EyeSift

Editorial & Research Team

How EyeSift audits AI detection models, evaluates false-positive risks, and maintains transparent client-side screening tools.

Sasha Brenner, Founder & Lead Computational Linguistics Editor at EyeSift

Sasha Brenner

Founder & Lead Computational Linguistics Editor

Sasha Brenner leads EyeSift's research into statistical text forensics and media provenance. Before launching EyeSift in 2025, Sasha worked in natural language processing and editorial verification workflows, studying how autoregressive language models leave measurable distributional traces in token predictability, clause cadence, and lexical diversity.

Frustrated by commercial AI detectors that hide uncertain classifiers behind paywalls and present probabilistic guesses as proof of misconduct, Sasha designed EyeSift around open stylometric telemetry: sentence-level burstiness mapping, C2PA JUMBF manifest inspection, and explicit false-positive warnings grounded in peer-reviewed benchmarks such as RAID (ACL 2024) and Stanford HAI's ESL evaluation studies.

  • Core focus: Stylometric feature engineering, C2PA/EXIF metadata forensics, and detector benchmark design
  • Verification protocol: Cross-testing commercial detectors against human academic, journalistic, and ESL corpora
  • Review cycle: Re-calibrating thresholds as frontier models (GPT-5, Claude 4, Gemini 2.5) evolve
  • Direct contact: editor@eyesift.com

How We Evaluate AI Detection Tools

Every guide, tool review, and heuristic analyzer published on EyeSift follows a strict forensic review protocol:

  1. Inspectable signals over black-box labels. Our browser-based analyzers expose exact sentence-length variance, trigram repetition, and metadata markers so reviewers can see why a passage or file was flagged.
  2. Empirical false-positive testing. We evaluate detectors not only on raw LLM output, but on human-written technical prose, non-native English essays, and heavily edited drafts where false accusations cause the most harm.
  3. Cryptographic provenance first. Wherever available, we prioritize verifiable C2PA Content Credentials, SynthID watermarks, and original draft history over statistical guesswork.
  4. Human-in-the-loop ethics. No statistical AI score should ever serve as sole evidence in academic discipline, hiring, or editorial termination.

Scope & Limitations

EyeSift provides educational benchmarks and browser-side screening utilities. Statistical text and media analysis is probabilistic by nature and does not constitute court-admissible digital forensics or definitive proof of authorship.

Reporting Corrections

If you identify outdated vendor pricing, a changed API endpoint, or a benchmark discrepancy, email editor@eyesift.com. Verified corrections are deployed within 48 hours and noted in the article revision timestamp.

Editorial Independence

EyeSift does not accept paid rankings, vendor sponsorships, or affiliate commissions from commercial AI detectors or writing tools. Revenue comes solely from programmatic display advertising on informational pages.

Related