Ethical Stack.

Meta Muse

Personal cloud agent that works through connected apps, browser tasks and long-running goals, using Meta Muse Spark.

Score calculation

(2 + 1 + 5 + 2 + 7 + 3) ÷ 6 = 3.3

Pressure target: −1 overall point. Overall score: 2.3/10.

6 of six dimensions rated. Unrated dimensions are excluded from the average.

Full scoring method ↗
Setup details

US rollout of Meta’s hosted consumer agent. Muse Secure VM runs in Meta cloud; connected app access and permitted actions are user-configured.

What fits your needs

Hosted personal agent using a dedicated cloud VM, Meta Muse Spark, connected apps and a browser; rollout stated for the US.

Everyday actions

Handles browser-based research, form completion, email and travel tasks, and can continue working after the app closes.

See the evidence

Connected accounts

Users select apps and choose whether access is read-only or may send on their behalf.

See the evidence

Availability and price

Meta says Muse is rolling out in the US, is free for most usage, and offers subscriptions for higher usage.

See the evidence
Tradeoffs to expect

Hosted execution

Agent and connected-service data run in Meta’s cloud VM; no local inference option is documented for Muse.

See the evidence
How to switch

Choose permissions

Connect only the apps needed, set read/send permissions individually, and disconnect services from Muse controls when access should stop.

See the evidence
Practical questions

Sensitive actions

Meta says Sentinel checks outbound internet actions and Muse requests approval before sensitive steps such as sending email or making purchases; an audit trail is available.

See the evidence

Memory and model use

Muse remembers user details; users can request that it forget specific items and opt out of interactions being used to train Meta AI models.

See the evidence

Energy practices

2/10↑

Muse is described as a hosted dedicated-VM agent; the product materials provide no scoped energy report, facility electricity mix or workload control evidence.

See the evidence

Meta’s stated dedicated-cloud-VM architecture gives a basic operational boundary for the agent. The reviewed product and safety sources do not quantify energy, identify the serving facility’s electricity or document a product control linked to resource reduction; the score is limited to boundary disclosure.

1 + 1 reporting + 0 electricity + 0 controls = 2/10. This grades documented practices and transparency.

VM region, model-serving route, task energy, facility electricity and resource-management controls are not disclosed in the cited product sources.

Reporting1/3
Meta identifies Muse as running in a dedicated cloud VM, defining a hosted operational boundary, but publishes no product-level energy report.
Electricity sourcing0/3
The product announcement names Meta cloud hosting but does not identify the VM’s facility electricity mix.
Resource controls0/3
The reviewed Muse materials provide no product-level token/context or workload resource controls tied to reduced physical demand.

Water disclosure

Water reporting has not been assessed for this choice.

Training data

1/10↑

Muse names Muse Spark, but the reviewed model materials do not disclose its training-source inventory, training-data rights or creator remedies.

See the evidence

Meta’s product and model announcements identify Muse Spark as the fixed underlying model. The reviewed source does not provide a training-content summary or rights ledger, and user controls over agent interactions don’t confirm permissions for the model’s training data.

1 + 0 source transparency + 0 permission + 0 creator controls = 1/10, with the permission cap applied.

Training source categories, named datasets, source permissions and release-specific creator controls remain undisclosed in the reviewed model announcement.

Source transparency0/3
Meta identifies Muse Spark as the model but the reviewed model announcement does not inventory source datasets or training data categories.
Documented permission0/4
The reviewed Muse Spark source doesn’t confirm permissions or a rights basis for the training data.
Creator controls0/2
The reviewed model announcement does not document a training-data creator opt-out, removal or retraining process for Muse Spark.
Data-rights rubric ↗

Your data & privacy

5/10↑

Meta describes a dedicated cloud VM, per-app permissions, disconnect controls and an opt-out from Meta AI model training; interactions are not used by ad systems.

See the evidence

Meta’s launch materials say each Muse runs in a dedicated cloud VM, users choose app permissions, may disconnect services, and may opt out of interactions being used to train Meta AI models. This remains a hosted product; later Confidential VM encryption was described as future, not generally available at launch.

5/10 reflects documented per-app permissions, action approvals and interaction-training opt-out, balanced against hosted cloud execution and unresolved retention and rollout details.

Current opt-out scope, retention/deletion details, availability of Confidential VM, and third-party connector/model processing require checking current settings and terms.

Openness

2/10↑

Muse is a hosted closed product; the agent runtime and Muse Spark model are not released as open source in the cited launch material.

See the evidence

Meta describes Muse as a hosted agent powered by Muse Spark and its dedicated cloud VM, without publishing the product code or downloadable weights in the cited launch documentation. Access controls and a model-training opt-out do not make the system inspectable or portable.

2/10 reflects a hosted consumer agent without a published source license or user-modifiable client/model route in the reviewed product materials.

No public model card, runtime repository, or downloadable Muse Spark weights were identified in the cited materials.

Worker protections

7/10↑

Meta now reports supplier surveys, independent factory audits and corrective-action programs, but recent reporting on its AI annotation partner Sama raised privacy and contractor-condition concerns before Meta ended that engagement.

See the evidence

Meta’s FY2025 anti-slavery statement says its supplier program used independent RBA VAP audits, anonymous worker surveys and focus groups, and tailored corrective-action plans with closure review; its program is targeted to selected manufacturing suppliers, not all annotation vendors. In 2026, SvD/GP reported Sama workers handling intimate smart-glasses footage and Meta said it paused then ended the relationship because Sama did not meet its standards; AP reported more than 1,000 Sama layoffs and ongoing Kenyan litigation over earlier moderator allegations, which Meta and Sama dispute. Meta’s supplier system merits credit, while the AI-annotation accounts and subsequent contract end leave actual coverage, remediation and worker outcomes unresolved. These findings concern the developer or operator. This product’s use in a particular deployment needs separate evidence.

1 + 2 pay & conditions + 2 worker voice & remedy + 2 evidence & follow-through = 7/10. This grades documented worker safeguards, not a measured rate of mistreatment.

Supply-chain worker systems focus on manufacturing sites. The 2026 Sama evidence concerns third-party AI annotation and sensitive footage; allegations are reported, investigations/litigation are ongoing, and Sama denies retaliation/monitoring claims. Meta said the work had user consent, paused while investigating, and ended the contract, but did not disclose findings.

Pay & conditions2/3
Meta reports audited supplier expectations and corrective-action systems and has made historical contractor commitments; current annotation-worker evidence is mixed and the Sama relationship was terminated after an investigation.
Worker voice & remedy2/3
Meta reports anonymous worker surveys, focus groups and grievance strengthening at selected supplier sites; public reporting also documents workers speaking out, but it doesn’t confirm effective channels or remedies for annotation workers at Sama.
Evidence & follow-through2/3
Meta reports third-party RBA VAP audits, corrective-action plans and closure audits for selected manufacturing suppliers; it has not published independent outcome findings on the Sama AI annotation concerns.
Workers rubric ↗

Military ties

3/10Limited defense work↑

Meta publicly describes Llama access for defense and national-security partners, including a deployed special-operations assistant; evidence supports defense use but not targeting, strike decisions, or combat intelligence.

See the evidence

Meta’s September 2025 national-security announcement says Llama is powering the USSOCOM SOFChat assistant and describes mission-support examples such as faster intelligence report production and video analysis/monitoring of field conditions. Meta also makes Llama available to selected U.S. allies. This establishes real defense support and not mere abstract availability, but the reviewed public sources don’t show that Llama selects targets, supports strikes, or supplies combat intelligence. Meta’s broad distribution limits visibility into downstream uses and the reviewed materials disclose no binding military weapons/surveillance restrictions or independent use oversight. Classify the named described use as limited-defense support, not direct combat use. These findings concern the developer or operator. This product’s use in a particular deployment needs separate evidence.

1 + 0 use restrictions + 0 enforcement & oversight + 2 relationship disclosure = 3/10. Known defense ties lower this rating; safeguards can improve it within the relationship’s cap.

Meta does not identify all deployed model versions, exact field use, downstream applications, operational safeguards, or review outcomes. No source establishes that Llama is used for weapons targeting or a particular strike.

Use restrictions0/4
The reviewed Meta defense announcements describe partner access and examples but do not disclose binding weapons, targeting or harmful-surveillance restrictions.
Enforcement & oversight0/3
Meta publishes no military-specific use review, enforcement mechanism or independent outcome evidence in the reviewed defense announcements.
Relationship disclosure2/2
Meta names the USSOCOM SOFChat deployment and gives examples of report-generation and field-video analysis, while not disclosing downstream auditability or mission-level limits.

Named special-operations support deployment and allied availability are documented, but no evidence reviewed ties Llama to targeting, strike decisions, or combat intelligence; limited-defense support classification applies. The relationship limits this grade to 6/10.

Military rubric ↗

BDS targets

Pressure target↑

Meta is listed as a pressure target in the BNC No Tech campaign; this is company scope, not evidence about Muse deployment.

See the evidence

The BNC No Tech campaign explicitly lists Meta among its technology targets. The match is to Meta as company; it doesn’t confirm an operational relationship between Muse and any named military program.

This is limited to the two official BNC campaign sources; it does not cover every local campaign, supplier, investment, or future change.

More from Meta

All sources

Side by side