AI-Assisted or AI-Written? I Think We’re Asking the Wrong Question
A few days ago, I saw that Anthropic had announced it was introducing machine-readable watermarks into text produced by future Claude models.
And, honestly, it sent me down a bit of a rabbit hole.
The watermark won’t be visible and it won’t add hidden characters to the text. It will allow detection tools to assess the likelihood that Claude was involved in producing it.
I understand why Anthropic is doing it. The company has said the change is part of its response to the EU AI Act, and transparency around AI-generated content clearly matters.
But what does “AI-generated” actually mean now?
If I write an email and ask AI to check the spelling, has AI generated it?
What if I ask it to make a paragraph clearer because I know what I want to say but can’t quite get the words right?
And what about meeting transcripts, summaries, suggested email responses and all the other AI features appearing in the software we already use?
Maybe I am overthinking it. But I keep coming back to the same point.
Knowing that AI was involved doesn’t tell us what it actually did.
AI-assisted and AI-written can mean very different things
“AI-assisted” and “AI-written” aren’t legal definitions. They are simply terms I find useful when thinking about how AI is being used at work.
We’ve accepted technological help with our writing for years. Spellcheck fixes errors, Word suggests grammatical changes and Teams transcribes meetings. Most people wouldn’t say Microsoft Word wrote their report because it corrected a spelling mistake.
Generative AI can go much further, which is where things become a little less clear.
Take a performance review.
Someone could open an AI tool and type:
“Write a performance review for this employee.”
Another manager might write the review themselves and then ask:
“Can you make this easier to read without changing what I’m saying?”
Both managers have used AI. A watermark or declaration could treat those two examples in exactly the same way.
From an HR perspective, they don’t feel the same to me at all.
In the first example, I would want to know what information the AI was given, where the comments came from and whether the manager had applied their own judgement.
In the second, AI has helped someone communicate their own thoughts more clearly.
There are plenty of grey areas between those examples, of course. That’s really the point. AI use doesn’t fall neatly into two boxes.
There are levels to how AI is involved
Sometimes AI is doing fairly basic administrative work. It might organise notes, transcribe a conversation, format information or correct spelling.
It can also act more like an assistant. You might give it your existing ideas and ask it to suggest a structure, summarise a document or point out questions you haven’t considered.
Then you have AI generating content. Perhaps it produces the first draft of a policy, report or employee communication. That could be perfectly useful, but someone still needs to check it properly. A well-written response can contain inaccurate information, poor reasoning or assumptions that don’t fit the situation.
The area that concerns me most is when an AI output begins to influence decisions about people.
It might help shortlist candidates, assess performance, identify employees for redundancy or recommend an outcome in an employee relations case.
At that point, I’d want to understand exactly how it was used. What information did it consider? Was anything important missing? Did someone challenge the recommendation? Who took responsibility for the decision?
These questions matter far more to me than whether AI contributed some of the wording.
Why this matters in HR
HR work is full of context, history and information that might not appear in the document sitting in front of you.
An employee relations investigation is a good example.
Using an approved AI tool to summarise 40 pages of meeting notes could save hours of administration. That time could be spent checking evidence, speaking to the people involved and considering what is fair.
I can see real value in that.
I would feel very differently about someone uploading the notes and asking AI what disciplinary sanction should be given.
In my own HR career, I’ve seen how much time disappears into lengthy notes, repeated drafting and administration. I want AI to help with that work. It has the potential to give HR teams more time for the parts of the role that need human attention.
I also know how easy it can be to rely on something because it looks polished and sounds confident. That’s where we need to be careful.
What does the EU AI Act say?
Article 50 of the EU AI Act became applicable on 2 August 2026. It includes transparency requirements for providers and deployers of certain AI systems.
Providers of generative AI systems are required to make AI-generated or manipulated content detectable in a machine-readable format.
There are separate disclosure requirements covering particular uses, including deepfakes and certain AI-generated text published to inform the public about matters of public interest.
I’m not offering a legal interpretation of Article 50 here. There are people far better qualified to do that.
What interests me is the conversation this creates for employers.
A machine-readable watermark might tell us that AI was involved in producing a piece of text. It can’t explain whether Claude wrote the whole thing, reorganised someone’s existing thoughts or corrected a sentence.
That missing context is important.
Should employees declare every use of AI?
I’ve seen AI policies that require employees to declare whenever they’ve used AI. I can understand why an organisation might take that approach. It feels clear and safe.
I’m less convinced it will work as AI becomes part of everyday workplace software.
Would someone need to declare a Teams meeting summary? What about a suggested response in Outlook, a Microsoft Copilot search or a grammar correction?
Where would the line be drawn?
There is also a slightly cynical part of me that wonders whether an outcome like this might suit AI providers.
Anthropic has introduced its watermarking approach to comply with the EU requirements.
That is entirely understandable. But the same machine-readable marker could appear when Claude generates an entire document and when it helps improve a single sentence.
I have no idea whether this formed part of Anthropic’s thinking. It does make me wonder what happens when almost everything produced with AI support carries the same marker.
Do people eventually stop paying attention to it?
We’ve all seen warnings so often that they become part of the background. You click, accept or acknowledge them without really thinking about what they mean.
The same could happen here. The requirement is met, the marker is present, and yet we understand very little about how the work was actually produced.
I’d rather see organisations explain the situations where AI use needs greater transparency or review. That could vary depending on the organisation, the information involved and the effect the work may have on another person.
What should employers be asking?
I don’t think there is one rule that will work for every organisation. These are some of the questions I would start with:
What was AI asked to do?
Was personal, confidential or employee information entered into it?
Was the tool approved by the organisation?
Did it create substantive content?
Could the output affect a decision about another person?
Has someone checked the accuracy, context, fairness and tone?
Could the person responsible explain how the final result was produced?
Who is accountable for the outcome?
The answers would help me judge whether a particular use of AI was fairly routine or whether it needed stronger controls.
Polished work isn’t always good work
One of the strange things about generative AI is its ability to make almost anything sound convincing.
It can turn limited thinking into a beautiful report. It can produce a confident recommendation based on incomplete information. It can make something sound professional even when the reasoning behind it is weak.
I don’t see that as a reason to avoid AI. It’s a reason to stay alert.
There is a lot of administrative work that AI can help us with. There are also plenty of situations where it can help someone organise their thoughts, get past a blank page or communicate more clearly.
Those uses can be genuinely valuable.
The danger comes when the polished output stops us asking whether the thinking behind it is any good.
The question I’m left with
I’m sure people will have different views on how much AI use should be declared and when a piece of work should be described as AI-generated.
I’m still working through some of those questions myself.
But if your organisation is developing an AI policy or guidance for employees, this is the question I would take back:
Are we governing how AI is being used, or are we getting better at detecting that it was there?
A watermark may help answer one of those questions.
I’m not convinced it answers the one employers most need to understand.
About HRnetics
HRnetics helps HR teams and organisations understand how AI is already being used, where it could genuinely help and where greater care may be needed.
If you’re beginning that conversation in your organisation, the free AI in HR Checklist can help you identify the tools, uses and potential risks that may already exist across your HR function.
Comments