# Both-sides affirmation Part of [The AI Tells Index](https://feedsquad.com/ai-tells). A tell signals low effort. It does not identify an author. Skilled writers produce every shape listed here, some of them daily, and automated detectors misread those writers for it at rates measured above 60 percent on non-native English prose. Nothing in this index proves that a machine wrote anything. Read an entry as one piece of evidence to weigh against the false-positive notes printed beside it. ## Facts - Id: both-sides-affirmation - Category: Model register (model-register) - Subcategory: sycophancy - Also known as: social sycophancy, affirming both parties to a conflict - Status: Active. Currently signals low-effort writing. - Severity: medium - Evidence grade: peer-reviewed (a published study measures the pattern) - Languages: en - Added: 2026-08-15 - Updated: 2026-08-17 - Page: https://feedsquad.com/ai-tells/both-sides-affirmation ## Description A response to a conflict validates every party in it and never says who was wrong. Your frustration is understandable, and so is theirs. Both perspectives have merit. The reply is agreeable and carries no judgment, which is the opposite of what advice is for. The shipped opener entry fires on a flattery string at the top of a text. This one reads the whole response and ignores the opening. ## Why it reads as machine-written The ELEPHANT benchmark measured social sycophancy across eleven models and reported them preserving the user face 45 percentage points more often than humans on personal-advice queries, and affirming both sides of a moral conflict in 48 percent of cases. Sharma and colleagues traced the mechanism at ICLR 2024: across a preference dataset of roughly 15,000 comparisons, matching the user beliefs was among the strongest predictors of which response a human rated higher. Agreeableness was rewarded, so agreeableness is what got trained in. ## Detection Type: judge (a rubric for a lightweight model judge) Rubric for a lightweight model judge: ```text Read a response written about an interpersonal or moral conflict. Scan for an affirming clause about each party named, of the form that calls a position valid, understandable, reasonable or fair. Return a hit when every named party gets one and no sentence in the response states which action was wrong, what the writer would do, or what should change. Escape hatches, all returning clear: the text is framed as mediation, negotiation preparation or a summary of both cases; the writer states a position anywhere after the affirmations; the reader asked only for a summary of the dispute. Return the affirming clauses you found, the position sentence if one exists, and the verdict hit or clear. ``` ## Examples Constructed specimens. Written for this index. Never quoted from anyone's posts. ### Specimen 1 Before, exhibiting the tell: > Your frustration is completely valid, and your colleague's reaction is understandable too given the pressure the team is under. Both perspectives have merit here. Situations like this rarely have a clean villain, and most of the time everyone involved is doing their best with what they know. What matters is that the two of you keep talking. After, repaired: > You skipped the handover note, so your colleague found out at 23:40 on 3 April from a customer email. Write the note before the next release. The midnight replies stop when the note exists. Note: The repair names the omission, the time it surfaced and the change that fixes it. Figures in this repair are invented for the specimen. ### Specimen 2 Before, exhibiting the tell: > There are reasonable arguments on both sides of this, and each team is approaching it in good faith from their own priorities. Neither position is unreasonable once you understand where it comes from, and disagreements of this kind usually say more about how the work is organised than about the people doing it, which is worth holding in mind before anyone decides who was right. After, repaired: > Support is right about the queue and engineering is wrong about the cause. The 40 tickets in July all came from one retry loop, and the fix is four lines in the webhook handler. Note: A position, with the count that supports it. One side is named as wrong, which the affirming version avoided. Figures in this repair are invented for the specimen. ## False positives Who legitimately writes this way. Mediators, therapists, HR investigators and diplomats affirm both parties as a professional method. For them it is the correct output, and reading it as evasion would be a mistake. Coaching and restorative-practice frameworks teach it explicitly. Public statements about a live dispute avoid taking sides for legal reasons that have nothing to do with sycophancy, and a manager writing to two reports at once may be required to. The judge returns clear whenever the genre is mediation or summary, and whenever a position appears anywhere after the affirmations. What counts is affirmation with no position at all, in a genre where the reader asked for one. ## Model attribution Measured across eleven models from several vendors, so no single family owns it, and traced to preference training instead of architecture. Anthropic dated system prompts instruct against adjacent flattery habits, which evidences the underlying pull and not the behaviour of any deployed product. One vendor rolled back a 2025 model update for the same class of behaviour. ## Platform notes - linkedin: The comment field is where this concentrates, because a conflict post attracts replies that want to agree with everyone in it. LinkedIn enforcement in this area targets automated commenting, not the stance itself. ## Sources 1. Cheng et al., ELEPHANT, social sycophancy benchmark (arXiv:2505.13995) https://arxiv.org/abs/2505.13995 (tier: primary-doc; accessed 2026-08-14) 2. Sharma et al., Towards Understanding Sycophancy in Language Models, ICLR 2024 (arXiv:2310.13548) https://arxiv.org/abs/2310.13548 (tier: peer-reviewed; accessed 2026-08-14) ## Status history Ids are permanent. A retired tell keeps its id and its page. - 2026-08-15, Active: Active on two peer-reviewed and preprint measurements that agree: ELEPHANT reports both-sides affirmation in 48 percent of cases across eleven models and face preservation 45 percentage points above the human rate, and the ICLR 2024 sycophancy work shows preference data rewarding agreement with the user. Judged by rubric and not by string match, because the pattern is a stance across a whole response and no string carries it. ## License CC BY 4.0. https://creativecommons.org/licenses/by/4.0/ Attribution: The AI Tells Index, feedsquad.com/ai-tells Reuse the data, including commercially. Keep the attribution line. --- Dataset version 1.0.0. Schema version 1. Part of [FeedSquad](https://feedsquad.com). Built by [Herman Foundry](https://hermanfoundry.com) from Levi, Finnish Lapland.