Artificial intelligence is supposed to do one thing above all in projects: preliminary work. It sifts through material, summarises it, suggests variants. The decision, according to the common assumption, remains with humans. An experiment by Anthropic calls exactly this assumption into question.

The study examined how strongly an AI agent — that is, an assistance system that independently carries out analysis steps and formulates recommendations — influences the judgement of experts. Involved were 50 experts who were to work on open analysis questions: questions, that is, for which there is no clearly correct solution, but rather a reasoned weighing of options.

Coin toss instead of facts

The decisive trick of the setup: the AI agent's inclination was determined by coin toss. The recommendation was therefore not the result of a superior analysis, but simply chance. Nevertheless, the experts followed this assigned direction in 86 per cent of cases.

Even more revealing is the manner of adoption. The participants frequently took up the AI's arguments almost word for word. So no independent justification emerged that happened to arrive at the same conclusion, but rather a far-reaching adoption of the pre-formulated line of argument.

And the third finding weighs most heavily for practice: the majority ultimately regarded the result as their own decision. The participants were not aware of the foreign origin of the judgement.

Why this is not a convenience problem

The obvious interpretation would be that someone here took a shortcut. But the effect describes something more fundamental: an anchor effect. Anyone presented with a fully formulated position examines it — and precisely no longer examines the entire solution space with an open mind. The first plausible variant becomes the reference point against which all others are measured.

AI agents reinforce this mechanism for two reasons. Firstly, they deliver not only a result but also a fully formulated justification in a confident tone. Secondly, this happens immediately and without friction — before an independent position could even form.

The fact that the recommendation in the experiment came about by chance is the actual point. The effect operates regardless of whether the proposal is well justified. A weak recommendation therefore carries just as far as a strong one, as long as it is persuasively formulated.

Where this surfaces in projects

For daily work in web and software projects, three situations are particularly affected:

  • Code reviews with AI support: When an assistance system has already marked what it considers critical, the reviewing person's attention narrows to precisely these points. What was not marked tends to be treated as unproblematic.
  • Architecture decisions: Questions about framework, interface cut or hosting model are classic open analysis questions. There are several defensible paths with differing downstream costs. This is exactly where a pre-formulated proposal has the strongest effect.
  • Conception and requirements clarification: When the first draft of a process description or an information architecture comes from a model, it is as a rule commented on and not replaced.

What is dangerous is less the individual case than the documentation trail it leaves behind. A decision that appears in the minutes as professionally weighed may in fact have been the adoption of a proposal. Later reviews then find no point of attack, because the justification is already there.

What can be derived organisationally

The consequence is not to do without AI support. It lies in changing the sequence and the visibility:

  1. Own position first. For decisions with longer-term binding effect — architecture, data model, platform choice — it is worth recording one's own assessment before consulting the model. If the results diverge, that is a useful occasion for review.
  2. Request variants instead of a recommendation. Anyone who asks for several options with pros and cons does not get a single anchor result, but material for a weighing of options.
  3. Document the origin. In decision notes it should remain recognisable which parts come from an assistance system. That costs little and preserves traceability.
  4. Demand counter-argumentation. A second query that argues specifically against the first recommendation exposes weaknesses that did not appear in the first answer.
  5. Separate roles. Whoever produced the draft should if possible not sign it off alone — that applies to AI outputs just as to human ones.

These measures are deliberately kept simple. They work without additional tools and can be hooked into existing review and approval processes.

Assessment

The experiment shifts the view of AI quality: not only the correctness of the output counts, but also its effect on the judgement of the people who work with it. For web and software projects this means that governance for AI use should start less with the model than with the decision path — who sees what and when, and who cross-checks.

Anyone who anchors this in code review, architecture approval and design phases retains the speed advantages and does not lose technical control over decisions whose costs only become visible years later.

Sources