Imagine a manager sends a thoughtful message to one of her hard-working employees after a particularly exhausting week in the office. The message communicates personal warmth, but when the recipient discovers it was generated by AI, it somehow feels less genuine.
In the past, the recipient would only have discovered AI use if the sender disclosed it or through a lucky guess. Now, the European Union’s new transparency rules require companies to watermark content produced by generative AI systems in a machine-readable format. Systems already on the market have until December 2026 to comply.
OpenAI reportedly developed such a watermarking system for ChatGPT years ago but held it back for fear of alienating users. Google now uses a watermarking system called SynthID in its propriety AI models. Recently Anthropic followed suit, announcing that content generated through its Claude models globally will contain an invisible watermark.
In other words, whether you reside in the EU or Canada or elsewhere, your AI use will no longer remain private or undetectable.
As researchers studying the impact of AI on learning, including the relational and emotional dimensions of that, this development concerns us deeply.
A culture of surveillance
Recent research demonstrates that although many people can’t tell whether a personal message was written by AI or not, their perception of the sender changes the moment they learn it was AI-generated.
We tend to read the time and effort someone puts into their writing as signals that we matter to them. When that work is outsourced to a machine, it feels like some of that care is no longer there, even when the underlying emotions are genuine.
But the process of watermarking AI also feeds into an already prevalent culture of surveillance. This is especially evident in workplaces, where employees are more willing to accept a manager’s use of AI for a routine scheduling note than for a personal message that offers praise or feedback.
In the latter case, the manager’s sincerity is more likely to be questioned. Watermarking makes this evaluative judgment available on demand. In a workplace where checking each other’s messages for AI fingerprints becomes a habit, even casual communication can become marred with suspicion and doubt.
The same dynamic applies to our private lives, where the emotional stakes are even higher. Research on apology shows that the same message may draw less trust and forgiveness when it’s attributed to AI.
An apology is meaningful to the extent that it genuinely reflects the vulnerability, effort and thought of the sender. As a machine bears none of these qualities, a watermark can exacerbate the perception that an apology lacks sincerity and authenticity.
Slanguage: How the use of AI for apologies could cause the ‘Canadian Sorry’ to lose its soul
What the watermark can’t see
The problem is that a watermark tells only part of the story — that AI was likely involved in producing a text. It doesn’t specify the nature or extent of that involvement.
A heartfelt message you have drafted yourself before using AI to proofread and polish it is different from one you delegated entirely to AI. The watermark cannot differentiate between the two possibilities and in certain relationships — for example in friendship or marriage — this matters.
The burden of watermarking may also fall unevenly. The mark itself is a function of statistical word patterning and any extensive rewriting or paraphrasing can override it.
The EU Commission’s Code of Practice reinforces this unevenness: text under roughly 150 words is exempt from the watermarking requirement altogether, and assistive functions that don’t substantially change the input, such as grammar and spell-checking, are exempt as well.
As a result, a two-line apology sent by text message may carry no watermark, while a long letter of condolence will. Watermarking will mostly impact those who write at length, and those who use mainstream, transparent systems rather than models they run privately.
(Getty/Unsplash+)
Relationships run on something else
Of course, AI may have a role in our interpersonal communications. But human relationships are nurtured and maintained through trust, care and honesty. Watermarking does nothing to sustain relationships or enhance accountability to one another.
We support an open form of AI-use disclosure, one where people explain when, how and why they use AI. This does come at a cost: across 13 experiments, disclosing AI use measurably reduced trust compared to saying nothing.
However, voluntary disclosure could allow a person to control their own account of their AI use, rather than surrendering it to a detector.
A conversation about AI use belongs within a relationship between humans, where context, intention and trust can all be weighed together. A watermark offers something different: a verdict delivered after the fact to someone who may already suspect they’ve been deceived.
The emotional work of a relationship does not live in words alone. It lives in our willingness to stand behind those words and take responsibility for them.
