Design Fiction: The Rise & Fall of Humanhood

Design Fiction: The Rise & Fall of Humanhood

Report from an imagined future, where LLMs force people to wonder whether a human is behind everything they’re reading…

The creation of the AI Attestation Codes

NY Times headline, February 28, 202X: “Our commitment to human writing.”

As the use of LLMs spread in the mid-2020s, huge rifts began to form in the online landscape. As companies started replacing more of their human writers with AI tools, a growing movement of internet users found this generated content off-putting. Similar to the movement towards “organic” food labeling in the 2010s, prestige newsrooms started adding certifications to their online articles:

This article was written without the use of generative systems.

This movement had various reasons for rejecting generated content. Some of them felt misled by machine-written articles. Some took a stance that they couldn’t trust non-human reasoning. Others felt like it was a harbinger of an AI-takeover of our minds – that if we let AI write for us, it would subtly influence our thoughts in ways we couldn’t control. And some were just still looking to read as a form of human connection.

On social media, people discussed the idea of developing a “Creative Commons License for generated content” – simple, open source tags that any writer could add to their blogs, newsletters, and posts to attest to its humanity. While these were initially bespoke and varied, before long a non-profit formed and launched the AI Attestation (AIA) Codes for a variety of use cases:

AIA-SA 1.0: Strict Attestation. The author of this content certifies that it was written without the use of any generative system.

AIA-GEN-FGC 1.0: Fully Generated Content. The author of this content attests that it was written entirely by a generative system, without human intervention.

AIA-GEN-EGC 1.0: Edited Generated Content. The author of this content attests that it contains content that was mostly written by a generative system, with human editing that may extend to clarity, conciseness, slight to moderate additions, rewrites, and other edits.

AIA-MIX 1.0: Mixed Content. The author of this content attests that it was created from a mix of human-generated and machine-generated writing, which could include sections exclusively written by human or machine, sections written by human and extensively rewritten by machine, or vice versa, or some other unspecified mix of contributions.

AIA-INSP 1.0: External Inspiration. The author of this content attests that sections of it were inspired by writing or other content created by a generative system, but no generated content appears in the published text.

Initially, these were viewed by general audiences as a bit of a curiosity, but were more of a niche among those online writers tapped into the LLM waves. But before long, the use case for AIA became clearer.

The brief heyday of the AIA

NY Times Headline, May 3, 202X: “Hundreds of blogs raved about this new hair straightener. The problem? They were all AIs.”

LLMs were hardly out in the wild for a couple of years before the scammers wrapped their arms around them with glee. For only a few hundred dollars, you could hire a company to spin up hundreds of review blogs, newsletters, and Facebook accounts. They’d have dozens of posts, all seemingly written by a human, talking about their life as a mom, or college student, or Tech Early Adopter, and occasionally some reviews. It just so happened that those reviews were glowing about the same products.

(Not to mention the really adept bot companies would throw in a couple of blogs that left lukewarm reviews for really dumb reasons, assuring that anyone looking for “negative reviews” would still be left with a positive taste in their mouths.)

Just like it became common knowledge that the foreign prince or long-lost relative emailing you about a $2 million inheritance is obviously fake, it soon became common knowledge that most glowing reviews you read online were fake as well.

For a brief moment, the AIA was held up as a remedy, a way to regain that trust. Seeing a blog with an AIA was a signal to its quality – it allowed people to trust what they were reading online as human once again.

The obvious flaw in self-tagging

The AIA was started with good intentions, and so it imagined that its users would similarly be good-intentioned. That is, they would accurately certify their content.

Unfortunately, that was not the case. In fact, as soon as it became apparent that people were trusting the AIA codes, it became heavily incentivized for those with bad-intentions to begin falsely certifying their content as human-generated. The AIA had imagined that an organization’s reputation would keep them honest (the fear of backlash, fallout, negative press, etc.).

But obviously, for those trying to misrepresent fake content as real, they didn’t care about reputation. So sure enough, the fake blogs started adding “AIA-SA 1.0: Strict Attestation” to their posts.

To add insult to injury for the AIA Codes, as LLMs pulled live content off the internet and recycled it in its generated responses, requests to “write a blog post that seems like it was written by a real person” started to automatically include “AIA-SA” as a part of the generated text!

The invention of Humanity Certificates

Humanity Certificate Authority Superbowl Ad, February 12, 202X: “You trust websites with your credit card. But would you trust that they’re humans?”

With the failure of the AIA Codes, a new non-profit rose to try to fill the gap. Though independent in theory, the founding was supported by several of the world’s largest internet companies, who worried that declining trust in online content could spell an existential risk to their business models.

Coming together in a matter of months, as opposed to the several years these types of standards committees usually take, the Humanity Certificate Authority (or HCA) proposed a new type of certificate that a domain could install that certified its content was written entirely, or mostly by humans. Certificates would require submission of personal information to confirm a real human was behind them, and the organization had a complaints procedure where violative certificates could be revoked.

Despite the huge PR push, money behind it, and swift integration into existing browsers, pushback to the HCA was intense and immediate. People from all corners of every political spectrum found it insulting to “prove” to some distant authority that they were humans.

The PHA Protocols

“The basic idea is simple: would you rather be roommates with a total stranger, or the friend of a friend? So why not apply the same principle to humanity on the internet?” – Danny Peet, keynote at Human Visions Conference 202X

In the pushback, several different organizations that came to be known as Peoples’ Humanhood Authorities (or PHAs) rose to prominence. Some of them were non-profits, others were business plays. Some used distributed ledgers on the blockchain, others used private databases. But they were all based around the same principle:

The best way to assess humanhood is by other humans.

(Humanhood was a neologism based on debates of whether “humanity” was an appropriate term for “likeliness that this entity is not a machine.”)

The PHA systems worked through a process of Social Attestation:

• Anyone who wrote content online could “sign” it and validate it against their unique Author ID in a PHA’s database
• The PHA assigned a “Humanhood” score to the writer, and therefore the content, corresponding to the confidence the author was a human
• Authors with IDs in the database could Endorse other writers as also being human – so not a self-attestation that I am a human, but that I attest this other entity is a human.
• The higher an Author’s own Humanhood score, the greater impact it had on the Endorsed Author’s Humanhood – preventing the mass creation of spam bots to artificially inflate scores

PHAs were seen as a big improvement over both AIA Codes, due to the Social Attestation angle, and the HCA, due to its more decentralized implementation.

The PHAs Evolve

But just like before, cracks started showing in the implementation of PHAs, leading the emerging industry through several rounds of modifications.

First, there were the Humanhood Attacks. In the heyday of social media, an army of trolls might try to kick someone off a platform through a deluge of false content reports, triggering some internal system to temporarily suspend an account. This transformed in the age of PHAs to Humanhood Attacks – when someone angered a group of opponents, they could tank their Humanhood score through Negative Endorsements (“I believe this content was written by a machine.”)

Since Humanhood scores had been integrated into scores of products – typically being used as a threshold for aggregators to include or exclude content from their feeds – the scores became a viable attack vector. As soon as the internet realized it was a weapon at their disposal, Humanhood no longer signified “Do other people think I’m a human?” but instead “Do other people think I’m a human, and like my content?”

In response, several PHAs started allowing reports of accounts that were abusing the system, banning their Author IDs and removing their influence from the system. As could be expected, those banned authors went and started up their own PHAs. Before long, choosing your PHA became part of your identity and politics.

The Humanhood Economy

“And, if you liked this article and want to learn more about how to write like me, check out my Writers Workshop link below, which includes the opportunity to get my PHA endorsement at the end of the course.”

Economic incentives followed close on the heels of PHAs. As readers were rejecting content with low PHA scores in droves, that meant advertisers and sponsors started scaling payouts against PHA. If your score was high, you became a Top Author, and could make a decent living just from publishing trusted content. It was a new spin on the Creator Economy, one where trust was at the forefront of a successful career.

In addition, people soon realized their Humanhood scores were a powerful form of currency. While “selling” a PHA Endorsement was usually grounds for suspension in a PHA’s Terms of Service, authors found creative ways around this. Many started selling “Writing Courses” where they’d doll out pre-recorded videos or articles of advice, and the expectation that you could “connect” with the author for a “possible” endorsement at the end of things.

Intermediary organizations formed – ones that charged a fee to “connect” you to various Authors. After a perfunctory conversation, the Author would then rubber-stamp a PHA endorsement, and get a kickback fee for doing so. This launched a tertiary industry: hiring low-cost labor to represent their personhood in these conversations, so that bot farms could apply it to their generated content.

In other words, people soon realized that the authenticity of PHA scores was getting watered down even further. In effect, it entrenched two classes for PHA:

• Seeing content with a moderate PHA score became fairly meaningless. It was assumed that something at this level could be bought or faked, and was disregarded as “essentially as trustworthy as an LLM” by readers, and therefore advertisers.
• The one thing that couldn’t be faked were the astronomical scores of Top Authors. While some newbies were still able to rise through the ranks, the status and importance of this high PHA segment essentially became the only meaningful economic metric.

The fall of Humanhood

NY Times headline, March 3, 202Y: “Updating our commitment to human writing.”

Of course at this point, Humanhood scores had naturally evolved away from their real purpose: distinguishing human-authored content from artificially-generated writing. As it stood, any halfway-decent deceiver could easily game a PHA for a reasonable score, and the PHAs couldn’t respond fast enough. But PHA had taken on so many secondary uses, from telegraphing affiliation to launching careers, that few people treated it seriously as its intended measure.

Where did this leave content on the internet? In some ways, for those seeking out human-created writing, the lines had become more blurred than ever. More writing tools were integrating AI-assisted research tools. Even standard note-taking apps and to-do lists had integrated generative content. Writers at some of the most prestigious publications weren’t going to abandon the best-in-class tools to do their work.

In many ways, the cycle of humanhood ended up going full-circle. The public had to generally trust that publications with reputation to lose wouldn’t misrepresent too much of their content (and the few major news cycles where publications did try to skirt the public’s trust acted as a sufficient fear).

A new internet literacy emerged, but in some ways, it was the same message that’s existed as long as the world wide web itself: there are no gatekeepers here, anyone can post anything they want. Anyone can claim they’re a person, everyone will claim they’re right. Proceed with caution – on the internet, no one knows you’re a bot.