AI Detection Can Help—But It Should Not Make Decisions Alone
You may already use AI to draft marketing copy, reply to customers, summarise documents, or prepare internal reports. Your staff may do the same. That creates a practical problem: if a customer, employee, or business partner asks whether a piece of writing was produced by AI, how can you check?
AI detection tools appear to offer a simple answer. Paste in the text, receive a percentage, and decide whether the writing is “human” or “AI-generated”. But that percentage can look more certain than it really is. A detector is making a prediction, not revealing the writing process with absolute proof.
This matters if you run a Malaysian SME. A false accusation could damage trust with an employee, freelancer, customer, student applicant, or business partner. A sensible process treats detection software as one signal among several—not as a final verdict.
TL;DR
AI detectors such as Pangram estimate the likelihood that text involved an AI model, but their score is not conclusive evidence.
Use detection results to start a conversation and review your process. Do not reject work, discipline staff, or accuse someone based on a score alone.
What This Means
Pangram is an AI startup that analyses writing and provides a best-guess percentage for how much AI may have been involved. According to the source article, the company has 24 employees and had raised US$13 million at the time of reporting. Source: WIRED
Its method includes “synthetic mirroring”. In plain language, the system studies human writing and asks language models to produce similar text. It then learns patterns that may distinguish machine-assisted writing from human writing. The company also describes “hard negative mining”, which means collecting examples that could produce false alarms and using them to improve the system. Source: WIRED
The important point is that detection tools do not watch the writer create the document. They inspect the finished text and compare its language patterns with examples in their training. A high score may mean the writing resembles AI-generated text. It does not automatically prove that an AI tool wrote the entire document, or that the writer acted dishonestly.
A detection score is a prompt for investigation, not a substitute for evidence.
Even the wider industry has reasons to be cautious. The article describes accusations involving novels, journalism, prize-winning stories, and online publishing. Some disputed texts received very high scores, while the authors or organisations involved challenged the conclusions or the way the scans were used. Source: WIRED
How This Applies to Malaysian SMEs
1. Staff documents and customer communication
Your team may use AI to improve grammar in emails, translate a message into Bahasa Malaysia or English, prepare a first draft of a proposal, or organise meeting notes. If you run every polished document through a detector, you may get an alarming score simply because the writing is formal, consistent, or heavily edited.
Instead of banning AI outright, define acceptable use. For example, staff may use AI for brainstorming and language improvement, but a human must check facts, adjust the tone, protect confidential information, and take responsibility for the final message. This approach is more useful than asking whether every sentence was typed without assistance.
2. Freelancers, agencies, and outsourced work
Many Malaysian SMEs work with freelance copywriters, social media managers, designers, translators, and virtual assistants. You may want original work, especially when the content represents your brand. An AI detector can be part of your review process, but it should not replace the contract, briefing process, or quality checks.
Ask for practical evidence instead. Request an outline, source list, draft history, or short explanation of the writer’s approach. Review whether the work matches your brief, reflects your industry accurately, and avoids copied claims. If the detector returns a high score, speak with the provider before withholding approval or ending the relationship.
3. Recruitment and job applications
Recruiters and business owners may be tempted to scan cover letters, written tests, or interview assignments. Be careful. A candidate who uses AI to correct English may receive a high score even though the ideas and experience are their own. Conversely, a person can produce weak or misleading work that passes a detector.
For important roles, use a short follow-up conversation or live task. Ask the applicant to explain their reasoning, improve a paragraph, respond to a customer scenario, or discuss the assumptions behind their proposal. This tests understanding directly and gives you better information than a percentage alone.
4. Sensitive business information
Before uploading contracts, customer records, financial documents, product plans, or employee information to any external tool, check its privacy terms and internal approval requirements. The source article notes that Pangram says its datasets are properly licensed, but this does not automatically answer every question about how your uploaded text is handled. Source: WIRED
Create a simple rule: remove names, account numbers, identification details, and confidential commercial information before using an external writing or detection service, unless your approved process clearly allows it.
Reading an AI Detector Score Responsibly
| Score or result | What it may suggest | What you should do |
|---|---|---|
| Low likelihood | The text does not strongly resemble the tool’s AI examples | Continue normal quality, accuracy, and originality checks |
| Medium likelihood | The text contains patterns the tool considers AI-like | Review the draft, ask about the process, and check sources |
| High likelihood | The text strongly resembles the tool’s AI examples | Request clarification or supporting drafts; do not treat it as proof |
| Any result | The tool may still be wrong or incomplete | Consider context, human review, and other evidence together |
This table is a decision guide, not a claim that a particular percentage has a universal meaning. Different tools may produce different results, and their models can change over time. The source article identifies other services in this area, including Originality.ai, GPTZero, and Turnitin. Source: WIRED
Practical Takeaways for Your Business
- Write an AI-use policy in plain language. Explain what staff may use AI for, what requires approval, and what information must never be pasted into public tools.
- Separate assistance from authorship. Grammar correction, translation, research organisation, and full content generation are different activities. Your policy should distinguish them.
- Keep human accountability. The person submitting a document remains responsible for facts, tone, confidentiality, and compliance.
- Do not use one score as disciplinary evidence. Combine the result with drafts, version history, interviews, source checks, and work quality.
- Ask for disclosure where appropriate. For client-facing content, decide whether your business should disclose substantial AI assistance.
- Check tools before adopting them. Review privacy handling, data retention, supported languages, reporting methods, and whether the tool explains its limitations.
- Use a second-person review for important work. A colleague can check factual errors, unnatural wording, copied phrases, and missing local context.
The Bigger Picture
AI-generated and AI-assisted writing will become harder to separate from traditionally written text. The source article reports that a survey of 1,481 working writers found 61 percent used AI tools, with 7 percent saying they had published AI-generated work. Source: WIRED Whether or not those figures represent your sector, the direction is clear: businesses need rules for responsible use, not just tools for suspicion.
For SME owners, the long-term question is less “Was AI involved?” and more “Was this work accurate, suitable, original enough for its purpose, and reviewed by someone accountable?” A customer usually cares whether your answer solves their problem. An employee needs to know what responsible assistance looks like. A manager needs a fair way to assess performance.
Detection technology may improve, but no tool should remove your judgement. Build a workflow that records who approved important content, where facts came from, and what review took place. That gives you a stronger business process than relying on an attractive percentage.
Start small: choose one content workflow, such as marketing posts or customer emails, and document your approved AI steps. Train your team to disclose substantial assistance internally, protect sensitive information, and challenge detector results respectfully. This lets you benefit from productivity tools while keeping trust at the centre of your business.
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 →