Two different questions
ChatGPT watermark vs AI detectors
A watermark check looks for an embedded provenance signal. An AI text classifier estimates origin from patterns in the writing. Their results are not interchangeable.
Embedded signal or predicted origin?
OpenAI distinguishes watermarking during generation from classifiers that analyse writing afterwards. Its older classifier was trained on human and AI examples; OpenAI retired it on 20 July 2023 because of low accuracy. That history concerns a different system from textGrain, not a performance test of the new watermark detector.
| Question | OpenAI watermark check | AI text classifier |
|---|---|---|
| What is examined? | A supported embedded statistical signal. | Text patterns used to estimate origin. |
| Is a watermark required? | A detectable matching signal is needed. | No embedded watermark is required. |
| What does a negative mean? | No supported signal was found. | The model did not classify the sample as AI-generated. |
| Does it establish the author? | No. | No; an estimate is not authorship evidence on its own. |
Sources: Earlier OpenAI classifier and its retirement ↗ · OpenAI provenance guide ↗
How can the results disagree?
As an illustration, a classifier might label a paragraph as AI-like while a watermark check finds no supported signal. Those answers address different questions, so they are not automatically contradictory. This example is a way to interpret the distinction, not a claim that SafeRender has tested or compared either tool.
- Name the tool and the question it was designed to answer.
- Record the checked version and the result’s limitations.
- Avoid converting a label into a claim about the writer.
Which method is more reliable?
There is no defensible winner for every document from the information available here. Compare methods on the same task, content, language and error thresholds. The textGrain report describes its generation and detection procedures and the assumptions behind its guarantees; that is not a guarantee for arbitrary student writing or every competing tool.
Sources: textGrain technical report (PDF) ↗
What should students and reviewers keep separate?
Provenance, permission to use AI, factual accuracy and assignment quality are separate review questions. A sensible review records the evidence for each one instead of using a single score for all four. For a student, this means keeping your drafting process and disclosure alongside the submitted work. For a reviewer, it means asking for context before reaching a conclusion.
Common questions
Is a watermark detector a general AI detector?
No. An OpenAI watermark check targets a supported OpenAI signal. It is not a universal check for text generated by every AI provider.
Does no detected watermark mean a classifier was wrong?
Not necessarily. A classifier estimate and an embedded-signal check address different questions. Compare their scope before comparing their labels.
Is textGrain the OpenAI classifier from 2023?
No. The retired classifier estimated text origin from examples. textGrain is a watermarking mechanism applied during supported text generation.
Does SafeRender offer either detection check today?
SafeRender does not currently offer a public text watermark detector or a general AI text classifier. Its watermark feature has a waitlist, with launch dependent on access and permitted use.
Sources and editorial notes
This is an independent explanation, based on the linked sources and their review dates below. OpenAI’s announcement describes a staged rollout; product documentation may describe the intended scope more broadly. We do not assume that every eligible output is already covered.
- OpenAI provenance guide ↗OpenAI Help Center · Checked
- Earlier OpenAI classifier and its retirement ↗OpenAI · Checked
- textGrain technical report (PDF) ↗OpenAI and research collaborators · Checked
Want to know when you can check your text?
Join the waitlist for a launch notification if SafeRender can offer the detector. No draft needed and no launch date confirmed.