What is a humanizer?

Last week, we announced that Pangram is now able to detect AI humanizers. But, you might be wondering, what is a humanizer? And who are they for?

Imagine this: You're a college student. You've just written an essay using ChatGPT or plugged your homework questions into an AI. However, you know their submissions will be checked for AI content. So, you're looking for a way to hide the true origin of the content.

Humanizers are a solution to your problem: you copy and paste your text through a humanizer to “hide” the fact that it was written by AI.

Multiple humanizers advertise bypassing Turnitin AI detection as a key feature. Turnitin is a tool designed to detect plagiarism and AI writing in schools and universities.

What do humanizers actually do?

Humanizers employ a broad range of techniques to modify, distort, and obfuscate AI text to hide its true authorship. To understand the effect they have on given pieces of text, Pangram researchers collected writing samples from 19 publicly available humanizers and manually assessed the effects humanization had on each sample. Here’s what we found:

Humanized Text Audit

During our humanized text audit, we picked up on several techniques that humanizers repeatedly use. These include:

Synonym Replacement:

Sometimes humanizers try to evade detection by swapping out words with their synonyms. However, this can damage the text, as synonyms don’t always have the same precise meaning. For example, take the following sentence:

Original Sentence: I need to get my car fixed because the engine is making a strange noise.

Synonym-replaced sentence: I require to obtain my vehicle repaired because the motor is creating a peculiar sound.

While each word obstensibly means the same thing, this damages the text and obviously lowers its clarity and fluency, sometimes to comedic effect.

Nonsensical Phrases:

Other times, humanizers will add nonsensical phrases to a piece of writing in the hopes that AI detectors will decide that the gibberish that is unlikely to have come out of an AI.

Example Nonsense:

...as a result, will put more effort into their studies. CGSizeMake pp 18-23. Last but not least…

Text Quality:

Example Original Text:

In an era dominated by technology, the convenience of cell phones has transformed the way we communicate and access information. However, this advancement comes with grave consequences, particularly when cell phones are used while operating a vehicle.

Example Quality Degradation:

Now a days technology let us convenience to communicate and recieves info more easly by cellphones, but this like to say the darkside of the technology for use this devices while is driving, because cellphones are a little from a great danger for everybody…

Parting Thoughts

Humanizers are a new class of tool designed to evade AI detection. However, the use of a humanizer is always a risk. New research conducted by Pangram shows that, with subtle adjustments, AI detectors can be trained to detect humanized text, even when the humanizer was built specifically to evade a particular model. Additionally, when you use a humanizer, you risk degrading your text and adding nonsense to your output.

Finally, here at Pangram, our model is robust to detecting more than 90% of high-quality humanized content. If you are interested in learning more, read our announcement here or reach out at info@pangram.com!