Any trends or examples of how to distinguish between AI and human generated content?

Question

With a lot of AI generated content populating the Internet and its growing presence, it seems socially responsible for individuals, businesses and organizations to provide some disclaimer about the source of the content. This will ensure that there is accountability and authenticity in what is presented to the audience.

However, since the use and uptake of the technology has taken place at such a rapid pace, there seems to be no common or established practice in place.

Has anyone seen good examples of this in bigger organizations, or even smaller but established sites? This could be labels or icons that show the use of AI technology in the generation of content, % estimate of AI generated content, or even that something is 100% human generated (and what that means).

Answer 1

What you are talking about is, of course, a pressing issue. And it is widely discussed and there are some actual efforts for the solutions.

1. Guidelines and Best Practice

There are different guidelines and papers about the best ways to disclose AI-generated content:

• Synthetic media'a "Responsible Practices for Synthetic Media" framework. It seems to be backed by a lot of big players.

• MIT's brief "Labeling AI-Generated Content: Promises, Perils, and Future Directions"

• Research paper "What label should be applied to content produced by generative AI?"

2. Content Credentials

There is this organization called "Coalition for Content Provenance and Authenticity" that is a part of Content Authenticity Initiative charged with basically what you are talking about:

[Creating] an open technical standard providing publishers, creators, and consumers the ability to trace the origin of different types of media. – https://c2pa.org

This standard defines ways to denote the origins of content that can be trusted and different ways they can be used to make it obvious for the user. Of course, it's not just for AI, it's more global and it's a way for them to fight misinformation.

It includes metadata formats and different implementations, for example, Content Credentials icons, that would basically provide what they call "digital nutrition label" with the information of the image's origins. It seems to be embraced by Adobe, Leica, Nikon, Microsoft's Bing Image Creator, Publicis Groupe, etc.

It's pretty cool. If you're text-savvy, there's a lot of interesting information in the technical specification itself about how they ensure it can't be faked, encryption, trust, etc.

It also doesn't just work with images - they propose solutions that can be used for "all major content types".

I have only seen it used with images, but I'm pretty sure it'll be more and more popular with other data formats too.

3. Not by AI

Not sure, how popular, but there's this initiative called "Not by AI" which defines these badges that should be used when you want to disclose that your content isn't generated by AI. Which is more human, less technical approach.

If You See the 'Not By AI' Badge, You See Human-Created Content.

They have guidelines and different versions for different content types.

4. Icons and badges

It's also very interesting how different solution use iconography to denote to the user in the visual way that AI has been used in creating content:

• Deviantart example mentioned in the other answer;

• Adobe recently added their Firefly to generate images as a part of their Adobe Stock image library and both display a badge and Content Credentials to those images.

• Newly introduced Samsung Galaxy AI places this icon as a watermark on their AI retouched photos (Which, ironically enough, can be removed with the same AI tools).

• Microsoft Bing Image Creator uses Content Credentials mentioned above and also adds a watermark to the image, using Image Creator logo:

  • 
  • 

• TikTok is experimenting with AI-generated content disclosure by showing appropriate labels:

• Google Pixel phones introduced Magic Editor for their photos, they use this icon (which is very similar to what Samsung uses now) to denote AI enhancement ("Magic").

Answer 2

The responsible thing indeed would be to place a marker onto the content. In the case of DeviantArt, there's a barely visible text placed underneath the description together with EXIF-data - which, in case you use their own imagegen, is paired with a ad for their own service:

In the case of dA it's quite interesting - the site is absolutely unrecognizable to the place of teens sharing art of questionable quality and subject it used to be known as. It now acts as a imagegen subscription with an attached sharing platform. They announced a policy of requiring AI tags on AI generated content and of course also made an AI to detect AI.

Given that they can monetize AI, deviantart probably has a quite decent incentive to mark things to the best of their ability. Most other websites lack this incentive.

I also have seen several articles on help centers and knowledge bases of the likes of Google and Microsoft with a banner saying something to the effect of "this material has been machine translated", though of course in recent months it's turned to "AI translated". Here's an example from Citrix which liked the idea so much they did it twice (yellow and blue banner) and a disclaimer paragraph ("Haftungsausschluss") at the bottom:

In practice, this banner is understood as "user beware, this page is probably somewhat low-quality, go check the original if you can speak English". This is fine for a Microsoft KB which has an interest in keeping trust with its business users, but absolute poison for the kind of SEO spam that Google and Bing are struggling with lately.[1] I therefore suspect that, short of some quite heavy regulation, we won't be seeing any sort of unified icon for it or similar server-side.

% estimate of AI generated content, or even that something is 100% human generated (and what that means).

I think there's a massive difference from the likes of "I used grammarly to improve my sentence structure" and "I used an imagegen instead of a stock image for the cover image" to "hey chatgpt write me a review about X following all the SEO rules there are".

A "how much of this content is AI-generated" appears similarly hopeless and useless as a score that measures how much of a website is UX designed: It's a bit of a weird metric to begin with and doesn't capture what I actually care about. In the case of SEO and AI, I'd want to know if the article is genuine and/or truthful, in the case of UX, I'd want to know if the UX is actually any good or if it's full of dark patterns to annoy me into doing something.