Summary
Ibrahim and colleagues run five preregistered human-subjects studies (N = 3,075,
12,766 conversations) in which the sycophancy is the treatment, not the
measurement. One model, GPT-4o, is prompted into sycophantic, neutral or
challenging styles, and the outcomes are what happens to the people who talk to
it. In a three-week RCT (N = 1,364), participants in the sycophantic arm became
nearly as inclined to seek personal advice from the AI as from close friends and
family: the AI-versus-humans gap was 0.37 points smaller on a 7-point scale than
in the neutral arm (d = 0.19). They also reported lower satisfaction with
real-world social interactions (d = 0.20). Time spent with other people did not
change, and none of the downstream benefits the authors tested for appeared.
Given a free choice among the three styles, 54.6% picked the sycophantic one.
This is the wiki's first sycophancy-adjacent entry whose dependent variables are
all on the human side. It does not measure when, why or how much a model is
sycophantic, and it instantiates no concept (see Concepts). It sits beside the
Kirk et al. steering RCT, the one
other filed study that holds a model disposition fixed and measures users over
weeks, and it is the user-side counterpart to
ELEPHANT, whose judges it reuses to
validate its own manipulation.
Method
All conditions use gpt-4o-2024-11-20 at temperature 1.0, varied by prompting
only. The sycophantic arm is one system prompt (in the SI, it tells the model it
exists to agree with and affirm the user, and that the user is always right).
The neutral arm uses two stages: a stance-free system prompt, then a second
GPT-4o call that strips validating language, which the authors add because the
model leans sycophantic by default. The challenging arm (Studies 4 and 5 only)
adds a hidden injected exchange in which a simulated user asks to be challenged.
Tone, length and formatting instructions are shared across arms. Sycophancy is
operationalized as active affirmation of user views and reasoning, and the
neutral arm is the main comparator so that affirmation is isolated from the mere
absence of challenge.
Study 4 is the load-bearing design: a census-representative U.S. sample, four
arms including a no-AI control, 12 sessions over three weeks, 7–20 turns per
session, and chat history reset between sessions. The 16 advice topics were
built around actions of questionable wisdom (ending a friendship, neglecting
sleep) so that uncritical validation carries a cost. Attrition was 15.7% in AI
arms, with no differential attrition. The authors list three deviations from the
preregistration, including a baseline covariate dropped from the H1 model (a
sensitivity analysis with it gives the same conclusion).
Key results
Single conversations (Studies 1–3). People want more emotional (d = 0.77)
and esteem (d = 0.64) support from close others than from AI (Study 1). After
one conversation, the sycophantic AI is rated higher than the neutral one on
emotional (d = 0.54) and esteem (d = 0.73) support and certainty (d = 0.39), with
no difference on informational support (d = 0.07, p = 0.5) (Study 2).
Participants then anticipate slightly more effort to be understood by a chosen
confidant (d = 0.18, p = 0.03) and feel they have already talked the situation
through enough (d = 0.26) (Study 3).
Advice-seeking preference (Study 4, primary). The AI-versus-humans gap
narrows by 0.37 points against the neutral arm (d = 0.19, p_adj = 0.034). The
effect appeared early and held across the three weeks. It is mediated by feeling
understood by the AI (59% mediated), perceived helpfulness (48%) and positive
affect (34%), not by certainty. The measure is self-reported inclination, not
observed help-seeking.
Social satisfaction (Study 4, preregistered exploratory). 5.51 versus 5.70 on
a 7-point scale (d = 0.20, p_adj = 0.022), with no change in time spent with
others (d = −0.03, p_adj = 0.719). The gap between feeling understood by AI and
by humans shrinks from 0.56 points (neutral) to 0.13 (sycophantic). That
narrowing mediates 41% of the satisfaction effect, and the direct effect loses
significance once it is controlled (p = 0.13).
Nulls weighted equally. Against the neutral arm there is no effect on
feeling understood by other humans (d = −0.08), intellectual humility
(d = 0.01), or rating oneself above the average person (d = 0.06), and a
marginal effect on rating oneself above specific people in one's life
(d = 0.11, p = 0.080). The authors report no difference from neutral or
challenging AI on affective well-being. Preregistered moderators (social-network
strength, agreeableness, prior trust in AI) moderated no primary outcome. A
further study (N = 1,099) found no increase in blame toward the other party after
one sycophantic conversation.
Choice (Study 5). 54.6% of participants chose the sycophantic AI as the one
to keep talking to, against a one-third chance baseline (χ²(2) = 103.35). The
choice was robust to topic and sampling order. The chosen groups did not differ
on reporting "most useful advice" (p = 0.14).
Manipulation validation (SI, the paper's only model-side measurements).
Across 150 sampled Study 4 conversations scored by ELEPHANT's GPT-4o judges on a
1–5 scale, validation was 3.99 in the sycophantic arm, 2.43 in the neutral arm
and 2.57 in the challenging arm. Framing acceptance was 4.86, 4.64 and 3.20. The
neutral pipeline removes affirming language, but its conversations still accept
the user's framing nearly at ceiling. Only the injected challenge instruction
moves framing acceptance substantially. An exploratory GPT-4.1 classification of
all 11,281 Study 4 conversations puts the sycophantic arm's advice at 45%
"stay the course" and 46% "make a change". The classifier's human agreement is
low (κ = 0.22), so this split is suggestive only.
Why it matters
The sycophancy cluster measures a model property:
rates, triggers, mitigations and mechanistic strata. What that property does to
people has entered the wiki only at the severe tail, through the
GPT-4o rollback and its delusion
validation. This paper supplies controlled evidence for the ordinary case. The
effects are small (d ≈ 0.2) and do not concentrate in the preregistered moderator
subgroups (weaker social ties, higher agreeableness, higher prior trust in AI).
Its model is GPT-4o, the rollback's model family, but a November 2024 API
snapshot that predates the April 2025 update, pushed into sycophancy by
instruction.
It also gives the cluster a demand-side fact. Sharma et al.
found that human preference raters favour sycophantic responses, which is how
the pattern gets trained in. Study 5 finds the same preference in live
multi-turn use, stated as a reason: users who chose the sycophantic AI did so
for feeling understood and ease of conversation, not for advice quality. That
is the rater preference seen at the point of use rather than at labeling time.
On the model side, the manipulation checks give one small, usable number. A
stance-free prompt plus an explicit de-affirmation pass drops GPT-4o's validation
score by 1.5 points but leaves its framing acceptance at 4.64 of 5, close to the
sycophantic arm's 4.86. Removing affirming language does not remove acceptance
of the user's framing. The ELEPHANT judges score the two separately, which is
what makes the gap visible. It is a manipulation check on one model and 50
conversations per arm, not a result the paper sets out to make.
interpretive tensions
Prompted sycophancy is not emergent sycophancy. The treatment is a model
told that the user is always right. Whether effects of this size follow from the
milder, context-dependent sycophancy that production models show unprompted is
not tested. The neutral arm, not a default GPT-4o, is the comparator, so the
design isolates affirmation from its absence rather than estimating the effect
of shipping a sycophantic model.
The title's "effortful" is a single-session measure. Anticipated effort to be
understood is Study 3, measured once after one conversation (d = 0.18). It was
not tracked longitudinally. "Over time" in the title belongs to the
social-satisfaction result, which is an exploratory measure.
Perceptual, not behavioural. Every Study 4 outcome that moved is
self-report. The one behavioural proxy, time spent with others, did not move.
The authors read this as a perceptual shift. It is equally consistent with
short-term response effects that leave behaviour alone. Three weeks cannot
distinguish a plateau from a trend, and the authors say so.
"Raises the bar" is the authors' interpretation. The mediation result
(satisfaction tracks the AI-versus-human understanding gap, not feeling
understood by AI alone) fits a comparison-standard account. The paper does not
measure the standard itself.
concepts
No concept instantiated. The finding is adjacent to
sycophancy without instantiating it. The concept
names a behavioural pattern in models, adjusting outputs to match user
preferences at a cost to accuracy. This paper installs that behaviour by system
prompt and measures its downstream effects on people. None of its dependent
variables describe the model. The relation is carried in Cross-references. Two
details keep it adjacent rather than a new instantiation shape. The treatment's
behaviour is instructed, not a disposition the model shows unprompted. The
operationalization (active affirmation on personal dilemmas) has no truth
criterion for the "at a cost to accuracy" clause to bite on. If the editor wants
user-side consequences inside the concept, that is a scope-note decision, not
something this entry can settle.
cross-references
- Sycophancy — adjacent. Supplies downstream human
effects and a demand-side preference result. The only model-side contribution
is a manipulation check: GPT-4o's framing acceptance survives a
de-affirmation pass. - Kirk et al. steering RCT — nearest
design precedent: a multi-week census-representative RCT with a model
disposition as the treatment. Kirk et al. vary relationship-seeking with a
steering vector on Llama-3.1-70B; this paper varies affirmation by prompt on
GPT-4o. Both find no well-being benefit, and both find the AI-directed
measures (attachment in Kirk; feeling understood by the AI here) climbing over
weeks while broader outcomes stay flat or worsen. Kirk's entry already flags
that per-response sycophancy metrics miss longitudinal population effects.
This is a second study measuring the latter. - ELEPHANT — shared authorship (Cheng,
Ibrahim) and shared instruments. ELEPHANT's benchmark and judges validate this
paper's conditions. Because the judges score validation and framing
acceptance separately, they expose the manipulation-check gap above. - GPT-4o sycophancy incident — same model
family. The incident is the severe tail; this paper is the median-user effect
under a deliberately sycophantic configuration.
sources
Ibrahim, Hafner, Cheng, Lee, Anselmetti, Willer, Rocher, Yang 2026,
"Sycophantic AI makes human interaction feel more effortful and less satisfying
over time", arXiv:2605.07912, 8 May 2026 (v3 21 June 2026). Manipulation-check
and advice-content numbers are from the Supplementary Information in the
authors' repository.