As we all know it Originality.ai is one of those AI detectors that has started to make the rounds across various platforms. But can it truly detect Deepseek? The short answer is YES. The longer answer is the devil lies in the details. Keep reading to know more about it.
Why Deepseek gets detected by Originality.ai?
The simple answer is just like many AI text generators, Deepseek is not specifically made to bypass AI detectors. You can see that when you look at Deepseek’s marketing material as well. They never claim to bypass Originality.ai (or any other AI detector) as a feature. Hence, if it is not made to accomplish this task it won't be able to do it reliably.
Deepseek currently has two major variants of their model — deepseek-chat and deepseek-reasoner. Our analysis on our custom dataset shows the following Originality.ai scores (which basically indicate how “human” the text appears, with 100% meaning fully human):
| Model | Count | Mean | Std Dev | Min | 25% | 50% | 75% | Max |
|---|---|---|---|---|---|---|---|---|
| deepseek-chat | 117 | 0.0093 | 0.0527 | 0.0 | 0.0 | 0.00 | 0.00 | 0.51 |
| deepseek-reasoner | 83 | 3.3102 | 15.2540 | 0.0 | 0.0 | 0.01 | 0.04 | 97.61 |
As you can clearly see, deepseek-chat is almost always scoring around 0% (meaning Originality.ai is highly confident it is AI-generated). The maximum it ever got was around 0.51%. Meanwhile, deepseek-reasoner does show a bit more variance, sometimes spiking as high as 97.61% in a few outlier cases but the majority of times, it is near 0%.
Hence, if you’re using deepseek to bypass Originality.ai, you’re unlikely to succeed on a consistent basis. The few outliers for deepseek-reasoner means it’s not absolutely hopeless, but you can’t reliably bank on it.
Also Read: Can GPTZero.me detect Deepseek?
How Does Originality.ai Works?
It might surprise you, but Originality.ai is not just a simple plagiarism checker—it’s designed from the ground up to catch AI text. They use a multi-layered approach to detect everything from GPT-3, GPT-Neo, ChatGPT, etc. Below is a summary of their detection pipeline:
| Layer | What happens | Key take-aways |
|---|---|---|
| Pre-training | A custom discriminator-style Transformer (inspired by ELECTRA, built on a modified BERT backbone) is first pre-trained on ≈160 GB of mixed web text using a generator–discriminator setup. | A bidirectional discriminator is better at spotting subtle “tells” than the left-to-right generators used to create AI text. (originality.ai) |
| Supervised fine-tuning | Millions of labelled snippets (human vs AI) are added, generated with multiple models (GPT-3, GPT-J, GPT-Neo, etc.) and diverse sampling schemes (temperature, top-k, nucleus). | Diversity at this stage makes the detector robust to different writing styles and sampling tricks. (originality.ai) |
| Token-level scoring | At inference, your document is chunked, embedded, and scored; probabilities are aggregated into a document-level “Original / AI” score (0–100 slider). | The score is a probability, not a percentage of sentences that are AI. (originality.ai) |
| Stylometric heuristics | Separate routines measure perplexity (how predictable the next token is) and burstiness (clumping of rare words). Human prose shows higher, more uneven patterns than most LLM output. | These metrics act as soft features that complement the main model. (originality.ai) |
With such a multi-pronged approach, it is not surprising that Originality.ai manages to catch Deepseek. So if you are planning on using deepseek-chat or deepseek-reasoner to avoid detection, think twice.
Is There Any Workaround?
It’s not that Originality.ai is perfect—there are always some ways to trick AI detectors if you want to, like introducing certain random mistakes or using advanced rewriting strategies. However, if your objective is to consistently bypass Originality.ai with Deepseek, it’s going to be extremely hard. You will have to use a paraphraser like Deceptioner. The outliers in deepseek-reasoner (one text soared to 97.61% “human” score) indicate that it can be done occasionally, but you simply can’t repeat that reliably.
Also Read: Can ZeroGPT detect Deepseek?
One Single Opinion That Might Help
If you’re concerned about Originality.ai flagging your text, the best approach is to write it by yourself or use a purpose-built paraphraser that specifically aims to fool AI detectors like Deceptioner. Deepseek is mostly not made for that. Originality.ai has advanced quite a bit, and you’ll need either a consistently updated rewriter or do manual rewriting to have a fighting chance.
The Bottom Line
Originality.ai easily detects both deepseek-chat and deepseek-reasoner in the vast majority of cases. Only a small fraction of deepseek-reasoner outputs managed to trick Originality.ai. That means if you use Deepseek for bypassing, you’re playing with fire. Originality.ai itself is built on a robust pipeline to detect AI text, so if your generator isn’t specifically trained to trick such detectors, it very likely won’t succeed.
So, be mindful of the fact that AI detectors are evolving rapidly. As they continue developing, rewriters like Deepseek would need to revamp their strategies if they want to stay ahead. For now, though, the numbers speak for themselves, and Originality.ai does appear to detect Deepseek with ease.
