Rules, Detectors, and Watermarks: Why AI Writing Still Slips Through

AI Text Detection: What Works, What Breaks, and What to Do About It

People and institutions want to know who wrote what. Teachers need to assess a student's actual grasp of a topic. Consumers want to know if an ad was written by a person or a model. Writing rules is easy. Enforcing them depends on something messy: deciding whether a piece of text was produced by AI.

Some studies show heavy AI users can spot AI-written text with strong accuracy, and panels of trained evaluators can beat automated tools in controlled tests (2). But that skill set doesn't scale across classrooms, journals, or marketing teams. So organizations turn to automated detectors-and run into hard limits.

How Detection Typically Works

The workflow sounds simple: take a text, run a detector, get a probability score, and act. The simplicity hides assumptions that matter.

• Which AI models could have produced the text?
• Do you have access to those models or their outputs?
• How much text do you have-one paragraph or a portfolio?
• Was watermarking enabled?

Watermarking is a special case. Some AI systems embed subtle markers so that later verification can confirm the text's origin using a vendor-held key (3). It's clean when available, but it depends on vendor cooperation and specific settings being turned on.

Three Families of Tools

1) Learned detectors. Train a classifier on labeled examples of human and AI text, then predict on new input. This can work without knowing the exact generator, as long as the training data is broad enough to cover many systems (1).

2) Statistical/model-based tests. If you have access to the model you care about, you can analyze how likely that model finds the exact word sequence. Unusually high likelihood can be a signal the model wrote it.

3) Watermark verification. With a secret key from the vendor, a verifier checks whether text matches a watermarked pattern. This is verification, not inference, and it relies on infrastructure outside the text itself (3). For a technical overview of watermarking, see Nature's coverage of scalable watermarking approaches here.

Limits You Need to Account For

• Drift hurts learned detectors. Performance drops when new text differs from the training set. As models change, detectors lag unless you keep retraining on fresh data (1, 4).
• Assumptions break statistical tests. Many methods rely on model access and stable behavior. Proprietary models, frequent updates, or unknown sources weaken these tests outside lab conditions.
• Watermarking is conditional. It only helps if it was enabled and supported by the vendor. It doesn't solve detection for general text.
• Arms race dynamics. Public detectors invite evasion. As generators improve and editing tactics spread, no detector stays ahead for long (4).
• Short, edited, or mixed-authorship text is hard. Small samples, heavy paraphrasing, translation, or collaborative drafts reduce accuracy across methods.

Practical Playbook for Educators, Editors, and Research Leads

• Use detectors as signals, not verdicts. Treat scores as one input. Never punish on a score alone.
• Set clear AI-use policies. Define what's allowed (idea generation, grammar checks, summaries) and what crosses the line. Give examples.
• Collect baselines. Keep representative writing samples for each student or contributor. Compare style and reasoning over time instead of relying on one-off judgments.
• Ask for process evidence. Drafts, citations, notes, prompt history, and version timelines help confirm authentic work.
• Triangulate. Combine a learned detector, a model-based test (if accessible), and human review. Log thresholds, rationale, and outcomes.
• Focus on quality criteria. Accuracy, originality, attribution, and method matter more than guessing the source.
• Provide an appeal path. If a decision affects grades, authorship, or careers, require additional evidence beyond detection scores.
• Calibrate and document. If you publish findings on AI prevalence, report false-positive/false-negative rates and uncertainty bands (4).

For Writers and Teams

• Disclose assistance. Note where AI helped and where human judgment led.
• Save artifacts. Keep drafts, prompts, and sources. This protects your credibility.
• Add non-generic value. Data, experience, methods, and original analysis are hard to fake and easy to verify.
• Self-check. Run your own content through detectors to see if it reads like generic AI output. Revise for voice, reasoning, and specificity.

What to Expect

AI text detection is easy to describe and hard to do well. Tools will improve, but certainty is unrealistic. The practical move is a layered approach: explicit policies, process evidence, multiple detection signals, and human judgment.

Tools and Training

If you're building team capability around responsible AI writing and review, explore curated learning paths by role and certification options:

• Courses by job
• Popular AI certifications

References

• 1) Junchao W, et al. A Survey on LLM-Generated Text Detection: Necessity, Methods, and Future Directions. Computational Linguistics 2025;51(1):275-338.
• 2) Russell J, et al. People who frequently use ChatGPT for writing tasks are accurate and robust detectors of AI-generated text. In Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics. 2025;1:5342-5373. Vienna, Austria. Association for Computational Linguistics.
• 3) Dathathri S, et al. Scalable watermarking for identifying large language model outputs. Nature. 2024;634,818-823.
• 4) Pudasaini S, et al. Benchmarking AI Text Detection: Assessing Detectors Against New Datasets, Evasion Tactics, and Enhanced LLMs. In Proceedings of the 1st Workshop on GenAI Content Detection (GenAIDetect). 2025;68-77. Abu Dhabi, UAE. International Conference on Computational Linguistics.