AI Watermarking

If you have travelled anywhere in Australia over the last few months, one song has dominated radios, shopping malls and events - a cover of Madonna’s 1989 hit ‘Like a Prayer’, by Australian DJ Josh Fawaz. It is the biggest song of the season there and yet it appears nowhere in the Australian music charts. The Australian Recording Industry Association announced that tracks wholly or largely generated by AI, like Fawaz’s song, would be barred to protect the livelihoods of Australian artists. They now require any song to be substantially human-made to be eligible to appear in the charts. This is a small indication of a new era of AI backlash that is starting.

 

As of 2nd August, all AI output now comes with a ‘watermark’. It was a requirement of the EU AI Act and came into effect over the last year. Thanks to the watermark, AI content will have nowhere to hide. The argument is not for an outright ban on AI content, but for transparent honesty when an artefact has little or no human input. Thanks to the watermark, tracking technologies will be able to uncover any deceitful practice. That is important news for anyone working in a public facing role and offers new options to university faculty who have struggled to effectively uphold academic integrity. 

 

A little background

Watermarking has a controversial history. It was originally proposed by the ethical AI research community as a solution to protect artists and human ingenuity and skill from widespread theft of their knowledge. The initial proposal was to add a watermark to artefacts at the input stage. Works of art and the collective knowledge and experience of billions of humans were being scraped and stolen at alarming rates. With a watermark, it would be possible to trace whose genius had been stolen and set up a system that restored some respect for copyright law. 

 

We often assume that the output of an AI must be a random collection of words gleaned from thousands of different sources before being pulled together into a completely new artefact. We now know that sometimes the machines often take shortcuts and have been caught plagiarising one or two relevant articles. The end result is the same - a highly useful end product. 

 

Unfortunately, the suggestion to watermark gave tech companies another idea. If they watermarked the output, instead of the input, they could protect their own copyrighted items, while ripping off everyone else. Many used watermarking voluntarily on any output that may have financial value in future. 

 

What changed on 2nd August, 2026?

The EU AI Act was drawn up by those same AI ethics advisors who were fighting to protect artistic talent and expertise around 10 years ago. The argument for respect for copyright appears to have largely been lost at this point. But there is another way to fight back - label AI generated content, ban it in some places and uses and find a technical solution to name and shame anyone who passes off AI content as their own work. 

 

Under article 50, which came into effect last month, providers of generative AI systems must ensure that synthetic or manipulated outputs, including text, audio, images and video are clearly detectable through machine-readable markings and technical watermarks. At the same time, they have been working with teams of academics to pilot tools that can trawl through and locate content, such as deep fake social media postings of a political nature, to allow the law to be enforceable. Anyone caught sharing fake content without clearly labelling it as AI generated in some use cases of public interest, will be breaking the law.  

 

Not every single byte of hidden AI generation globally will instantly carry watermarks overnight, though any major providers operating in the EU are bound by it.

 

What is a digital watermark? 

Watermarking will put an end to anyone passing off AI generated content as their own. A watermark is not an easy thing to remove. As the AI generates text, for example, word by word, a pseudo-random algorithm uses a secret key to split vocabulary up. Technically, it has been used to steer output in the past. The key encodes words from two different lists - green for vocabulary that is encouraged, and red for vocabulary that is discouraged, such as offensive words or naming real people. The frequency of green words is far higher than would occur if the text was written by a human. 

 

That same process is now being used to leave behind a key that makes imperceptible changes to pixel colours in images, or frequencies in waveforms of audio. Because the mark lives inside the data, copying or sharing the file will keep the watermark intact. None of this can be seen by the average user and will be almost impossible to remove, given that it is encoded at the pixel level. Combined with other layers of identification, such as file metadata with provenance tags and more human-visible watermarks, we will soon have an AI ecosystem where AI use can no longer hide in the dark. 

 

What uses of AI are covered by the new laws? 

AI is everywhere these days and the act does not cover all uses. Some AI tools are excluded from the new watermarking requirements. Spellcheckers, assistive screen reading tools and standard editing tools will be exempt. 

 

Key targets for identification, intervention and litigation include:

 

Political content - deepfake videos and social media posts with content of a political nature.

Health advice - any social media used to provide guidance on supplements, medicines and other health benefits claims.

Consumer safety - guidelines on food traceability, known hazards.

Environmental or climate claims - carbon reporting, material use and environmental impact.

 

Will this change the AI industry and social media?

Many years ago, I spoke with one of the first members of staff employed by social media companies to identify political misinformation on social media. State sponsored fake content, designed to destabilise countries, was rife. From 2016 onwards, we had shocking and incontrovertible evidence of the scale of interference. I expected not to like this person, as the social media companies seemed complicit. They may well still be.

 

The individual I met was earnest and wanted to do the right thing. He made the very valid point that ‘fake political content’ is not a black and white issue. How do you draw the line between fake content and an individual’s right to express their views? In the end, they gave up trying to answer that philosophical question and started asking advertising sponsors to give a physical address. It worked, but not for long. 

 

This updated EU legislation will make synthetic, AI-generated content far easier to identify. Fictional online personas, like the white American nurse, Jessica Foster, can be fronts for fictional personas, operated overseas to generate engagement and ad revenue under false pretenses, or to achieve political objectives. They will face much greater scrutiny, potentially limiting their viability.

While public reactions to synthetic media often vary, ranging from concern over deceptive identity spoofing to appreciation for the technical quality of satire or parody, mandated disclosure ensures that audiences can clearly distinguish between human-created and machine-generated content regardless of its source.

Social media companies may have larger issues to focus on right now, though. Meta agreed last week to a deal around their handling of content for young users. They have agreed to pay record compensation and change their rules for displaying content, under intense scrutiny in the US. Backlash against data centres is a growing political issue and there are many other large and complex challenges in the industry. With so many issues vying for their attention, accepting paid sponsorship of fake content may well drift down the list of concerns for many. 

Is this a new solution for Higher Education and Human Resources managers? 

The new watermarking laws may well offer new solutions to some of the chaos university faculty and recruiters have been navigating since the pandemic. When universities can once again provide undisputed evidence of plagiarism, full academic standards can return to our educational institutions. 

 

Currently, many students submit assignments that are unquestionably AI generated, but without proof, the university cannot fail the work. Reliable AI detection will soon be added to plagiarism checkers. The experiment with AI content may even benefit future students, as many universities have responded positively, reviewing the suitability of their assessment practices for an AI world. We may very soon get better, stronger and more robustly tested cohorts of students from academic institutions than we have ever seen before. 

 

It may also offer a lifeline for Human Resources. Job application spamming has become the norm. The number of applicants per role has increased at shocking rates, leaving recruiters often struggling to separate skills and talent in a hundred almost identical, AI generated applications. It is hard enough to predict future performance from the brief application process, without the interference of tools that are designed to persuade. 

 

Many on both sides of the recruitment process would prefer to return to a system where real talent and the individual unique contribution each applicant brings are upfront and visible.. The capability to screen out AI content early will very soon be added - ironically - to AI resume scanning tools. The current hiring market is dysfunctional for applicants and recruiters alike. 

 

The IoA Portfolio offers credibility in a sea of AI generated CVs

At the IoA we have developed assistance for our members to make credible job applications with the addition of a dynamic evidence-of-skills portfolio. Every member can build and share their own digital portfolio, which is organised into the 4 sections of a gold standard industry framework: The IoA Analyst Competency Framework. This essentially builds your personal brand and can be shared via a link with recruiters or put into your CV, where employers can interact with your various forms of real-world evidence. Stand-out: Don’t just tell them about your skills - show them in your portfolio. Find out more here: Institute of Analytics (IoA) | Data Careers & Membership