FAQ
Yes. A 2023 result by Sadasivan and colleagues showed that no statistical detector can reliably distinguish AI-generated text from human-written text once the AI text has been sufficiently paraphrased or rewritten. This is a fundamental limit, not a flaw of any specific tool.
What this means in practice:
Cert is most useful when:
Cert is least useful when:
For investigative use against motivated adversaries, treat Cert as one signal among many. Not as definitive evidence.
Death Certificate provides calibrated confidence intervals rather than single accuracy numbers. This is intentional. Single accuracy percentages obscure important per-domain and per-condition variation.
In-domain performance refers to documents matching our calibration distribution: English prose, 2018-2024, mixed sources, no adversarial intent. Cross-domain covers underrepresented domains. Adversarial performance covers paraphrased and obfuscated text.
Each Cert reports its own confidence interval. Wider intervals indicate more uncertainty for that specific document. We prefer explicit per-document uncertainty over a global accuracy number that might mislead.
Phase 0 empirical validation will produce specific accuracy numbers. These will be added to the methodology page when validated.
Different tools for different questions.
Commercial AI detectors classify individual documents as AI-generated or human-written, often producing point estimates without confidence intervals. Documented false positive rates range 5-20% even for the best-performing systems. Applied to people (employment screening, academic discipline, content moderation), these rates produce significant harms.
Deadmeter's Pulse measures populations: aggregated estimates with calibrated confidence bounds. The question is “what proportion of this stream shows AI patterns” rather than “is this specific text AI”. Aggregation across many documents washes out individual classification noise.
Deadmeter's Cert applies to single documents but presents results as diagnostic signals with explicit confidence intervals and honest limitations disclosed upfront. Cert results are versioned and cryptographically signed for verification.
Both kinds of tool have their place. They are not substitutes.
Twitter (now X) closed its public API to non-enterprise tiers in 2023. Available pricing for compatible API access exceeds our operating budget. Alternative paths through scraping or third-party gateways do not meet our stability requirements for weekly publication.
The methodology page documents this exclusion. When public API access conditions change, or when an alternative data path matches our requirements, we will revisit.
Bluesky inclusion is under evaluation for v1.1, conditional on 24+ months of post-launch baseline data, separate baseline methodology, and explicit experimental-stream marking.
Email signup on the homepage or on the /pulse page. We send the weekly Pulse and nothing else. No marketing. No tracking pixels. No promotional sequences. Unsubscribe with one click in any email.
Semantic versioning combined with SHA-256 hashes. Patch increments (v1.0.1) for bug fixes without AUROC change. Minor increments (v1.1) for refinements with under 2 percentage points AUROC change. Major increments (v2.0) for breaking changes that affect historical comparability.
Every Cert and every Pulse publication includes the active methodology version and hash. Historical methodology pages remain accessible at version-specific URLs.
Quarterly transparency reports document any methodology changes.
No. None of our tools produce evidence admissible in legal proceedings. Each tool's disclaimer states this explicitly.
For legal proceedings, you need expert witness testimony based on multiple corroborating evidence types, chain-of-custody documentation, and methodology validated through peer review and court precedent. Our tools provide diagnostic signals appropriate for investigative workflows, not legal verdicts.
Deadmeter does not measure or report on political content. No elections. No government information operations. No active conflicts. No partisan advocacy. No identity politics.
This is implemented through stream selection (excluding political forums), NER-based filters in BS Meter and Dead Internet Index (rejecting political-entity-dense inputs), sponsor category restrictions (auto-rejecting political organizations), and methodology disclosure.
The scope deliberately privileges defensibility over breadth. This is a permanent commitment, not temporary positioning.
Subscriptions and sponsorship. Not AI labs. Not government agencies. Not political organizations. Not venture capital.
The single-operator structure (Pavel Ishchin solo founder, Daniil testing, Evelina marketing) keeps overhead minimal. Founder Lifetime supporters provide bootstrap funding. Pro subscriptions provide recurring revenue.
When sponsor relationships exist, they are disclosed in each Pulse publication with category and amount visibility. Sponsorship does not influence methodology, measurements, or coverage decisions.
Email feedback@deadmeter.com with the issue or correction request. Include URLs of affected pages, Cert hashes if applicable, and specific concern.
Issues affecting methodology accuracy are addressed through the versioning process. Issues affecting tool behavior are addressed through bug fixes. Quarterly transparency reports document any corrections made.