AI assistants and search decision
Winston AI
A competent developer can build a useful sentence-level AI detector, OCR, and reporting stack in ~30 hours and run it cheaply, but reproducing Winston AI’s claimed accuracy and proprietary dataset-backed advantage (and enterprise polish/support) is unlikely without their data and sustained R&D.
Visit website↗Open-source builds that already do this
Every project below is open source and already does this job today. Fork one, self-host it, or take the parts you need — the build prompt further down assumes an empty file, and this is the shortcut past that. Licences differ; check the one on each card before you ship.
What a replacement has to do
- Accept text or file input → extract text (OCR for images) → run detection model to produce sentence-level AI/human scores and plagiarism matches → generate shareable report / API response
What it still won’t have
- Winston AI’s claimed proprietary training dataset and any accuracy advantage from it
- Weekly model updates and ongoing model-research improvements
- Enterprise-grade SLAs, dedicated support, and any validated institutional certifications
- Pretrained integrations and polished UX (reports, team management, certification)
What remains hard
- Proprietary data
Trained on the largest dataset of human reviewed data to minimise false positives.
- Brand trust
Trusted by 10M+ users
First-year cost
Keep paying
Paying is—cheaper in year one.
On cash alone, building overtakes the subscription at 9 seats.
Money you would actually spend
Time you would spend
—
What you would spend
What we assumed
The verdict above measures whether you could build it. This one is only about money.
Runnable build prompt
Build a self-hosted AI-content-detection web service using Python (FastAPI), a small vector DB (Postgres+pgvector), Tesseract for OCR, and an open-source/small hosted classifier for sentence-level AI/human scoring. Scope: accept pasted text, URL import, and .docx/.png/.jpg uploads; extract text with OCR; split into sentences; compute embeddings and run a classifier to produce per-sentence AI vs human probability and plagiarism hits (nearest-neighbour on embeddings); generate a shareable PDF/HTML report and a JSON API. Out of scope: training large custom LLMs from scratch, enterprise SSO, and multi-tenant billing. Include input validation, rate-limiting, error handling, and unit tests for extraction, inference, and report generation. Provide a Docker Compose setup and deployment docs for a $20/month VPS and a $100/month inference budget.
How we checked
How the score was reached
- Partly verdict base52
- An open-source build was found+5
- 4 cited sources+3
- Price verified on pricing page+3
- 3/3 assessment runs agreed+4
- Hard moats found in the evidence-3
- Evidence score64
The base comes from the verdict. Everything under it is a check that either happened or did not, and each one is a fact frozen in this record rather than a judgement made at render time — so the same evidence always produces the same number.
How scoring works →Cited sources · 4
Every page the run actually retrieved.
- official productWinston AI — homepage
- official pricingWinston AI — pricing
- open sourcearc53/DocsGPT
- open sourceaingdesk/AingDesk
Integrity checks
What held up, and what did not.
