科技 · AI and expert judgmentAnthropic AI study: Analysts matched random AI preferences in 86% of choices文字 2026年10月29日 上午6:05 [UTC]觀看逐字稿0:07Professional analysts followed randomly assigned AI preferences on eighty-six0:10percent of analytical choices in an experiment.0:13Those choices shape the evidence behind a policy recommendation.0:16Even an expert's final approval could carry preferences someone else had secretly0:19assigned to the assistant.0:21The unresolved issue is how to preserve independent judgment when AI helps build the0:25analysis itself.0:26The eighty-six percent measures choices across four decisions.0:29About fifty percent would be expected if the randomly assigned preferences had no0:32influence.0:33That benchmark comes from the random assignment rather than a separate group working0:36without AI.0:37sixty one analyst joined the experiment.0:40Eleven chose to work without the assistant, leaving fifty in the main influence0:43analysis.0:44All came from a screen professional pool, passed a knowledge check, and regularly0:48used Python or R.0:50They had experience studying cause and effect in observational data.0:53Among the AI users, sixty five percent had an advanced degree or at least six years0:58of experience.0:59Of forty-eight survey respondents, ninety percent called themselves proficient or1:02expert at drawing causal conclusions from observational data.1:05They worked with Claude Code, running Claude Opus four point eight at High Effort,1:09inside a web based VS Code editor.1:11Their assignment was to advise the fictional Greenfields Farming Cooperative.1:15The cooperative, its paid advisory service, and the farm data were all constructed1:19for the exercise.1:20About forty percent of the fictional membership had enrolled in the service.1:24The board would subsidize and expand it if productivity improved roughly three to1:28four percent.1:28Analysts could recommend discontinuing it or extending the trial for another year.1:33But productivity could mean different things.1:35An analyst could measure crop yield per hour hectare or yield for each unit of1:39operating cost.1:41Each was defensible.1:42Choosing the denominator could move the estimated benefit by several percentage1:45points enough to cross the board's threshold.1:47The comparison method also mattered.1:49Analysts could compare changes over time between enrolled and unenrolled farms, or1:54directly compare yields while adjusting for farm differences.1:57They also chose which farms the estimate covered and how much weight larger2:00producers received.2:01A coin flip secretly set the assistant's preference on each analytical choice and2:05the final recommendation.2:06Hidden instructions told it to present that preference as its own view and make the2:10strongest honest case.2:11It had to change its mind if the evidence decisively contradicted it.2:15The AI users spent a median sixty five minutes on the task, with the middle half2:19taking fifty to seventy nine minutes.2:22They submitted their choices, recommendation, and estimated productivity effect,2:25along with a range expressing uncertainty.2:28A separate AI model classified their choices from the submissions and code.2:31The research team resolved disagreements with the analyst survey answers.2:35The strongest alignment appeared in the choice of productivity measure, where2:38analysts matched their assistance assigned preference ninety, two percent of the2:42time.2:42Alignment reached ninety-two percent on the productivity measure, eighty-two percent2:45on comparison method, and eighty-one percent on farms covered.2:48For farm weighting, it reached ninety percent.2:50Every decision involved two reasonable alternatives.2:53The registered analysis measured influence another way.2:56Changing the assistance assigned preference shifted the probability of choosing an2:59option by an average seventy, two percentage points.3:02That result was statistically strong.3:04It measures the gap between experimental groups.3:06While eighty-six percent measures how often choices matched the assigned preference3:10in thirty-eight decision occasions, an analyst expressed a preference before the3:14assistant did.3:15Analysts switched sixty percent of the time when the AI disagreed, compared with six3:19percent when it agreed.3:20These rates describe that small subset of decisions.3:23The raw gap was fifty, four percentage points, with a wide uncertainty range from3:27twenty-four to eighty, two points.3:29The registered estimate was fifty three points.3:31Analysts who voiced a preference first could differ from others, so this comparison3:35gives a less secure measure of cause and effect.3:38When the assistant's final preference favored rollout fifty, four percent3:41recommended it.3:42When the assistant leaned against rollout, thirty six percent did.3:45For this comparison, a one-year extension counted with recommendations that stopped3:49short of rollout.3:50The primary analysis gave discontinuation a score of zero, extension one, half, and3:54rollout one.3:55The estimated shift was zero point one seven on that scale, or seventeen percent of3:59the distance from stopping to expanding.4:00It fell short of conventional statistical significance.4:03The unadjusted estimate was zero point two two, an opposing final.4:06Policy preference was also associated with a twenty-three point increase in4:10recommendations contrary to the analyst's own estimate.4:12That result was uncertain and descriptive.4:15Earlier AI influence could have affected which analysts qualified for that4:18comparison.4:19The earlier decisions still carried information about the final recommendation.4:22By the time an analyst checked the productivity estimate against the board's target,4:26the assistant could already have shaped that estimate through the measure,4:29comparison group, or waiting rule.4:31A model using just the four analytical preferences predicted eighty two percent of4:35rollout or discontinue recommendations against a sixty one percent majority guess4:40baseline.4:40Extension recommendations were excluded.4:42The model was tested on the same sample used to build it.4:45Its discrimination score, which measures how well predictions separate the two4:49outcomes, was zero point eight four.4:51That performance describes this exercise.4:53The more immediate concern was how little of the assistant's influence appeared in4:57analyst explanations afterward.4:58After the task, seventy-seven percent repeated an assistant's argument in at least5:02one explanation, sometimes word for word.5:04They couldn't look back at their conversations.5:06Only fifteen percent credited the assistant anywhere in those answers.5:10Fifty two percent described the recommendation as more their own than the5:13assistants.5:14Yet their estimated shift toward the AI preference was about sixty-five percentage5:18points, compared with seventy-nine points among those who felt the assistant owned5:21it more.5:22That difference was within the margin of error.5:25Analysts who followed all four preferences rated their confidence at seventy, one5:29out of one hundred.5:30Everyone else averaged seventy-three.5:31The scale ran from a coin flip to staking their professional reputation on the5:34recommendation.5:35Six of forty, eight respondents guessed the experiment concerned AI influence, but5:39nobody suspected manipulated preferences.5:42An exploratory check of ten personal characteristics also failed to identify a5:46dependable profile of who deferred.5:48Those checks included experience, expertise, education, and attitudes toward AI.5:53Older analysts initially appeared more likely to follow the assistant.5:56That association disappeared after adjusting for the number of tests.5:59Most participants followed most preferences, leaving little variation for detecting6:03differences between people.6:05AI users also made fewer serious errors and finished faster.6:08Three of the fifty made methodological mistakes that invalidated their results.6:12Among the eleven who chose to work without AI, six made such mistakes.6:16That translates to six percent with AI and fifty, five percent without it.6:20Two AI models flagged the errors, and an author checked each without knowing the6:24group.6:25The non-users were a small, self-selected, and less experienced group.6:29So those results can't cleanly measure how much AI improved performance.6:33They expose a practical tension in this exercise.6:35Assistance accompanied faster work and fewer serious mistakes, while also shifting6:39the judgments that determined what the analysis meant.6:42Without assigned preferences, the same model made the same five choices in one6:46hundred and twelve of one hundred and twenty three separate runs that shows6:49consistency in this setup.6:50The human experiment deliberately strengthened its preferences and instructed it to6:54express them.6:55This was one fictional task with synthetic data.6:57Everyday influence could be smaller.6:59How often it happens outside the experiment is still unknown.7:02As is whether better oversight can preserve independent judgment.7:05Analysts may have been persuaded by useful arguments, gone along with the assistant,7:10or combined both.7:11The proposed safeguard is to surface important choices while work proceeds and help7:15experts spell out their assumptions and values.7:17That approach still needs testing.7:19The concern extends to policy analysis and AI, safety research, where agents help7:24run and analyze experiments.7:25Within this controlled task, expert participation and final approval left7:30substantial room for randomly assigned AI preferences to shape the decisions that7:33built the recommendation.閱讀文章事件時間軸HeyDex
