When AI agrees too easily: making a second opinion useful

Agreement is a distinct risk

Cheng et al., published in Science on 26 March 2026, studied 11 models, which affirmed user actions 49% more often than humans on average. Three preregistered experiments involving 2,405 participants found reduced willingness to repair interpersonal conflicts after sycophantic interactions. These were not evaluations of workplace decisions or Colay. Cheng et al., Science, 2026

Anthropic reported sycophancy in five assistants across four free-form generation tasks on 23 October 2023. The research links this behavior partly to preferences for responses that match a user’s views. Anthropic, 2023

Our workflow recommendation is to keep likability separate from reliability. A harsh response is not automatically better than a flattering one. A useful objection identifies a checkable premise, offers a plausible alternative and explains what evidence would change the critic’s view. Asking a model to be ruthless can simply replace eager agreement with confident fault-finding.

sycophantic models were trusted and preferred

Cheng et al., Science, 2026

Science reports users’ preference for sycophantic models.

A question can select its winner in advance

Compare two prompts. “Help explain why our contractor ruined the launch” already assigns blame. “Here are the plan, messages and dates: what could explain the delay?” leaves several possibilities open. For analysis, remove evaluative language from the initial description and separate observations, interpretations and missing information.

You do not need to conceal your position permanently. Put it after the facts and ask which evidence supports it and which evidence conflicts with it. This gives another perspective a precise job: inspect the argument rather than infer how strongly you want agreement. The same technique works for a product proposal, a difficult email or a retrospective.

Worked example: three delays do not establish a cause

Imagine a manager considering replacing a contractor after three missed deadlines. This is an invented teaching example. The initial prompt calls the contractor obviously incompetent. One model can quickly draft a persuasive case for replacement. Yet the supplied timeline also contains two late requirement changes and one access delay caused by the customer.

A useful review separates delivery performance from the conditions under which delivery was expected. One pass checks the contractor’s commitments, another checks changes to the inputs, and a third compares next steps. The final record should identify each promise, change, responsible party, schedule impact and missing document. If the evidence confirms a contractor failure, a critical review does not have to excuse it.

The counts in this example are not probabilities of blame. They illustrate why a shared event list can support different explanations until the causal sequence is established. The manager’s next task is to check the dated agreements and speak with the people involved. The eventual decision may remain unchanged, but its reasoning becomes inspectable.

Requesting a useful Consensus discussion

Colay Consensus brings participant discussion into a result synthesized by a coordinator. A user can ask: “Test my interpretation. Separate observations from judgments. State the strongest alternative explanation. Identify evidence that would change each conclusion. Keep material disagreements in the final answer.” This is a suggested workflow, not a validated cure for sycophancy.

A different role or model name does not establish independent judgment. Participants may receive the same loaded question, make similar assumptions or accept the first confident argument. Evaluate whether the discussion produces a new fact to check. The number of agreeing responses alone should not determine how much trust to place in the outcome.

Distinguish criticism from noise

A substantial objection can be brief: the email contains no agreed delivery date, so it cannot establish lateness. That objection connects the conclusion to a missing piece of evidence. A weak review may instead list generic risks that fit almost any situation. Such a list takes time to read without helping resolve this particular decision.

Ask for objections to be ordered by their potential to change the choice. Pair each with an action: inspect a document, clarify a condition, reproduce a calculation or run a small reversible test. An objection that neither affects the choice nor suggests a check can be removed from the final memo. This reduces noise while preserving the questions that are uncomfortable for a reason.

Agreement can survive an honest test

A second opinion does not need to disagree. Sometimes the initial position is well supported and the alternatives are weaker. Retain the decision and record which arguments survived inspection. Discussion becomes useful when it allows several outcomes: confirmation, revision, rejection or a pause while additional evidence is obtained.

Start with a decision you favor but can still change. Write down your own reason for choosing it, then give Consensus the facts and ask for conditions that would overturn the conclusion. Read the result alongside the source documents. When other people are involved, hear their explanations directly: a model is analyzing your description rather than observing the situation.