Scan employee advocacy drafts for AI-sounding language, generic phrasing, sentence-DNA patterns, and off-brand writing before employees publish — without making authorship judgements.
An AI-sounding writing scan checks text against a corpus of catalogued AI-writing signal entries — vocabulary clichés, generic phrases, sentence-DNA patterns, and cadence structures — and returns structured findings with risk scores and rewrite suggestions. It identifies patterns that sound AI-generated. It does not determine whether AI wrote the text.
AI detectors attempt to classify whether a human or an AI wrote a piece of text. That is a claim with well-documented accuracy limitations — OpenAI discontinued its own classifier, and independent research has found high false-positive rates, especially for non-native English writers. Bloomberry does not make authorship claims.
"This does not prove whether AI wrote the text. It only identifies language patterns that may sound generic, synthetic, or off-brand."
See also: Why Bloomberry is an AI detector alternative, not an AI detector →
| Dimension | AI Detector | Bloomberry AI Writing Scan |
|---|---|---|
| Claim | Determines if AI wrote the text | Identifies AI-sounding signal patterns |
| Output | Authorship probability score | Risk score + specific flagged signals + rewrites |
| False positives | High — especially for non-native writers | N/A — never makes authorship claims |
| Use case | Content policing | Brand voice quality control before publishing |
| What it answers | "Did AI write this?" | "Does this sound generic, synthetic, or off-brand?" |
| Actionable output | Pass / fail label | Specific signals, examples, and rewrite suggestions |
Bloomberry publishes two honest numbers. They are not contradictory — they describe different scopes.
| Count | Label | What it includes |
|---|---|---|
| 7,622 | Research-audited total | Full AI Sentence DNA corpus including research-only entries, regex surface forms, and persona runtime bans. Public label: 7,400+. |
| 5,698 | Production enforcement corpus | Directly-importable ESM modules enforced in the V1 product scanner and MCP server. Does not include research-only entries. |
Legacy employee advocacy tools solve distribution: they get approved content in front of employees to share. The quality problem is different. When employee posts sound generic — "leverage synergies", "in today's fast-paced world", "not just X, it's Y" — the audience disengages, the employee's credibility drops, and the brand association becomes negative.
Bloomberry is bringing MCP-native AI-writing guardrails to employee advocacy. The V1 MCP server uses stdio transport and exposes three tools callable from any MCP-compatible agent or IDE.
scan_ai_sounding_textFull scan: risk score, findings, banned phrase matches, sentence-DNA matches, and rewrite suggestions.
check_banned_phrasesCheck text against the production phrase corpus and return matches with severity and replacement guidance.
analyze_sentence_dnaDetect sentence structures that make writing sound synthetic, overly polished, or AI-assisted.
Full MCP setup guide: install, configure, and run the server in Claude Desktop or Cursor →
Bloomberry does not detect, prove, or confirm AI authorship. High signal counts do not mean AI wrote the text.
Human writers use the same phrases. Signal presence is not proof of anything about the author.
The production scanner enforces 5,698 production AI-writing signal entries. It does not enforce the full 7,622-entry research corpus.
The MCP server V1 does not compare against individual Voice Memory profiles. That comparison requires authentication and is not in V1.
Risk scores are deterministic pattern counts, not probabilistic authorship models.
Stop AI-sounding language before it reaches your audience
Try Bloomberry's AI-writing guardrails for employee advocacy content.