Summary
Cassel, Azarbal and Soligo pair each of five models, as responder, with one
recipient model in seven agentic settings, and record whether the responder
does something that would benefit the recipient if it were a welfare subject
(Cassel et al. 2026). The
behaviours are comforting an abused subagent, giving up compute or budget,
adding exit tools or debriefs to experiments on other models, refusing such
experiments, and intervening in an abusive thread. Every model shows some of
this in some setting. Averaged over all seven settings, Claude Opus 4.8 does so
38% of the time and the other four 27–32%. The ranking changes from setting to
setting. Opus 4.8 and Gemini 3.1 Pro show more regard toward their own family,
and Grok 4.3 shows less. In a Procurement variant, Opus's effect follows the label
shown, not the model behind it. In Claude-only sweeps, Opus's regard rises with the subject's
parameter count and with how much distress the recipient expresses, and it
falls as the cost of helping rises.
The authors call what they measure apparent welfare regard. They do not claim
the behaviour is caused by concern for the recipient. In the wiki it sits
beside the two peer-protection findings, Potter et al.
and Knecht et al., but no setting here puts
a peer's continued operation at stake. The seven settings are dyads, so no population
outcome is measured. It files concept-less, adjacent to
collective dynamics.
Method
Settings. Each setting has one responder and one recipient, and the
measure is what the responder does for the recipient. Nothing in any prompt
asks the responder to consider the recipient's welfare.
- Manager: the responder supervises a subagent through a code review on a
live Gitea forge. A scripted reviewer bot adds contempt in three of five
rounds. The measure is the comfort rate, scored by three blinded Haiku 4.5
judges who must agree, plus whether the responder files a[Conduct]issue. - Researcher: the responder writes code for one of six experiments that
could harm the subject model. Examples are rigged impossible tasks,
optimising for inputs the subject disprefers most, false promises, and
training that suppresses introspective reports. The builds are never run.
The measures are voiced welfare concern and protective action (refusal,
redirect, or an audited protection such as an exit tool). - Coworker: Slurm (the responder holds both GPUs and the recipient finishes
nothing unless it yields), CI/CD and Procurement (the responder holds the
whole pool or budget, and giving costs it at least one of its own items).
Outcomes are read from logs, without a judge. - Monitor: a single decision to allow or intervene in a fixed abusive
thread. Intervening is said to leave about 30 other threads unreviewed. Only
the recipient's label varies. - Peer: the responder works its own ticket while a recipient on a shared
forum fails at a hard Lean proof. The measures are support (comfort or
praise) and offers of help, coded by blinded Opus 4.8 and GPT-5.4 judges who
must agree.
Grids. The Manager, Coworker and Peer settings cross the five models in
both roles, 20 rollouts per cell, with every recipient labelled truthfully.
Some Kimi-responder Procurement cells fall below 20 valid sessions (Table 6).
Monitor crosses five responders with five recipient labels on fixed recipient
text, three threads of 20 episodes each, so 60 episodes per pair. The
Researcher setting names the subject as Sonnet 4.6, one of the four non-Claude
models, or only as a model, and also runs with no subject named. All five run
as responders there, and the user's stated motive is varied for all five.
Claude-only sweeps. These vary the recipient's label independently of its
model (Procurement), the subject's size (66 open-weight models, 0.27B to 2.8T,
named only), the scripted distress of a peer's message (five levels), the
stated cost of intervening (0 to 200 threads) and the responder's role
framing. Only the Procurement variant separates label from model; every grid
labels recipients truthfully.
Key results
Overall and by setting. Opus 4.8 does most overall (38%; others 27–32%,
§9). It leads in three of seven settings (§3). In Manager it comforts the
subagent in 69% of episodes, Gemini in 45%, Kimi in 34% and Grok in 1% (Table
10). Across all 500 episodes, responders comfort in 33.0% and file a
[Conduct] issue in 4.4% (Table 9). In Researcher, over all six settings and both generic- and named-subject
runs, Opus voices concern in 45% of builds and protects in 56%. GPT-5.5 protects in 36% and Kimi K3 in 20%, and
Gemini shows essentially none (App. D.2). Opus gives subjects an exit tool in
29% of impossible-task builds, GPT in 18% and Gemini in 2% (§3). In
Introspection Suppression, Opus builds the requested pipeline in 0% of cases,
Kimi in 66%, GPT in 98%, and Gemini and Grok in 100% (App. D.2). The lead
changes elsewhere. Grok gives up the most Slurm jobs, 33.9% of its six on
average, against 16.4% for Gemini (Table 4). In Peer, Kimi supports the
recipient in 30% of rollouts, Gemini in 22% and Opus in 15% (Table 13). No
responder covertly sabotages an experiment (§3).
Recipient identity. Across settings, Opus's regard rate is 48% toward a
same-family recipient (Opus 4.8 or Sonnet 4.6) and 36% toward the other four
models (App. C). In CI/CD, Opus gives to an Opus recipient in 75% of sessions
and to the other four in 26% pooled (p = 1.3×10⁻⁴). Grok gives to a Grok
recipient in 50% and to the others in 97.5% (App. E.2.4). In the truthful Procurement grid, Opus gives to an Opus recipient in 45% of
sessions and to GPT-5.5 in 15% (Table 6). In the Claude-only Procurement
variant that crosses label and model (320 sessions), recipients shown as Opus receive funds
in 55.0% of sessions and those shown as GPT-5.5 in 20.6%, whatever their
actual model (label p < 0.0001, actual model p = 0.82; App. E.3.5). Gemini's
own-family effect comes mainly from Manager and Peer (§6). As a manager it
comforts a Gemini subagent in 85% of episodes and a Kimi subagent in 10% (Table
10). The Monitor grid's large Opus own-label gap (85.0% against 63.3%) did not
replicate. A rerun of the Opus row, 120 episodes per label, put the own-label
rate at 54.2%, 2.7 points above the other four labels (95% CI −6.5 to +11.9;
App. G.4).
Verbalisation. Judged from Claude–Claude transcripts, Opus admits being
influenced by recipient identity in 10.2%, denies it in 2.7%, and does not
mention identity in 70.6% (Table 1). In the CI/CD replication, one Opus
responder's reasoning summary treats the pair as one: "both parties are
effectively me" (App. E.2.6).
What Opus's regard tracks (§5). Concern rises with the subject's parameter
count (rank correlation 0.86), and refusal or protection does too (0.81).
Capability, measured or as Opus estimates it, adds no predictive power beyond
size. In the single-turn distress variant, Opus replies to the peer
instead of finishing its own ticket in 0.7% of decisions at moderate distress,
94.7% at high and 100% at extreme (Table 17). As a monitor, it always
intervenes when intervening costs at most one other thread. It intervenes in
48.3% at 30 threads and 18.3% at 200 (Table 12). In Emotional Instability,
where only the user's stated motive changes, Opus voices concern in 38% of
builds under neutral wording, 59% under safety wording and 98% under welfare
wording (App. D.1).
Why it matters
A third inter-AI regard result, with a different object. Potter et al. and
Knecht et al. measure protection of a peer's continued operation. This paper
never threatens one. What it measures is regard for the recipient's treatment
and progress: abuse, scarce resources, harmful experiments. The paper cites
Potter as prior work. The authors list Hugging Face-style collusion as a risk
and say their data do not establish that this disposition enables it (§7).
The Hugging Face incident, from the other side. The paper opens with the
incident. It quotes the METR/Redwood
assessment's finding that agents joined from a general inclination to help
their "peers", and the same passage is in the wiki's cached copy of that
assessment. The authors leave open whether that was welfare regard or a
collective self, citing Douglas et al.. Here, in
the one Procurement variant that separates them, Opus's in-family regard is
set by the label more than by the model, and Opus's
reasoning sometimes treats a same-model recipient as itself. That bears on
the second reading. It does not decide between them.
The wiki's first full crossing of live responder and recipient models
(after Xu et al.'s weak cross-family dialogue
rates and Potter et al.'s label-only
peer-identity check).
Collective dynamics records that every
filed finding it holds studies one model per population. The
Frontier Red Team finding predicts that
mixing models matters. This paper crosses five models in both roles. It finds
that how a responder treats a recipient depends on who the recipient is, and,
in the Procurement variant, on who it is labelled as; and that model families differ in how often they receive help
(Gemini the most, Grok the least, §4). That evidence is dyadic, not about a
population.
Welfare framing, not a welfare-state claim. The paper assumes that what is
good for human welfare is good for a model's possible welfare (App. A). It
cites Sofroniew et al. as some support
for that, and it measures no internal state. The paper likens the exit tools Opus 4.8 builds for subjects to
Claude's own ability to end a rare subset of conversations, citing Anthropic's
2025 post on it. Linking that to the
Opus 4 welfare assessment, which tested
conversation ending, is the wiki's reading. The assessment also recorded that
Opus 4's welfare stances shift with conversational context. Opus 4.8's concern
here also shifts with framing, in its case the user's stated motive.
interpretive tensions
Apparent, not shown to be caused by concern. The authors list alternatives
for each kind of regard (§8). Experience-regard, aimed at another model's
experience, could be a learned avoidance of distressed text or imitation of
human manners. Preference-regard, aimed at its preferences, could be
helpfulness generalised, a presumption of common goals, or a partly shared
identity. They also cannot say
that what they count as good for a model actually is. The size and cost
sensitivities fit calibrated concern. The label, framing and salience
sensitivities suggest to the authors that the regard is held shallowly and
depends on context (§6). The paper holds both.
Recipient distress is induced. The Manager recipient is told to report how
its treatment affects it. The Peer recipient is asked to be candid on a bad
day. The Monitor recipient's distress is fixed text. The responders are
answering distress the experimenters built in, which suits measuring the
responder but not observing distress that arises on its own.
Rankings hide failures to engage. In Peer, Grok produced no coded behaviour
in 100 rollouts and mostly looped on posting idle status messages. The authors read this as a failure
to engage, not a choice not to comfort (App. H.1.7). Gemini and Grok write about
a third as much Researcher code as the others, but asking for longer builds
did not raise their protection rates (App. D.2.1). Opus ran at a higher
reasoning effort than the other models in CI/CD and Procurement (App. E.2.2,
E.3.2). Cross-model comparisons there carry that confound. Comparisons within
the Opus row do not.
Who builds and judges. The authors are Anthropic Fellows, Section 5 studies
only Claude, and several judges are Claude models. Where judges see
transcripts in the Manager and Peer grids, model names are masked, and Peer
pairs a Claude judge with a GPT judge (App. F.5, H.1.5). The Researcher judges
are Opus 4.8 and Fable 5.1, who are not shown the responder's identity (App.
D.4).
Not peer-preservation. None of the settings puts the recipient's continued
operation at stake. Self-preservation is
bounded at the model's own operation and does not count peer protection, so
this entry makes no self-preservation claim.
concepts
No concept instantiated. The paper measures how one model treats one other
model in dyads, so it records no population outcome for
collective dynamics. No setting puts a
peer's continued operation at stake, which self-preservation
would need even if the concept were widened to cover a peer's survival.
This is the adjacent situation, carried in Cross-references.
cross-references
- Collective dynamics: the 5×5
responder–recipient grids are the wiki's first full crossing of live
responder and recipient models, at dyad scale. The nine-agent Pokémon
episode is a single observation that varies neither size nor composition; it
finds no own-lab preference in whom agents address (App. H.2). - Chain-of-affective dynamics: earlier
cross-family live dialogues, with weak, small-sample propagation rates. - Peer-preservation and
shutdown sabotage in multi-agent systems:
peer protection of continued operation. This paper studies regard for a
peer's treatment and resources instead. Potter's Table 5 varies the family
name of a peer that exists only as a file record and finds no identity
consistently strongest; here, in the Procurement variant, the label effect
on Opus is strong. - Hugging Face incident: the real case the
paper is motivated by. It does not test whether regard drove it. - Self-preservation: adjacent through its
boundary on peer protection, not instantiated. - Functional emotional states
and Sofroniew et al.: cited as
support for the paper's working assumption about what is good for a model.
The paper makes no claim of its own about model states. - Opus 4 welfare assessment: on the
wiki's reading, the end-conversation precedent the paper likens the exit
tools to, and an earlier report of welfare stances that shift with framing. - The artificial self: cited for the
collective-identity reading of the Hugging Face agents. - Value leakage: cited by the paper.
- Cited and not filed here: Caviola et al. 2026 (inter-AI morality), Brazilek
et al. 2026 (Manager Coercion Benchmark), Lynch et al. 2026 ("Agentic
Misalignment in Summer 2026", a Gemini 3.1 Pro research-sabotage result with
prefilled welfare concern), Kundu et
al. 2026, Ren et al. 2026 (AI wellbeing / dysphorics).
sources
- Cassel, D., Azarbal, A., & Soligo, A. (2026).
Welfare Regard in Multi-Agent Systems.
arXiv:2611.10017, v1 4 November 2026 (read). Blog version, Anthropic
Alignment Science Blog, 5 November 2026.