TrueSeeker AI · Verified claim report Case 994148b769 · 2026-09-24

§ Claim under review · Business

"Meta reportedly tested human contractors to secretly handle some phone calls made through its new Muse AI assistant, according to internal company posts seen by Reuters." (Post overlay text: "Meta reportedly used humans to secretly handle calls made by its new Muse AI agent")

Circulating claim, as submitted.

Verdict

Mostly accurate

Confidence

High
§

Summary

This post is mostly accurate. Reuters did publish an exclusive on September 22, 2026, reporting that Meta tested having human contractors place and handle some of the phone calls requested through its new Muse AI agent, based on internal company posts. The details in the caption check out: the feature was internally called a "human concierge," it was switched on for about half of Meta employees with an opt-out, staff raised privacy concerns about contractors hearing sensitive information, and a Meta vice president wrote internally that starting the test without proper disclosures "was a miss." Two things are slightly overstated. The word "secretly" is stronger than what was reported, since Meta did announce the test to employees; the disclosure failure was about not telling people at the time of the calls. And the rollback was described as "for now," with Meta saying publicly it still plans to ship the calling feature once it is ready and properly disclosed. The post also leaves out Meta's own on-record response, which called this routine pre-release testing and said employee feedback was positive. The internal documents themselves have not been seen by anyone outside Reuters, so the specific figures, including a claimed 95 to 98 percent success rate for human-handled calls, cannot be independently checked.

§

The readings

key figures from the evidence
half of employees

share of Meta staff enrolled in human concierge test

2.5 million downloads

Muse app downloads reported by Sensor Tower

§

Why this verdict

The full Reuters wire text was retrieved verbatim through licensed syndication and the Instagram caption's five factual bullets each map onto a specific sentence in it, with the subject declining to deny the test on the record. As of 2026-09-24, the only gaps are the word "secretly" where Reuters wrote "quietly," the dropped "for now" on the rollback, and a VP's internal post presented as a company acknowledgment, none of which change what a reader takes away. "Accurate" was rejected because "secretly" and the omitted "for now" do slightly harden the story beyond the source. "Partially accurate but misleading" was rejected because no cited source contradicts the operative proposition, which Meta itself effectively conceded. "Credibly reported but unconfirmed" was rejected because the subject responded on the record and did not dispute that the test happened. Confidence is High on the claim-to-source comparison, since the source text is in hand; it does not extend to the internal posts, which no independent party has seen.
§

Evidence

The Reuters exclusive is real, dated September 22 2026, bylined Katie Paul, and the Instagram caption tracks it closely. Reuters reported that Meta has been testing a "human concierge" for its new personal AI assistant, Muse, which entails having human contractors quietly handle some of the phone calls placed via the digital agent, according to internal company posts seen by Reuters , and that the company told employees about the test last week, shortly after publicly launching a phone calling feature for Muse .

On the scale of the test: Meta enabled the human concierge feature, also referred to as "human agent calls," for half of its employees last week, according to the internal posts , and employees who did not want to be included could join an opt-out group .

On privacy: some employees raised privacy concerns, warning the approach could result in sensitive information being shared unintentionally with contractors in call centers . One employee wrote that "It's baffling to me why we think this feature is worth the risk" .

On the rollback: a vice president in Meta's SuperIntelligence Labs unit acknowledged in one of those posts that "it was a miss" to start testing the contractor-placed calls without proper disclosures and said the company had "rolled back this feature" for now .

On performance: the VP noted some tests indicated having humans make the calls could get their success rate up to the 95% to 98% range, instead of the "lower percentage of AI calling" .

Meta's on-record response did not deny the test. A spokesperson said the response from employees has been "overwhelmingly positive" and that the purpose of the test was to "get feedback so we can implement safety and privacy protections and improve features before we release them publicly."

"We're working with merchants to continue improving this potential calling feature, and will only roll it out when it's ready and with the proper disclosures," said the spokesperson .

Two details the Instagram post omits: Reuters reported that an employee who asked Muse to call a cable provider to negotiate his bill said a transcript showed the human contractor had made a racist reference during the call , and separately that the agent has topped US app download charts with more than 2.5 million downloads per Sensor Tower .

§

Findings

✓ What's accurate 7

  • Reuters published this as an exclusive on September 22 2026, sourced to internal company posts. The attribution in the post is correct.
  • The feature name "human concierge" is Reuters' reported internal terminology, not the post's invention.
  • "Enabled for about half of Meta employees during testing" matches Reuters.
  • Employee privacy concerns about contractors receiving sensitive information are reported accurately.
  • A rollback and an admission that testing began without proper disclosures were both reported, in an internal post by a Meta Superintelligence Labs vice president.
  • "Meta said human-assisted calls performed better than AI-only calls in some tests" is a fair, correctly hedged restatement of the 95 to 98 percent figure.
  • Muse is a real product, launched September 8 2026, with a real phone-calling feature.

≈ What's misleading 4

  • Exaggeration: the post says contractors handled calls "secretly." Reuters says "quietly." Reuters also reports that Meta announced the test to employees and offered an opt-out group, so the arrangement was not concealed from the tested population as a group. The disclosure failure Reuters documents is narrower: the test began without proper disclosures, and call recipients and individual users were not told at the time. "Secretly" upgrades a disclosure gap into deliberate concealment.
  • Omitted qualifier: the VP said the feature was rolled back "for now," and the spokesperson said Meta is still working on the feature and will roll it out with proper disclosures. The post's "later rolled back the feature" reads as termination.
  • Temporal overreach: the image overlay says Meta "used humans to secretly handle calls," present and completed. The caption's own wording ("tested") is more accurate than the headline it sits under. The overlay drops the test framing entirely.
  • Attribution imprecision (no canonical name fits): "The company later rolled back the feature, acknowledging that testing began without proper disclosures" presents an individual VP's internal post as a corporate acknowledgment. Meta's actual on-record statement conceded no error and framed the episode as routine dogfooding. The post omits that statement entirely, which removes the subject's side of the record.

? What's uncertain 5

  • The internal posts themselves were not retrieved. Everything about the test's scope, the rollback, and the success-rate figures rests on Reuters' reading of documents no one else has seen.
  • The 95 to 98 percent versus "lower percentage" comparison is an unpublished internal figure from an interested party. No methodology, sample size, task mix, or baseline number was published.
  • There is a spokesperson-name discrepancy across outlets: Reuters attributes the statement to Daniel Roberts, Gizmodo to Dave Arnold, with near-identical text. This does not change the substance but is unresolved.
  • Whether the human-concierge test ever touched non-employee public users is not established by the reporting. The evidence describes an internal employee test.
  • Whether the rollback is permanent is explicitly open; Meta says the feature is still under development.
Distortion flags exaggeration omitted qualifier temporal overreach
§

Sources

5 of 5 linked to records
[1]

Reuters wire article, "Exclusive-Meta testing a 'human concierge' for its new personal AI agent, Muse," by Katie Paul, NEW YORK, Sept 22 2026, retrieved in full verbatim via licensed syndication

secondary named-outlet accountable journalism
https://www.ksl.com/article/51627159/ ↗
[2]

Meta on-record spokesperson statement, obtained directly by Gizmodo (spokesperson named as Dave Arnold; statement text matches the one Reuters attributes to spokesperson Daniel Roberts)

secondary named-outlet journalism
https://gizmodo.com/report-claims-metas-ai-agents-sometimes-tagged-in-human-contractors-and-that-one-of-them-got-racist-2000815808 ↗
[3]

BNN Bloomberg syndication of the same Reuters text, with the "half of employees" and opt-out detail

secondary syndicated wire copy
https://www.bnnbloomberg.ca/business/artificial-intelligence/2026/09/22/meta-testing-a-human-concierge-for-its-new-personal-ai-agent-muse-reuters-exclusive/ ↗
[4]

TNW and PYMNTS write-ups of the Reuters exclusive

tertiary tech press
https://thenextweb.com/news/meta-muse-ai-agent-human-concierge-calls ↗
[5]

Background on Muse existence and launch date (Sept 8 2026)

secondary Axios, TechCrunch, Bloomberg
https://www.axios.com/2026/09/08/meta-debuts-muse-personal-ai-agent ↗
How links are chosen. A source is linked only when the address comes from the investigation's own retrieval or from a registry lookup (PubMed, Crossref) that matches the citation's title and year. Author lists shown as registry-verified come from the registry record, not from the report text. Citations that cannot be matched are labeled, never guessed.
This is one case on the record See the full case, browse the archive, and search every checked claim on TrueSeeker AI Open on ai.trueseeker.com →