How could Google recognize AI content?
Recent advances in machine learning put search engines and the web in front of new challenges.
When the internet, it mostly meant playing online computer games with my friends for me. Back then, I couldn’t grasp that there might someday be companies that aggregate the whole web or my friends and make them easily accessible. Aggregators build an inventory (Google: websites, Facebook: people) and make it searchable and egagable. The internet and its network effects make it possible: the whole world has access to the same network and common needs.
The next stage of the web is being able to create anything with ease. Instead of having to search the web for information to write my newsletter, I’ll be able to give some form of machine learning tech a few inputs and it will write the newsletter for me.
Now, we can do the same with images. In January of 2021, OpenAI released DALL-E, a GPT-based machine-learning technology that can create any image based on text input. Recently, OpenAI demoed what it can do. You want an avocado GPU? DALL-E got you covered.
What DALL-E means for the future of content creation
OpenAI is using the same Transformer technology for DALL-E that Google uses for MUM and GPT-3. As a result, DALL-E understands the implied context of inputs very well. For example, it understands the difference between “wine glasses on a table” and “wine on a glass table” and returns very different images as a result. It also understands the context of languages and time. It's well suited to provide imagery for any application. [x]
GPT-3 can create text based on a few inputs. You tell it to write about “Napoleon's strategy for invading Russia” and it will write a full essay about the 1812 war. DALL-E, a combination of Pixar’s WALL-E and artist Salvador Dali, does the same for text-to-image.
Creative SEOs can imagine a future in which DALL-E creates images for blog articles automatically. You just write the text, tell DALL-E how many images you want and in what style, and it will create them based on the context it finds in your writing. Or it might create stock creative for landing pages. It could even place your logo or product in a totally different context and make it look real. No more design support needed.
Someone might build a service that creates any image you want based on a simple text or voice input to replace or enhance sites like Unsplash, Pexels, or Shutterstock. You wouldn’t have to look for license-free images for your slide deck anymore. Google or Microsoft could provide the service for free right in the software. “Insert image of smiling business people in suits here.”
It seemingly fits that Google and Microsoft are the two biggest investors in AI research. In fact, Microsoft paid billions to get unique access to GPT-3’s source code while everyone else uses an API. The competitive advantage of owning a tech like GPT-3 or DALL-E could be hard to beat. [x]
But it doesn't come for free. Just a single training run for GPT-3 cost OpenAI $12 million and pushed computational cost by 300x. Ongoing training of AI is expensive, not even speaking about the cost of computational power to run it. The bar to entry is high.
As a result, cost for end-users might be low in the beginning since companies want to gain market share quickly. But, just like UBER or Lyft, eventually cost will have to rise. AI might be able to create lots of text and images at scale but cost is a limiting factor. [x, x, x]
Can Google recognize AI-generated content?
Recently, Google’s John Mueller mentioned that Google will likely consider AI-generated content to be spam. If Google’s webspam team finds AI-generated content, it might slap a site with a manual penalty. Mueller made the argument that AI-generated content is akin to content spinning. [x]
That leads us to two fundamental questions. First, how should we use AI to generate content? Second, how can Google detect AI content?
For the first question, it really matters how much AI is used to generate the content. There is 100% AI-generated content and content created with the help of AI, for example. Modern SEO editors and content tools can outline content for you and give suggestions for topics to cover. They’re not yet in a place to fully write the text but might get there in the next few years if the technology keeps progressing at the same pace.
It also depends on the use case. For years publishers from the New York Times to the Wall Street Journal have been using AI for all sorts of things like automated journalism or robot journalism, in which publishers use AI to summarize news. Which human really wants to write that kind of stuff anyways? Give it to machines! The same applies to meta descriptions (which Google likes to ignore) or product descriptions. [x]
The second question is much more complicated. Content from previous versions of NLG (Natural Language Generation) models like GPT or GPT-2 can be identified. But more sophisticated models like GPT-3 are much harder to identify because they mimic human style so well. [x]
The question of how Google can detect AI content is the same question as how to separate fake news from actual news. AI might make the fake news problem even worse, since it can now be created automatically and therefore, at scale. That was the original reason for why OpenAI didn't publish GPT-3 and it does the same with DALL-E.
One version of how this plays out is that Google might treat AI content similar to backlinks. Technically, buying links or trading links is against Google’s guidelines. Practically, Google cannot always understand whether the link was set naturally or part of a trade. Only when patterns become obvious (say, when the backlinks are irrelevant), Google can take action. The same could apply to AI content, where Google won’t see the difference as long the style is consistent, the article has a byline, and it’s of high quality.
Another possible solution is the Blockchain. Google might validate and track human-written content through a ledger, similar to how it used authorship markup, and display a tag in the search results. It might even rank human-written content higher, depending on the use case. The Blockchain is the antidote to fake news. It might very well be the path to trustworthy results.
From aggregation to creation
I like Gedankenexperimente - thought experiments - and one of my favorites is "2nd order consequences". Instead of thinking about consequences, you think about consequences of consequences. When I think about AI, content, and SEO, two questions pop into my head: ”what will differentiate sites when everyone has the same content?” and “will sites even be needed at this stage?”
Google looks at a lot of different factors besides content but has also said that content is one of the strongest ranking signals. If everyone can create text at scale and low cost, Google might have to revert to other signals. In e-commerce, it could be availability or shipping time. In SaaS, it could be features. What else?