AI Detection and the “Humanization” of Writing

Writers are trying to sound “more human” with the rise in AI-detection software, why that sucks, and why some AI-detection companies might be acting unethically

Modern writers — from student writers in academia to content creators in the marketing world — are facing a new threat: AI-detection software.

This software promises to detect AI-generated text in various forms of writing, from student papers to website copy.

The problem: the software doesn’t always work, and one of the biggest issues — for writers, in particular — is the possibility for false positives, in which the software accuses a piece of writing of being AI-generated when it’s not.

In Geoffrey A. Fowler’s article for The Washington Post, he talks about Turnitin (a company that primarily creates plagiarism-detection software) and their recently developed AI-detection software. The software is flagging some students’ papers as “likely” being AI-generated even when they’re not. The article talks specifically about a star student’s completely original paper being flagged as a likely creation of AI.

The article talks about how detectors are being released without being widely vetted in order to keep up with the evolution and increased usage of AI text-generators like ChatGPT. “Detectors are being introduced before they’ve been widely vetted,” says Fowler. He mentions that Turnitin says that their software should be used as an indication — not an accusation — of cheating. But in the hands of teachers who don’t fully understand AI text generators or AI detectors, how will the detection software actually be used in the classroom?

This is just one of many articles that has recently been published on this issue, and this issue isn’t limited to the classroom. Professional writers of all stripes are facing issues with AI-detection software in their jobs.

I’ve been following Reddit threads in which professional writers 1) complain of being fired for using AI-generated text even though they haven’t, 2) are fearful of being accused of using AI and are trying to sound more “human,” and 3) are contending with AI-detection software and its problems.

For example, the original poster of this Reddit thread, who claims to be in “managerial position in a content writing and digital marketing agency,” talks about working with Originality.AI, which markets itself as being the most accurate AI-detection software on the market for use outside of academia. The poster accuses the software of not working well at all, with one of the main issues being that it flags most content as being generated by AI when there exists solid evidence to the contrary (the poster mentions that the software flagged some of their own original writing as being AI-generated). One surefire way to prevent a piece of writing for being flagged for using AI, the poster says, is to introduce grammar errors into the writing, since the detection software clearly operates on the assumption that AI-generated text is error-free. (I find this laughable.)

So how does AI detection software work?

AI-detection softwares basically reverse engineers predictive text patterns, looking for language patterns that are more likely to be produced by an AI than a human, as human-generated text is generally less predictable. (Notice the keyword here: more likely.) AI are typically more consistent than humans in the text they generate, and they tend to use certain words more than humans. Of course, this gets more complicated, and you can read more about methods of AI detection here.

So what’s the issue? Why doesn’t this software always work?

If the bolded text up above doesn’t already clue you in, there is a more solid explanation. The answer, in my mind, is primarily twofold. AI text-generators like ChatGPT have learned to create text by being fed human-created text, so the text they generate is fundamentally “human” text even though it’s typically more “robotic sounding.” Also, humans have been taught to write like robots for years, so distinguishing between “human” and “robotic” writing isn’t actually all that easy.

For example, students are taught to follow very formulaic writing structures. SEO writers are expected to follow a slew of robotic writing conventions. Even certain creative writers must keep their creativity (AKA their “more-human” writing tendencies) in check; for example, the genre-fiction mass market rewards easy-to-read/easy-to-binge books that use simple sentences and word choice.

How human writers became robot writers

As discussed above, humans have been taught, encouraged, or required to write in ways that might be deemed “robotic” for years to make their writing more understandable and accessible.

I used to teach college-level first-year writing courses. Many instructors in these classes are expected to teach students genre conventions but in a very formulaic way. (For example, the five-paragraph essay still dominates in first-year composition classes, and it’s a favorite structure of ChatGPT.)

Some aspects of this formula involve:

· A clear thesis, which is usually placed at the end of the introductory paragraph.

· Clear topic sentences at the beginning of each paragraph.

· A conclusion that repeats/reinforces all the main points without introducing new information.

The thinking behind this is that helping students to master the formula will give them the knowledge of conventions they will require in order to defy these conventions later on.

(Also, this must be said: having students write using a formula helps teachers to read and understand student writing so they can give students the fairest grades for their efforts. A lot of student writing is not easy to read. Seasoned writers are better at making their writing understandable while straying from typical writing formulas.)

Student writing isn’t the only writing that’s formulaic. Anyone who’s ever written content that’s optimized for search engines knows the SEO best practices include the use of many writing conventions that AIs like ChatGPT excel at, including:

· Short sentences.

· Active constructions.

· Simple word choice (think: middle-grade level).

· Keywords, keywords, keywords.

SEO copywriters might be asking themselves: should I include longer sentences with passive constructions and complex word choice just so I’m not accused of using AI?

This is one of those situations that is equal parts amusing and upsetting.

What does the future of writing look like?

This “future” of writing, in which humans who have learned to be robots and attempt to unlearn roboticisms, and in which robots learn to be human, has already begun.

Writers are going to continue their attempts to not sound like robots.

For better or worse, many writers’ new writing goal is going to be to sound more human. This could be a good thing. Many writers complain of how writing professionally has made them worse writers, so giving themselves permission to “be more human” might help them return to their former glory.

However, SEO writers are having to find ways to reconcile optimizing their content for search engines and making content that doesn’t seem like it was generated using AI — which can prove a difficult undertaking. But some of the attempts to avoid false detection, like incorporating grammar errors into one’s writing, seem counterintuitive and overall just bad for the world of writing.

Writers are going to find ways to game the system.

Writers are already finding ways to make AI like ChatGPT sound more human. For example, one method of doing this is using strategic prompting. Through cleverly written prompts, writers can coax AI into creating text that sounds less robotic.

This can be tricky. For example, one can’t simply tell ChatGPT to introduce grammar errors into its responses:

But one can tell it to adopt the dialect of a specific person, and such a dialect may contain grammar errors.

The very act of writing is becoming a performance.

Employers might want to do more than just use AI-detection software to prevent their writers from using AI. For example, Originality.AI encourages employers to use a Chrome extension as part of these prevention efforts, which allows employers to “watch writers write” to confirm that writers did not use AI text generation. This is yet another form of workplace surveillance.

I don’t know about other writers, but my writing process is messy, and the thought of anyone peeking at the evolution of one of my drafts fills me with a quiet dread.

Clients and employers who do allow writers to get assistance from AI expect heightened productivity.

There are several clients and employers who have accepted AI as a legitimate writing aid, but many of these might expect their writers’ productivity to increase dramatically with the use of these aids.

In my experience, these aids are certainly helpful, but they don’t increase my productivity dramatically by any means. Sometimes, I find they actually end up slowing me down.

Finally: Are AI-detection software developers engaging in ethical marketing practices?

Personally, I don’t think we can ever count on AI-detection software and that AI-detection software developers are over-exaggerating the accuracy of their services, and, in some cases, the need for their services. I think AI-detection software companies are relying on their potential users’ misunderstanding of how large language models generate text compared to how humans generate text, making it seem like there is a huge difference when there isn’t always, and making it seem like detecting these differences is simple when it’s not.

These detection software companies often give people very surface-level descriptions of how their software works, typically explaining how documents are “scored” (as being AI- or human-generated) rather than how the actual detection works. While this might be because potential customers wouldn’t understand the technical processes, I suspect it might also be because the explanations of those processes could reveal the potential flaws: namely, the absolutely unavoidable false positives and false negatives.

I highly suspect that any internal studies these companies done to prove the accuracy of their services have been very selective. (The study Originality.ai has done is used to make the claim that their AI-detection software is “96% accurate”). There is a huge need for more independent external studies of the accuracies of these softwares.

I also think that AI-detection software companies are unfairly capitalizing on unfounded fears and false beliefs. For example, Originality.AI’s website says:

“As you know Google has reconfirmed that it does not want your AI generated SPAM. It has also confirmed that it is capable of detecting AI generated content. Within the wordpress editor you can check your content in the Originality.AI chrome extension to determine if Google will think it is AI generated.”

This manipulatively rhetorical text uses a straw-man fallacy to make people unfamiliar with Google’s actual AI policy fear that Google will label any AI-generated text as spam. Google’s actual policy (at the time of writing) is that they “[reward] high-quality content, however it is produced.” They admit that “AI has the ability to power new levels of expression and creativity, and to serve as a critical tool to help people create great content for the web.” They do contend that they will continue to actively fight spam, including AI-generated spam, but AI-generated content doesn’t automatically qualify as “spam” simply because it’s AI-generated. (Their definition of spammy AI-generated content is very specific.)

The aforementioned Washington Post article quotes Ian Linkletter, the emerging technology and open-education librarian at the British Columbia Institute of Technology, who says the following about Turnitin and their AI-detection software:

“I am worried they’re marketing [the software] as a precision product, but they’re using dodgy language about how it shouldn’t be used to make decisions. They’re working at an accelerated pace not because there is any desperation to get the product out but because they’re terrified their existing product is becoming obsolete.”

Originality.AI also appears to be marketing their software as a precision product while simultaneously pushing the false idea that Google considers all AI-generated content to be spam. They are instilling fear into web marketers and then capitalizing on that fear. Their business is built to address a problem they are misrepresenting through fear-based tactics, and their product is marketed to address this problem with a level of accuracy that might not reflect reality. (Please note: I have not tested Originality.AI myself, so all my ideas about the accuracy of their AI-detection software are based on hearsay.)

Perhaps they, too, are worried about their product becoming obsolete, especially since there are many companies or other employers who are welcoming AI as a productivity-boosting tool and encouraging or even requiring their writers to collaborate with it. Some teachers, too, recognize the futility of trying to prevent students from using large language models to write, and instead are opting to incorporate these models into their instruction. Perhaps AI will be collectively embraced as a writing tool, and AI-detection softwares truly will become obsolete. Or, as the Washington Post article suggests, perhaps models like ChatGPT will improve so much that it will become impossible to tell the difference between AI- and human-generated text even with the “best” AI-detection software.

For now, AI-detection software seems to be becoming more popular among many different entities.

While writers claiming to have generated text that was actually generated by AI is certainly problematic, attempts to catch these writers might be doing more harm than good. Many will argue that these softwares serve to deter writers from using AI-generated content. Some might view the false positives as collateral damage, necessary for catching all the true cheaters. (Also, there are other potentially good reasons to use AI detection software, including discovering whether misinformation was AI generated as part of a scam or disinformation campaign, for example.)

But what unintended harms will arise from the haphazard use of this software?