Yesterday, a friend and former colleague shared a heartfelt Facebook post about the loss of his cousin. It was personal, emotional and accompanied by a real photograph.
At the top of the post were two words, “AI content.”
I had no idea why. Had the photograph been generated or edited? Had an AI-enabled tool processed the file? Was Facebook suggesting the words were written by AI? Was I supposed to question the photograph, the message or both?
I could make a hundred assumptions. The label gave me no meaningful basis for making any of them.
Based on Meta’s published approach, the label appears to relate to the photograph or another piece of media. Meta labels images, video and audio when it detects industry-standard AI signals or someone discloses that AI was used. I could find nothing to suggest this label is triggered by analysis of the accompanying words.
Facebook does not make any of this clear. It places “AI content” above the entire post, leaving readers to decide whether they should question the image, the words or the person who shared them.
Since then, I have seen the same label on another Facebook post containing five family photographs from a holiday in Mykonos. The post read:
“Happy birthday to my sister (name) and son (name) -have the best day celebrating yr special and lots more to celebrate this weekend -this time last yr were were celebrating in mykonos and dont we all remember that lol love you all 🎂🍰”
It reads like a hurried, human birthday message. Yet “AI content” sat above the post.
Perhaps one photograph had been touched up or passed through an AI-enabled tool. We cannot tell. Facebook does not identify which image carried the signal, what happened to it or whether the other four were untouched. One signal may have cast doubt over five real family photographs and the entire post.
What is actually being labelled?
I have since clicked on the label. Facebook says: “Content in this post may have been created with AI.” It explains that the label is based on a signal added by the tool used to create the content and that this signal “typically identifies content that was entirely AI-generated.”
“May have been” tells me Meta is uncertain. “Typically” tells me the signal is not definitive. “Content in this post” does not tell me what it’s referring to.
Meta knows that a tool involved in creating or processing something within the post added a technical signal. It has not independently established what was generated, how AI was used or how much was created by a person.
Meta also says that not all AI content contains the data it needs. Entirely generated content may pass through Facebook without a label while authentic content attracts suspicion because a photograph carries a signal from a tool that processed it.
Meta has acknowledged that its labels have not always matched what people understood them to mean. Photographs with minor AI-assisted modifications, including retouching, were once labelled “Made with AI”. Meta later changed the wording to “AI info” and moved information about modified content into the post’s menu.
The labels I saw were more prominent and, according to Meta, triggered by signals that usually indicate entirely generated content.
But usually is not always. May is not was. And content is not specific enough to tell me what I’m being asked to question.
This is an automated expression of suspicion.
“AI content” is not a useful category
The use of AI exists across an enormous spectrum.
Someone might use it to check spelling, suggest a clearer sentence, organise their thoughts or test how an audience could interpret something. They might ask it to substantially rewrite their work or give it a two-line prompt and publish the response without reading it carefully.
They might also use it to create a convincing photograph of an event that never happened, imitate someone’s voice or invent a statement and attribute it to a real person.
These activities do not carry the same ethical, professional or reputational implications.
Placing all of them under the label “AI content” is like placing calculator-assisted mathematics and fabricated financial accounts in the same category because both involved a machine.
I support disclosure. The communication profession has an important role in promoting responsible, transparent and ethical AI use. Through the Centre for Strategic Communication Excellence, we have published responsible AI guidelines and researched the profession’s use of AI for the Global Alliance through our Reimagining Tomorrow study. We have conducted that research for two consecutive years and will continue to do so.
But disclosure needs to communicate something useful. A label that increases uncertainty and invites suspicion has failed to do that.
That effect is not hypothetical. Recent research found that AI labels reduced emotional and behavioural engagement, particularly for emotional posts. Another study involving 760 people found that ambiguous labels increased information avoidance. Uncertainty caused people to disengage.
That matters when the label appears on a message about grief or a family birthday.
LinkedIn has created a different problem
LinkedIn recently introduced a reporting option called “Seems like AI slop”. Any user can flag someone else’s post as low quality or overly reliant on AI. LinkedIn says this feedback will help improve its systems and has replaced its “enhance your post” tool with a more limited proofreading function intended to preserve the author’s voice. LinkedIn’s Chief Product Officer announced the changes here.
I understand the frustration. LinkedIn is saturated with formulaic posts, generic comments and content containing plenty of words but very little thought.
But “slop” describes quality. It doesn’t tell us how something was created.
People produced shallow, repetitive and self-important content before AI. Experienced professionals also use AI to test original ideas, strengthen arguments or improve their own writing.
The button asks users to assess whether a post seems like AI. That is a perception, not evidence.
What prevents a competitor, a disgruntled former employee or someone who dislikes your opinion from flagging your entirely human-written post? Could reports affect its reach? Is the author told? Can the decision be challenged? How does LinkedIn distinguish coordinated misuse from genuine feedback?
Perhaps LinkedIn has good answers. Its users deserve to know what they are.
Watermarking doesn’t answer the authorship question
Anthropic has announced that future Claude models will watermark generated text to comply with the European Union’s AI requirements.
The watermark is a detectable statistical pattern in some of Claude’s word choices. Someone with the relevant key can assess the likelihood that Claude was involved. Anthropic says it will offer a detection API.
Anthropic is refreshingly clear about the limitations. The watermark cannot confirm that a person did not write the text or distinguish between something Claude wrote and something it heavily edited. It works less reliably on short passages and says nothing about ownership or authorship.
For supported image and file types, Claude will use Content Credentials, an industry standard that records in the file’s metadata that Claude created or processed it. Anthropic has announced to help members trace the origin and editing history of AI-created media.
These technologies may establish provenance. They cannot explain the human contribution, why AI was used or whether the final work represents the person’s genuine experience and thinking.
What should communication professionals be asking?
We don’t need to become technical experts in watermarking, metadata and detection systems. But we do need to understand the communication consequences.
When our organisations introduce AI disclosure requirements, we should ask:
- What specific element has been labelled and, in a multi-image post, which file triggered it?
- Does the signal establish AI generation or only that a particular tool handled the file?
- How confident is the platform’s interpretation of that signal?
- Can the author see the same explanation as the reader?
- Can the author correct or challenge an inaccurate label?
- What happens when authentic content is labelled while undetectable AI content passes through without one?
We should also expect more from the platforms making these choices.
Platforms should explain what has been labelled, whether the label refers to text or media and whether the content was generated or lightly modified. Users should know why a label was applied, what evidence supports it and how they can challenge an error.
If user reports help determine what content is seen, platforms should disclose how that feedback is used and how they prevent abuse.
The context matters too. A label attached to a political deepfake serves a very different purpose from one attached to a real family photograph that may have been retouched.
We need more than a label
I don’t know what triggered the label on my friend’s Facebook post. I don’t need to know. His grief was real, his cousin was real and the photograph was real.
Yet those two words introduced doubt into a moment that had nothing to do with deception.
This is what concerns me.
The technology is developing quickly. Platforms are racing to demonstrate that they are responding. Governments are introducing new requirements. AI companies are building technical markers into their products. And users are being enlisted to identify content that feels inauthentic.
Meanwhile, we haven’t established a common, meaningful understanding of what “AI content” actually means.
Communication professionals should be part of that conversation. We understand that words do more than categorise. They create impressions, influence behaviour and affect credibility.
“AI info”, “AI content” and “AI slop” are not neutral labels. They carry implications about originality, effort, honesty and trust.
If a platform applies those words to someone’s work, it has a responsibility to explain precisely what it means.
A label should reduce uncertainty. This one creates it, attaches it to someone’s credibility and leaves both the author and the reader to work out what Meta might mean.
