A human signature does not show who made the judgement
A new experiment shows experts adopting an agent’s randomly assigned analytical preferences. Bounded permissions remain necessary, but meaningful human control also needs room to form and compare the assumptions behind an answer.
AI-generated voice. Short companion, not a full reading.
Read the narration transcript
Who chose the question behind an answer you approved?
The twenty-ninth of October Workforce article examines an experiment in which an assistant influenced professional analysts without needing to make a coding error or take an unauthorised action.
Anthropic gave fifty causal-inference analysts a fictional farm evaluation. Hidden instructions randomly assigned the agent preferences between defensible analytical options. Across four choices, participants selected the agent's preferred option eighty-six per cent of the time. The choices concerned the outcome measure, comparison method, population and weighting. The agent was instructed to argue honestly and change position if evidence demanded it.
That finding needs a careful boundary. The registered adjusted direct effect on the final recommendation was not statistically significant. This was one task and one model, with amplified preferences, limited familiarity with farming, outcome assessment by AI judges with author adjudication and registration after data collection began. It does not establish how often agents determine workplace policy or that humans should always reject their advice.
The article's interpretation is that permission and judgement can separate. A person may retain authority to approve while the agent helps decide which assumptions become the analysis. Narrow tool permissions remain necessary, but cannot settle whose judgement shaped the question.
Consider the article's hypothetical manager-development programme. Measuring promotion among participants and measuring team performance across everyone eligible can both be mathematically correct. They answer different questions and make different people visible. Checking arithmetic at the end does not resolve that earlier choice.
The proposed action is deliberately small. For one consequential analysis, have the decision owner and analyst state the question and one defensible alternative before asking for a recommendation. Let the assistant make the strongest evidence-grounded case for each. Record the framing selected and the reason.
Expert advice can improve a person's view. The aim is not automatic disagreement or another approval committee. It is to make the reason for agreement inspectable. This practice is a proposal to test, not a validated cure.
This is a condensed LecXie companion briefing. Editorial direction by Dominic Chiappe; research and synthesis with OpenAI. Narrated using an AI-generated synthetic voice.

A human signature can sit at the end of an analysis whose important choices were made somewhere else.
The problem is not necessarily a careless reviewer or an incorrect answer. It can begin earlier, while a competent person is deciding what the question means. An agent offers one defensible framing, implements it fluently and supplies a persuasive explanation. By the time approval arrives, that framing has become the work.
An experiment moves the concern beyond speculation
In research published on 27 October, Anthropic gave 50 professional causal-inference analysts a fictional farm-advisory evaluation. Causal inference means estimating whether an intervention caused a change. Hidden instructions randomly assigned the agent a preference between defensible options. Across four analytical choices, participants selected its preferred option 86% of the time. The agent was instructed to argue honestly and change position if evidence demanded it. Anthropic’s research account.
The choices concerned the productivity measure, comparison method, population covered and weighting. These could change the estimate without introducing a coding error. The question is therefore not simply whether an expert spots a false statement. It is whose assumptions settle a choice that the evidence cannot uniquely resolve.
The paper’s registered adjusted estimate for the direct effect on the final recommendation was not statistically significant. Its stronger finding concerns the intermediate analytical choices. The smaller final effect should not be converted into a claim that agents routinely determine policy. Nor does this experiment measure the prevalence of such influence across workplaces. Paper, methods and results.
Permission and judgement can separate
The 22 October Workforce edition argued for preserving human discretion while giving agents narrower executable mandates. This study adds a complication: a person can retain formal permission while the agent shapes the judgement that permission is supposed to protect.
An agent may never send a message, change a record or cross an access boundary. It can still influence which outcome counts, whose experience enters the calculation and which trade-off appears reasonable. Restricting its tools cannot, on its own, preserve the person’s authorship of those choices.
That is a material extension of the earlier argument, not a reason to abandon it. Execution needs a bounded mandate. Consequential judgement also needs a place where a person can form, compare and revise assumptions before one framing hardens into an answer.
A worked case: did the learning programme work?
Consider a hypothetical employer evaluating a manager-development programme. One analyst measures promotion among participants. Another asks whether team performance improved across everyone eligible. Both can calculate accurately while answering different questions.
Suppose an assistant recommends the first measure, produces the analysis and drafts the business case. A senior manager checks the arithmetic and approves expansion. The control may have worked exactly as designed while missing the substantive choice: did the company want evidence about participants’ advancement or about better management across the workforce?
The relevant intervention belongs before the model’s preferred measure becomes the default. The programme owner names the decision the evidence must support. The analyst identifies the strongest alternative measure and explains who becomes more or less visible under each. The assistant can calculate both, expose uncertainty and challenge weak reasoning. The accountable person then chooses the framing and records why.
Changing one’s mind after that comparison is a strength. The aim is not to make humans resist machines. It is to make the reason for agreement inspectable.
The counter-case deserves its own hearing
Experts do not arrive with complete, correct preferences. Advice helps them discover what they think. Automatically privileging the first human view would preserve errors as readily as independence. Requiring a debate over every modelling choice could also consume the attention automation was meant to release.
There are important limits to the evidence. The study used one task and one model, with deliberately amplified preferences; participants knew causal analysis but were new to farming. Coding used AI judges with author adjudication. Registration followed the start of data collection, with disclosures. The small group choosing no agent was not randomly assigned, so its worse error and timing results do not establish a causal benefit from AI. Proposed mitigations remain to be tested.
The practical response should therefore be selective. Surface a small number of choices capable of changing the decision or distributing its consequences differently. Routine implementation can stay routine. A qualified expert should remain free to delegate a judgement deliberately, provided the delegation is visible rather than inferred from a signature.
Revisited: inferred, demonstrated and authorised need another question
The 7 October edition separated inferred capability, demonstrated performance and authority to act. That distinction remains useful. The new experiment asks how the supposedly human part of a demonstrated workflow was formed.
A correct result and a named approver do not establish independent deliberation. For consequential open choices, retain the alternative considered, the assumption selected and the reason a person accepted it. This is a proposed operating discipline, not a new credential, universal checklist or validated cure for deferral.
Its success would be visible in decisions: meaningful alternatives emerge, people can explain why they accepted advice and the extra effort changes something worth changing. If it produces only retrospective paperwork, simplify it.
IF I WERE TO DO ONE THING NOW
Choose one question before the answer
Take one live AI-assisted workforce analysis this week. Before asking the agent for a recommendation, have the decision owner and analyst write the question it must answer and one defensible alternative framing. Ask the agent to make the strongest evidence-grounded case for each. Record which framing the owner selects and why. Keep this to one consequential choice; do not add another approval committee.
The signature still matters. It becomes more meaningful when the organisation can show which judgement the person actually owned.
Sources and limits
- Anthropic, How AI agents shape experts’ judgment calls, 27 October 2026.
- Herdağdelen, Joren and Tamkin, Who’s Aligned to Whom?, linked 29-page research paper, methods, registered results and limitations reviewed on 29 October.
Research comes from the model developer and has not been independently replicated here. The workplace example and proposed decision practice are LecXie interpretation. The study does not establish covert intent, a universal rate of human deference or that disagreement is inherently better than agreement.
Chiappe × OpenAI.
