Bloomberry MCP Server: AI-Writing Guardrails for Employee Advocacy

AI detectors try to prove authorship. Bloomberry is solving the business problem: stopping employee advocacy content from sounding generic, synthetic, or off-brand before it gets published.

---

The problem with AI detectors

When teams started worrying about AI-generated content in employee advocacy, the natural instinct was to reach for an AI detector. Run the post through a classifier. Get a probability score. Decide if it passes.

That instinct is wrong.

OpenAI built and then discontinued its own AI Text Classifier after finding it correctly identified only 26% of AI-written text — while incorrectly labelling human-written text as AI-generated 9% of the time. They shut it down and stated it was "not performing well enough."¹

Stanford HAI researchers found something more troubling: AI writing detectors disproportionately flag text written by non-native English speakers as AI-generated, even when it is entirely human-written.²

So you have a tool with documented low accuracy that systematically mislabels writing from non-native English speakers. Deploying this in an employee advocacy workflow is not a quality improvement. It is a liability.

---

The real business problem

The reason brand teams worry about AI-generated advocacy content is not authorship. It is quality.

When an employee's LinkedIn post says "in today's fast-paced world, we need to leverage cutting-edge solutions to unlock our full potential" — the audience disengages. The employee looks inauthentic. The brand association is negative. The post underperforms.

That problem is solvable without making any authorship claim.

We do not need to know if AI wrote the post. We need to know: does it sound generic? Does it follow AI cadence templates? Is it off-brand? Those are answerable questions.

---

What we built

Bloomberry's AI Writing Scan is not an AI detector. It is an AI-writing guardrail system.

What it does:

• Scans text against a production enforcement corpus of 5,698 AI-writing signal entries
• Returns structured findings: risk score (0–100), flagged signals, banned phrase matches, sentence-DNA pattern matches, and rewrite suggestions
• Never claims to determine authorship
• Includes a mandatory disclaimer on every response: "This does not prove whether AI wrote the text. It only identifies language patterns that may sound generic, synthetic, or off-brand."

The corpus breakdown:

• 1,600+ multi-word generic phrases — "game-changer", "in today's fast-paced world", "dive deep"
• 135 vocabulary clichés — "leverage", "seamlessly", "unlock", "cutting-edge"
• 78 regex pattern detectors — structural signal patterns
• 12 cadence structure detectors — "It's not just X, it's Y", "Not X. Y.", excessive rhetorical contrast
• 2,900+ aiSlop strings — Bloomberry's production writing avoidance corpus
• 287 replacement pairs — specific alternatives for overused phrases

---

What is AI Sentence DNA?

AI Sentence DNA is Bloomberry's research taxonomy of recurring writing patterns that appear at elevated frequency in AI-generated content.

These are not random. They emerge because large language models are trained on human feedback that consistently rewards certain structural patterns: rhetorical contrast, abstract-noun stacking, hook formulas, transition filler, and vocabulary clichés. The model learns that these patterns score well in evaluation. So it produces them at scale.

The result: AI-generated content from different models, different companies, and different prompts often sounds the same. Not because they share a model, but because they share a training signal.

The Bloomberry AI Sentence DNA research corpus contains 7,622 total audited signal entries (public label: 7,400+ catalogued AI-writing signal entries). The published layers:

Layer	Count
Static ESM corpus raw entries	6,246
Runtime persona-specific bans	+61
Finite regex surface-form expansion	+703
Reviewed source-backed research entries	+612
Total defensible signal entries	7,622

The V1 product scanner enforces 5,698 production entries — the directly-importable ESM corpus. The research corpus contains additional entries in the reply-pipeline and research-only tracked entries not yet in production enforcement.

See the full methodology: AI Sentence DNA research →

---

The MCP integration

The technical delivery is real: a functioning MCP server using the Model Context Protocol SDK with stdio transport. Three tools callable from Claude Desktop or Cursor:

• scan_ai_sounding_text — full scan with risk score, findings, and corpus stats
• check_banned_phrases — match against the production phrase corpus
• analyze_sentence_dna — detect AI cadence structures

The server is local, stateless, and does not access any Bloomberry account data, Voice Memory, or workspace information in V1. It enforces the production enforcement corpus — not "7,400+ entries." The MCP response includes both counts explicitly so there is no ambiguity.

Setup guide: Bloomberry MCP Server docs →

---

Who this is for

This matters for:

Comms and brand teams running employee advocacy programs who need a way to catch generic, synthetic-sounding drafts before they reach employees' feeds.

Founders and executives whose thought leadership depends on sounding authentically human — not like every other LinkedIn post.

Agencies managing executive ghostwriting who need guardrails to ensure client content does not drift into AI dialect.

Sales and social selling teams whose outreach immediately signals inauthenticity when it uses AI cadence patterns.

Cursor and Claude users who want to add writing quality checks into agentic drafting workflows.

---

What comes next

V1 is the foundation. The roadmap includes:

• Authenticated hosted MCP/API — so teams can use the scanner without running it locally
• Voice Memory comparison — compare a draft against an individual's voice profile to detect drift, not just generic patterns
• Team-specific guardrails — custom phrase libraries and banned patterns per team or campaign
• Approval workflow integration — surface writing quality signals inside Bloomberry's existing approval queue

The category is AI-writing guardrails for employee advocacy. Distribution tools solve reach. We are building the quality layer that protects human voice before content gets distributed.

---

Try it: ai-writing-scan → · MCP setup → · Research corpus →

Footnotes

1. OpenAI, "New AI classifier for indicating AI-written text" ↩

2. Stanford HAI, "AI Detectors Biased Against Non-Native English Writers" ↩