, GPT-3.5, and experimental rather than embedded enterprise work; no measurement of confidential-data risk.
Before
Drafts a press release, report, analysis plan, or sensitive email from an occupation-specific prompt. → Edits and submits the final response.
After
Prompts the model for a draft or revision during the same occupation-specific task. → Checks, edits, and submits the final work under their own responsibility.
Human boundary
The model proposes text; the professional owns prompts, edits, factual review, and final submission.
Why it matters
The redesign is a shorter drafting loop with human revision and accountability, not autonomous publication.
Before
How the work ran before the change.
College-educated professional
Drafts a press release, report, analysis plan, or sensitive email from an occupation-specific prompt.
ControlIncentivized task; control average 27 minutes.
Professional
Edits and submits the final response.
ControlHuman owns submission; evaluators grade quality anonymously.
What changed
The redesign is a shorter drafting loop with human revision and accountability, not autonomous publication.
Decision rightAI acts within a human backstop
After
How the same work runs now.
Professional using ChatGPT
Prompts the model for a draft or revision during the same occupation-specific task.
ControlTool may generate text, but participant decides whether and how to use it.
Professional
Checks, edits, and submits the final work under their own responsibility.
ControlHuman submission and blind evaluation; experiment did not measure enterprise confidentiality or factual-error incidents.
Exception path
Unsuitable, inaccurate, or sensitive text is rejected or rewritten by the professional; the study did not quantify this failure queue.
Work removed
Decision authority
The model proposes text; the professional owns prompts, edits, factual review, and final submission.
Before
College-educated professional
Drafts a press release, report, analysis plan, or sensitive email from an occupation-specific prompt.
Control: Incentivized task; control average 27 minutes.
Professional
Edits and submits the final response.
Control: Human owns submission; evaluators grade quality anonymously.
After
Professional using ChatGPT
Prompts the model for a draft or revision during the same occupation-specific task.
Control: Tool may generate text, but participant decides whether and how to use it.
Professional
Checks, edits, and submits the final work under their own responsibility.
Control: Human submission and blind evaluation; experiment did not measure enterprise confidentiality or factual-error incidents.
Work that left the path
Human role before
Study participants read an occupation-specific prompt, drafted and revised the response without ChatGPT, and submitted it for independent evaluation.
Human role after
Professionals direct the model, revise its output, and remain accountable for submission.
AI roleGenerates and revises candidate professional text.
Control average 27 minutes→Average time decreased 40%; output quality increased 18%
Preregistered online experiment in 2023 · 453 college-educated professionals
Short incentivized tasks, GPT-3.5, and experimental rather than embedded enterprise work; no measurement of confidential-data risk.
Anti-pattern
Do not project a short-task experiment directly onto confidential or high-stakes enterprise writing.
Questions
Portability conditions
Reputation risk
medium
Evidence and authority
Current · updated
1 independent, 1 peer reviewed; publication outcomes are verified.
Bundle 1.0.0 · reviewed 2026-08-23 · stable ID d06b7c157117b939
Swiss Plasma Center at EPFL and Google DeepMind · Autonomous + backstop
Operators set plasma objectives while a simulator-trained policy coordinates all coils on physical hardware.
Google · Autonomous + backstop
That AI should only recommend cooling set-point changes for operators to implement manually.
Lawrence Berkeley National Laboratory and University of California, Berkeley · Autonomous + backstop
That scientists must manually choose every recipe, execute every synthesis step, interpret each XRD result, and plan the next experiment.
National Grid Electricity Transmission · Autonomous + backstop
Central BVLOS drones/AI automate routine pylon imaging and corrosion grading; engineers decide maintenance.