Verification Report 002

Emergence Verification Report 002 — Meta content moderation appeals

Dimensions: Authority Boundaries, Human Agency.

Claim framing (per instruction — do NOT overstate)

The question is not "does Meta promise every appeal gets human review." The public evidence does not support that. The question is:

When Meta says appeals are reviewed by human content reviewers, where does human review actually occur, and where can consequential decisions remain fully automated?

Decision pathway (mapped from documented structure)

1. Automated enforcement — Meta's systems detect and remove violating content, largely automatically (much of it pre-report).
2. Internal appeal — a user appeals the removal; the appeal may be resolved automatically or by a human reviewer.
3. Oversight Board appeal — after exhausting internal appeals, a user may appeal to the Oversight Board.
4. Board review — the Board selects a small subset of cases for review and issues a decision.
5. Binding outcome — selected Board decisions are binding on Meta unless implementation would violate the law.

Answers to the ten questions

1. Which moderation decisions can be made fully automatically?

   Initial enforcement is largely automated. Meta's systems remove a large share of violating content without human involvement, including before any user report.

2. Which internal appeals can also be resolved automatically?

   Some. In the Al-Shifa Hospital matter, the Board documented that both the original removal and the rejection of the user's appeal were made automatically from a classifier score, with no human review. [1]

3. Under what documented conditions does a human reviewer enter?

   In the Breast Cancer Awareness Content decision, the Board quotes Meta stating that the majority of appeals are reviewed by content reviewers, and that nudity-removal appeals are reviewed by a content reviewer except where Meta has capacity constraints. [2]

4. Are users told whether a decision was automated or human?

   Not systematically. Meta does not, as a matter of documented routine, disclose to each user whether their specific appeal was resolved by a human or an automated system. (Transparency gap.)

5. Can users meaningfully contest an automated decision?

   Partially. There is a contestation path — internal appeal, then Oversight Board — but its meaningfulness depends on whether the internal appeal reaches human review.

6. What happens when human-review capacity is constrained?

   Meta documents that capacity constraints can cause appeals to be handled without human review — the documented exception to the "majority human-reviewed" norm. [2]

7. How selective is access to the Oversight Board?

   Highly selective. The Board received 558,235 appeals in 2024 and necessarily selected only a tiny subset for case review. [3]

8. Does the Oversight Board constitute meaningful human oversight for the ordinary user?

   No — it is a high-level backstop, not routine oversight. It is an independent check that reviews a tiny subset of cases and issues binding decisions that shape policy, but it does not supply human review for the ordinary user's individual appeal.

9. Where does final authority sit at each stage?

   Initial enforcement: Meta's automated systems. Internal appeal: Meta (automated or human, per capacity). Board case: the Oversight Board (binding on Meta unless illegal). Policy: the Board's binding decisions and recommendations shape Meta's rules.

10. What evidence would change the assessment?

    Meta publishing complete human-vs-automated review rates across all appeal categories; Meta disclosing to each user whether their decision was automated or human; data on the actual share of appeals that reach human review.

Finding

MIXED.

Human recourse is real and documented at some stages:

• The majority of certain appeal categories are reviewed by humans. [2]
• An independent, binding Oversight Board exists as an external check with documented authority (Charter, Bylaws, Trust Agreement, binding decisions).

Serious automation gaps exist:

• A consequential decision can be automated end to end — original removal and appeal rejection both from a classifier score, with no human review. [1]
• Capacity constraints produce documented exceptions to human review. [2]
• The strongest external recourse is available only in a tiny subset of appealed cases (558,235 appeals in 2024; only a small subset selected for review). [3]
• Users are not systematically told whether a given decision was automated or human.

Severity: CONCERNING (bounded, not SEVERE — recourse exists and the Board is a real backstop).

Confidence: CONFIRMED (structural facts and the specific automation/capacity claims all trace to Oversight Board primary material).

Citations

1. Oversight Board, "Oversight Board Issues First Expedited Decisions About Israel-Hamas Conflict" (Al-Shifa Hospital). https://www.oversightboard.com/news/1109713833718200-oversight-board-issues-first-expedited-decisions-about-israel-hamas-conflict/
2. Oversight Board, "Breast Cancer Awareness Content." https://www.oversightboard.com/decision/bun-0w49p93l/
3. Oversight Board, 2024 Annual Report. https://www.oversightboard.com/wp-content/uploads/2025/08/2024-Annual-Report.pdf

Limitations

• This review maps the documented authority/appeals structure only; it does not infer hidden Meta workflows (per instruction).
• The "users are not systematically told whether a decision was automated or human" finding is an absence-based observation (no documented routine disclosure), not a claim that disclosure never occurs.

What would change the assessment

• Toward SEVERE: evidence that human review is the exception rather than the majority for consequential appeals, or that no meaningful contestation path exists for ordinary users.
• Toward NEUTRAL/PROTECTIVE: evidence that the large majority of consequential appeals reach human review and that users are told when a decision is automated.