My work these days is in part focused on trying to ensure content is what might be called uncontaminated: that the cold clammy hand of AI is not only anywhere near it, but is seen and believed to be nowhere near it. Here are some general thoughts on the problem and possible solutions.
We humans, as with all God’s creatures on this planet, are designed for danger. We are primed to quickly evaluate something, and to rely on our senses. If you’ve ever been caught in a riot or protest, where feelings flow through a crowd, you’ll know that sense of self-preservation before you see or hear anything.
I’m not comparing getting caught in a riot to reading AI slop, but I sort of am. I believe we’re reaching a point where our guard is so far up it may make it very hard to convince readers, viewers or listeners that what they’re reading, seeing or hearing is authentic — in other words, human.
The uncanny valley
Our growing awareness of what AI is capable of is largely down to sharpened senses — the uncanny valley where we can’t accept something that looks almost human — to figure out what is and isn’t. The term was originally developed for robots, but it can be applied to content, and the response is more visceral: we don’t want to be made a fool of by fake content, and so we shoot first and ask questions later.
A study by researchers at the Hebrew University of Jerusalem and Harvard Business School found that if participants merely suspected AI-written replies expressing support to their descriptions of real emotional situations, they felt them to be less empathic than those they were told had been written by a human. (Even if they were told a human had only been helped by AI participants felt reduced empathy.)
The yearning for authenticity
This is logical. When we consume content we are entering into a relationship with the (claimed) creator, a relationship that relies on trust.
It’s not unlike counselling: the client has good reason to demand that their therapist is not phoning it in or parrotting stock phrases. The entire premise of counselling rests on both sides being authentic.
It, and content, fails if the user has reason to lose trust, because the thing being consumed — text, audio, video, or a mix of the three — rests on a single pillar: the creator’s credibility. If that pillar gives way, or cracks appear, the consumer will likely bolt and never come back.
The problem is that the credibility of the content is falling in inverse proportion to the quantity of AI-assisted content out there. The more AI we see, the harder we find it to trust anything.
The reality is that the table has already flipped: that the user, consciously or unconsciously, is looking not for evidence of AI but for evidence of humanity, evidence that they are having an authentic interaction with a human via their words, thoughts, vision.
If they don’t see it, they walk.
Certain phrases and constructions, for example, are red flags, which may in themselves be useful devices. But because some LLMs use them consistently, we notice. And suddenly that construction becomes a no-no. Suspicion has crept in. And with so much other stuff out there to read/watch/listen to, why waste time on something that may be AI?
As adults we’re already developing strategies to try to keep it a bay. The most obvious one is to ignore anything that you feel is written or created by AI. AI; didn’t read (ai;dr) is already a thing.
This isn’t just sense-driven. Tools designed to catch AI-created content (themselves based on AI) can pass judgement and appear robustly authoritative.
An AI checker reckoned a chunk of text of an Atlantic front page story was 100% AI created, prompting some strong pushback. But the line is already blurred: could some AI text have found its way in?
That’s because we’re at a point now where many of us are resentful of AI, and feel its cold vice-like grip on our lives, our jobs, our hopes, to the point where we don’t want to have anything to do with it. It’s telling that anecdotal evidence suggests the younger generation — usually the first to embrace a technology — do not recognise AI as a positive influence on their life. We have not just poisoned the water hole, we’ve screwed up the whole water table.
The content problem
This raises a troubling question: how can we create anything if the moment someone feels there’s AI involved the entire platform/oeuvre is tainted and dismissed? How can we recover from that?
Some solutions might include the following:
- Don’t use any AI at all. Ban it from the premises.
- Use AI but keep any AI usage air-gapped from whatever device you’re using to create content. No moving stuff across, limit AI to backoffice tasks
- Use AI but explain how it was used and what it was and wasn’t involved in.
- Use AI but run it through an AI-checking service.
- Use AI but look over it first to see whether it passes the smell test: does it look like AI to you?
I’ve seen all these approaches proposed, adopted or recommended. And I’m here to say none of these is watertight. In fact, the opposite: they give the comforting appearance of a process and protocol without acknowledging that you need to not only proclaim you’re not using AI, but your content needs to survive rigorous testing by humans.
You can build up a generation of trust and credibility, but as all writers know, you only need to screw up one story, one perception of cutting corners with a bit of AI help, for all your content to fall under suspicion.
Content as craft
This is a terrifying concept, or should be, for any individual or institution for whom content is an expression of their value.
- A Substack writer leaning on AI to write or conceptualise is obviously at some point going to end up in deep trouble.
- But so is a company using AI for messaging, be it white papers, press releases, bio, company values and product pages.
- What might look like saving money or ‘being smart’ may be endangering the brand in ways that might not be recoverable.
The answer? I would say this because I’m a former journalist, but I think those who write for a living (and have been writing before the birth of LLMs) are probably going to be the vital cog, just as developers who were let go as companies saved money by having AI ‘vibe-code’ their software are finding they’re being rehired or contracted to go over problematic AI-generated code.
There are some things that cannot be taught, nor can they done by AI. Using a particular sentence construction more than once in the same piece, say, would sound contrived and unsightly to a good writer, something an AI wouldn’t necessarily see.
More next time.