On April’s Fool last year, Google Search Analyst John Mueller had this to say when asked about his thoughts on websites hosting AI written content:
"For us these would, essentially, still fall into the category of automatically generated content, which is something we’ve had in the Webmaster Guidelines since almost the beginning.
And people have been automatically generating content in lots of different ways. And for us, if you’re using machine learning tools to generate your content, it’s essentially the same as if you’re just shuffling words around, or looking up synonyms, or doing the translation tricks that people used to do. Those kind of things.
My suspicion is maybe the quality of content is a little bit better than the really old school tools, but for us it’s still automatically generated content, and that means for us it’s still against the Webmaster Guidelines. So we would consider that to be spam."
As innocuous as this sounds on what is essentially a day for pranks, things quickly spiralled when a blog post categorically concluded that Google will impose penalties on all AI written content, no matter their quality or fluency. As a cooldown measure, several SEO companies and AI copywriting businesses have since jumped out to claim otherwise, insisting that their AI writers can bypass the Google algorithm’s discerning eyes.
Today, marketers and writers alike are still understandably confused. Are humans allowed to co-write articles with AI? Will AI written content be penalised by Google’s algorithm?
Let’s find out.
What did John Mueller actually say?
Despite not having said much, it didn’t stop anyone from misinterpreting John Mueller’s remarks that day. Let’s take a closer look at some of his statements and address them one by one.
Statement 1
"For us these would, essentially, still fall into the category of automatically generated content, which is something we’ve had in the Webmaster Guidelines since almost the beginning."
Here, John Mueller restates Google’s stance on auto-generated content, which is explicitly listed in their Webmaster Guidelines under “spammy automatically-generated content”. Such content includes “text generated through automated processes without regard for quality or user experience”, which at a cursory glance appears to be a condemnation of AI writing and paraphrasing tools.
Yet, people forget that Google has had these guidelines laid out “since almost the beginning”. The Google algorithm had long learned to regard low quality, nonsensical pages and articles as spam in tandem with the development of black hat SEO techniques (the Panda and Penguin updates way back in 2011 and 2012). Similar low quality content over the years were flagged and penalised as such, no matter if they were written by a human or not. As such, it’s clear that so long as the Google algorithm notices that a certain piece is unhelpful to users, they will be treated as spam and ranked lower or entirely removed.
Statement 2
"And for us, if you’re using machine learning tools to generate your content, it’s essentially the same as if you’re just shuffling words around, or looking up synonyms, or doing the translation tricks that people used to do. Those kind of things."
While first generation AI tools could only manage simple translation and word shuffling, current generative AI tools possess vastly superior capabilities as compared to their predecessors. A quick search on Google will return countless reports of GPT-4 (an advanced machine learning model capable of mimicking natural human language) tricking humans into believing that they were reading human-written text or conversing with an actual person.
The best AI writers deployed commercially today produce written content that is almost indistinguishable from human written work. To categorically decide if a piece of content you read was written by an AI or human would be a huge challenge—but not entirely impossible.
Regardless, Google’s interest lies not in drawing the line between human and AI. Their business banks on a separate but specific object: you.
Google’s interest: you
The August “Helpful Content” algorithm update said as much—Google’s only interest is in you.
Google’s main business thrives in the attention economy. Of the US$256.74 billion in revenue the company reaped in 2021, a staggering US$148.95 billion came from search ads. This includes ads on their search engine results page and other products like Gmail, Google Maps, and Google Play.
Now, ask yourself: what would happen to Google if they allowed spam content to overrun the search engine? Or if an article written two decades ago was returned to you as the top search?
If the above situations occur, you would most likely stop using Google Search. Worse still, you might convince all your friends and colleagues that Google Search is awful, and suggest them to take their searches elsewhere.
Taking a step back, asking questions about how Google’s search engine works offers us indirect clues into their motivation. In the attention business, users’ interest are often inextricably tied to the company’s interest. Thus, the question that Google must answer is this: what is in users’ best interest, and therefore indirectly, in Google’s best interest?
Everybody wants quality content
Indeed, everybody appreciates quality content. And because users like us do, so does Google.
As I browsed through various pages of Google’s Webmaster Guidelines, it seems clear since the start that Google’s intent had always been on ranking helpful, quality content for the benefit of users. Recent changes to the Google Search algorithm also signals their steadfast commitment towards promoting “helpful, reliable people-first content” by focusing primarily on user-based metrics. This means that whatever works for the benefit of users will be prioritised and ranked according to how useful the content actually is, and not whether it’s AI-generated. The most recent September 2023 updates indicate a shift in Google’s attitude towards AI-generated content, making it more consistent with other advice on creating helpful material.
The original guidance: “Google Search’s helpful content system generates a signal used by our automated ranking systems to better ensure people see original, helpful content written by people, for people, in search results.”
The updated guidance: “Google Search’s helpful content system generates a signal used by our automated ranking systems to better ensure people see original, helpful content created for people in search results.”
An earlier comment made by John Mueller in 2021 also suggested that Google has long anticipated the shift towards human-AI co-authorship in the near future:
…my feeling is at some point that is going to shift a little bit in the sense that we'll focus more on the quality rather than how it was generated.
So some mix of maybe automatically generated content and human curated content I imagine will become normal.
The takeaway here is simple: Google has no intention of cracking down on all AI-generated content across the world wide web. Removing non-spammy generated content simply because of its artificial origins is neither in the interest of Google, nor is it helpful in any way to users who could have benefitted from the high-quality content. Their December 2022 algorithm roll-out further clarified high-quality content as “content which demonstrates experience, expertise, authoritativeness, and trustworthiness”, neatly captured by the playful acronym E-E-A-T.
So, the verdict’s in: Google doesn’t care, nor does it intend to care in the future. But wait, even if Google did care eventually, does it have the capacity to do so?
Can Google distinguish between AI and human content?
It has been long speculated that the hallowed Google Search algorithm—using over 200 ranking factors—is potentially worth US$180 billion dollars. Surely then, if the rumours are true, such a sophisticated algorithm would easily differentiate between AI and human content, right?
Well, only partially.
As we have repeatedly established, the detection of low-quality, AI-generated content is easy for anyone. Currently, Google’s built-in spam filters are smart enough to block 99.9% of spam, phishing, and malware from reaching your inboxes. By employing machine learning models that learn from user feedback to identify common patterns, these AI-driven filters can be said to be experts at their jobs—much more efficient than manual human filtering.
Nevertheless, as AI spam detection tools get smarter everyday, so is every other machine learning model out there. This is especially the case for generative AI models, an elite subset of unsupervised (or semi-supervised) machine learning algorithms that allow supercomputers to study existing content to create new content from scratch.
In the case of text generation, this means human-like text generation unlike anything you’ve seen. In fact, you’ve probably already come across AI-generated text without noticing. According to experts at the Copenhagen Institute for Future Studies, up to 99.9% of the Internet’s content will be AI-generated by 2030, particularly if generative AI models achieve wider adoption. When that happens, differentiating between automatically generated content and human written content would be practically impossible—and entirely meaningless.
So, Google can’t do it. But someone else can.
Is it possible for anyone to detect AI-generated content?
The short answer is yes, but it might not matter. Allow me to explain.
In the last decade, training machine learning models has become increasingly expensive—we’re talking state-of-the-art data infrastructure in the tens of millions. This means that any business not at the scale of Google, Meta, or OpenAI can forget about building an in-house model good enough to compete with the giants. As a result, using off-the-shelf models like GPT has become the norm in the industry.
Herein lies the trick: anyone in possession of these off-the-shelf models can find common patterns in the output produced and thus detect if a certain chunk of text was AI-generated or not. To draw an analogy, this is akin to watching a magic show while being a magician yourself. You already know the trick, so you can’t be tricked.
But there’s a caveat. If you learned the same trick and improved on it using your own private techniques, the other magicians won’t know what to expect. Over time, if you put enough of your own spin to it, you can be that one magician who always has good tricks up their sleeves.
This is the same thing that’s being done at some of the best AI companies now. By applying in-house proprietary machine learning techniques, startups like Hypotenuse AI add their own flair to the output of the machine learning model while retaining the human-like fluency and uniqueness of the foundational model.
To summarise, off-the-shelf models can be detected by insiders who know what to look out for. This includes AI text detection tools like GPTZero or OpenAI’s own text classifier that identifies patterns in text to tell if something has been AI-generated or not, though the detection is not always 100% accurate.
Choose models that employ proprietary machine learning techniques to avoid getting caught!
Just give it to me—will I get penalized by Google for using AI content?
As we’ve extensively discussed, it really depends on the model you’re employing.
If you’re using off-the-shelf models, there’s a high chance that the writing quality may not be good enough for Google to be rated as helpful. By using these base models, you run the risk of your content being penalized by spam detection filters like Google’s Spambrain System.
On top of their own research into AI chatbots, Google’s stance on AI content is aptly captured in their February developers blog:
“…it's important to recognize that not all use of automation, including AI generation, is spam. Automation has long been used to generate helpful content, such as sports scores, weather forecasts, and transcripts. 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.”
To avoid generating spam or low-quality content, it’s important to do your own research on different AI content generators and find one with special “tricks” up their sleeve: Hypotenuse AI applies proprietary techniques to generate high-quality original content backed up by factual information sourced directly from Google.
Oh well, talk is cheap—try it out for yourself sometime!
