Before You Trust an AI Detection Score, Read This
You may already use AI to draft marketing copy, reply to customers, summarise documents, or prepare internal updates. At the same time, your customers, employees, suppliers, and business partners may be wondering whether the words they receive were written by a person or generated by software.
That creates a practical problem. An AI detector can appear to offer a simple answer: this document is “78 percent AI-generated” or “100 percent AI-generated”. But a percentage can look more certain than the evidence behind it. If you use a score to reject a job applicant, accuse an employee, challenge a supplier, or question a customer’s submission, you could damage a relationship based on an imperfect result.
The recent attention around Pangram, an AI-detection startup, shows why caution matters. The company has been used to assess books, journalism, prize-winning stories, and online writing, while critics have questioned how detector results are interpreted and shared. The lesson for your business is straightforward: an AI score should start a conversation, not end one.
TL;DR
AI detectors make educated predictions by comparing text patterns with writing produced by language models. They do not provide conclusive proof of who wrote a document.
Use detection tools as one review signal, alongside drafts, version history, interviews, source checks, and clear disclosure rules. Never make a serious people decision from a score alone.
What This Means
Pangram describes its approach as “synthetic mirroring”. In plain language, the system studies human writing and compares it with similar text produced by AI models. It also looks for examples where detectors wrongly identified human writing as AI-generated, then uses those examples to improve its model. The source article says Pangram uses licensed datasets and serves sectors including education, legal work, recruitment, and creative writing. Source: WIRED
This is a classification exercise, not a time machine. The detector cannot watch your employee type. It cannot know whether a writer used an AI tool to produce an outline, corrected grammar with software, translated a passage, or copied a complete answer. It studies the final text and estimates how likely certain patterns are to resemble AI output.
That distinction matters because writing style varies. A short, formal email may look machine-like even when a person wrote it. A fluent writer may produce text that resembles AI output. A person can also edit AI-generated text until it looks more natural. A percentage therefore represents a model’s confidence, not a proven record of authorship.
Use AI detection to ask better questions—not to replace human judgement.
The controversy around Pangram also highlights a second issue: process. The article describes how a manuscript was scanned and discussed publicly after being obtained from a piracy website, raising questions about consent, data handling, and responsible use. Source: WIRED For your business, the same principle applies. Before uploading confidential contracts, customer records, product plans, or employee writing to an external tool, check what happens to the data.
How This Applies to Malaysian SMEs
1. Recruitment and job applications. You may receive a cover letter, writing sample, sales proposal, or customer-service scenario from a candidate. An AI detector may flag the document, but that does not automatically mean the candidate is dishonest. They may have used software for grammar correction, translation, or organising ideas. Instead of rejecting the person immediately, ask them to explain their approach or complete a short live task related to the role. For example, a candidate for a sales position could respond to a realistic customer complaint in their own words. Compare the reasoning, clarity, and ability to answer follow-up questions.
2. Marketing and social media. Many Malaysian SMEs now publish content in Bahasa Malaysia, English, Mandarin, Tamil, or a mixture of languages. Text written or translated with AI may be helpful, but your brand still needs human review. A detector score does not tell you whether a post is accurate, culturally suitable, or aligned with your tone. Your better control is an approval checklist: verify claims, check names and locations, remove exaggerated promises, and make sure the final version sounds like your business. If a customer asks whether content used AI, prepare an honest internal policy rather than relying on a detector to prove anything.
3. Customer service and sales communication. AI can help your team draft replies more quickly, especially for frequently asked questions. However, a polished response can still be wrong or insensitive. A detector is not designed to check whether the reply promises something your company cannot deliver. Train staff to verify order status, warranty terms, delivery details, and personal information before sending. In many cases, accuracy and accountability matter more than whether a machine helped create the wording.
4. Supplier documents and proposals. You may receive a quotation explanation, technical proposal, tender response, or business plan that appears heavily AI-assisted. Instead of treating that as misconduct, test the substance. Ask the supplier to clarify assumptions, explain implementation steps, identify risks, and provide relevant examples. A company that understands its proposal should be able to discuss it clearly. A detector score alone cannot tell you whether the supplier can perform the work.
5. Internal employee work. If you suspect a staff member used AI against company policy, preserve the original document and review the work process. Look at version history, notes, references, and the employee’s explanation. The important question may be whether confidential data was entered into an external service, whether the information is accurate, and whether the employee followed your instructions. Handle the matter privately and consistently. Public accusations based on an automated score can harm morale and expose your company to unnecessary disputes.
Practical Takeaways
- Write a clear AI-use policy. State what is allowed, what requires review, and what must never be uploaded to external tools.
- Define acceptable assistance. Distinguish between spelling correction, translation, brainstorming, summarising, and fully generated content.
- Keep human accountability. The employee who sends or approves a document remains responsible for its accuracy.
- Use detectors only as a secondary signal. Treat a high score as a reason to investigate, not as proof.
- Ask for supporting evidence. Request drafts, research notes, version history, references, or a short verbal explanation.
- Test capability directly. Use live exercises, follow-up questions, or demonstrations relevant to the role.
- Protect confidential information. Review the detector’s privacy terms before uploading business or personal data.
- Record decisions consistently. Apply the same process to every employee, applicant, and supplier in comparable situations.
- Review multilingual content carefully. AI performance and detection reliability can differ across languages and writing styles.
A simple review process
| Step | What you do | Why it matters |
|---|---|---|
| 1 | Scan only when there is a legitimate reason | Prevents routine surveillance and unnecessary data exposure |
| 2 | Record the tool, date, document version, and result | Keeps the review transparent and repeatable |
| 3 | Check drafts, sources, edits, and context | Provides evidence beyond a statistical prediction |
| 4 | Speak privately with the person involved | Allows misunderstandings and legitimate assistance to be explained |
| 5 | Make a decision using several signals | Reduces the risk of an unfair conclusion |
This process does not require a large compliance department. You can put it into a one-page standard operating procedure, assign one manager to handle reviews, and keep a basic record in your document system. The key is consistency. If one employee is questioned for using grammar software while another is praised for using the same tool, your policy will quickly lose credibility.
The Bigger Picture
AI detection will probably become more common as businesses, schools, publishers, and online platforms try to understand how content was produced. The source article reports that Substack integrated Pangram so readers could assess possible AI use, while Pangram also serves education, legal, and recruitment sectors. Source: WIRED This points to a future where disclosure, provenance, and review processes matter as much as the final wording.
But detection tools will remain limited because generation and editing methods keep changing. A text can pass through several stages: a person writes an idea, AI creates a draft, a worker edits it, translation software changes the language, and another person approves the result. Asking whether the text is simply “AI” or “human” may be less useful than asking: Who approved it? What information was used? Was confidential data protected? Can the author explain the content? Is the result accurate and fit for purpose?
For a Malaysian SME, trust will come from your process. Tell your team what responsible AI use looks like. Give customers confidence that important communications are checked by people. Ask suppliers and applicants for evidence of understanding rather than relying on a mysterious percentage. Most importantly, do not let an automated label replace a fair, documented decision.
The practical standard is simple: use AI where it helps your team work better, keep humans responsible for the outcome, and treat every detection score as an uncertain signal that needs context.
Ready to Streamline Your Operations?
Your business should run itself. AutoRunBiz deploys AI agents to automate your daily operations — WhatsApp orders, invoicing, customer follow-ups, and accounting. Book a free 15-min ops audit to see where automation fits your business →