Who is reading AI-generated text?

GPT-3 and similar AI language generation models are fascinating. The use case from the writer’s perspective is clear — write more, faster, and with less effort. My question is — who is reading AI-generated text? And why?

I went through the list of GPT-3 applications compiled by Alex Schmitt. In my mind, there are ten groups of AI-generated text readers, roughly ordered below by increasing commercial opportunity.

No one

Shakespeare-style poems, love letters from a toaster, or AI-generated essays are equivalent to “Hello world!” program in any new programming language. It is easy and fun to do, giving exciting and unexpected results, but has no value whatsoever. Still very cool nevertheless — see the wisdom tweets by @ByGpt3.

Fantasy fans

One of the issues with the AI-generated text is that it can get too creative and unrealistic. But what if too creative is good? Excellent examples of AI language model applications in the fantasy text games: Dungeon AI or LitRPG Adventures. I haven’t seen it yet, but children fairy-tales should work too, assuming someone checks that they do not unexpectedly go too dark.

Struggling writers

Coming up with plausible but unexpected creative ideas is challenging. Maybe you don’t want AI to write the whole essay or a blog, but only to give you a head start. I can see how an ad agency could ask for 20 inspirational ad copy ideas for a bike lock. “You don’t lock your bike because it’s a bike. You lock it because it means something to you”. Fair enough.

No one, but still must be written

Not to diminish the importance, but some texts have lower readership than others, while they still need to be written. AI text generator can be a massive help here. Product return policies, “Risk Factors” sections in the financial filings or legal fine print are a few examples that come to my mind.

Knowledge seekers

Information crunching AI is a lower novelty use case since knowledge reciting and Jeopardy playing AI has been around for a while. Nevertheless, natural language capabilities can create even more interactive and engaging ways to consume Wikipedia.

Unsuspecting recipients

This use case applies to emails, CVs, and similar. It helps craft elaborate sentences and paragraphs, which historically indicated that an author put a lot of effort into the work and, therefore, respects and cares about the recipient. It is a natural extension to what Gmail or Grammarly are already doing and will remain important in the same way as photo filters are essential for selfies.

We will need to revisit the whole approach as soon as AI-generated emails start getting AI-generated replies, making the entire exercise rather pointless. Then we can switch to JSON instead and move on.

Google indexing bots

If you are using AI text generators to get the full articles rather than just ideas, the ultimate reader is likely Google indexing bot. Automatically generated content marketing listicles are not for reading; they are for traffic generation.

Eventually, I suspect this will transform into the same arms race as with video deep fakes. Some AI models will be auto generating blog articles, and others — filtering them out.

Advertising audiences

A/B testing is a powerful tool but generating thousands of ad copies is a time-consuming endeavour. Ability to create multiple versions of an ad or a landing page to see what works best is an undeniable use case. Unfortunately, GPT-3 cannot be fine-tuned on the results and cannot know what worked and what did not. Once it becomes an option, I am curious to see what kind of click-baits GPT-3 will generate.

Support seeking customers

We all been there — trying to navigate complex help documentation, getting frustrated by poorly written support chatbots, and then calling help centre just to be told by an automated voice to go back to the website. Customer support service is often still frustratingly bad, therefore using GPT-3 can only be an improvement.

Other computers

This use case makes the most sense to me, assuming it can work well enough. The AI model takes natural language and translates it into a computer code (SQL, JS, or any other). Most importantly, differently from humans receiving an AI-generated love letter, computers would not mind. Even more so, such capabilities would play nicely into the current low-code / no-code trend.

My bet is on the language models writing computer code and functioning as customer support bots. The rest will likely appear in Microsoft Office products as neatly integrated features, which explains $1bn Microsoft invested in OpenAI.

And finally, an AI model writing AI code accidentally left on loop, is exactly how machines can take over the world. Let’s all agree not to do that ;)