aigenerative aitechnologyai detectionsocial media

The Nuance Of AI Detection

Off-Topics / Bloqdigital @bloqdigital.pckt.cafe

Sep 26, 20267 min read

ack in April 2024, I wrote a piece entitled [Artificial] Intelligence, which looked at the less than delicate "diversity overcorrection" of AI generative outputs from the likes of Google's Gemini.

Now in 2026, despite the release of the EU's AI Labelling system we are still seeing a different kind of sensitivity issue. This time it's a lack of one, from various third party AI detection tools.

You see, whilst digital infrastructure is very much binary, the level of application a user / designer / creator might utilise AI tools for in their process can be very nuanced. The trouble is that these AI detection tools don't do well with nuance. Either you did use it, or you didn't. Or in some cases you did even if you didn't at all.

Remove a speck of dust from a photograph and you might find yourself having to convince people that you did, in fact, take the photograph. Dictate your own thoughts and ask a tool to organise them, and somebody else’s software may decide how much of the result belongs to you. Write without any assistance at all and you can still receive a remarkably confident assessment that a machine did the work for you.

Now, AI detection in principle is important. We are surrounded by convincing images of things that never happened, automated commentary and articles assembled with very little care for their accuracy. Readers deserve to know what they are looking at. Creators deserve protection from impersonation. Platforms need ways to deal with the sheer volume of material being uploaded.

The concern is what happens when a reasonable ambition becomes a small badge with far more authority than the evidence behind it deserves.

Earlier in June 2026, Christopher Penn described testing LinkedIn posts with Originality.ai, GPTZero and Winston AI. One contributor, identified as "Becca", dictated her material and used Claude to assemble it from her words. Penn reported that her posts were flagged five to six times as often as those of comparison writers. His experiment also reported roughly 15% false accusations across the detectors tested.

In Becca's case, a person had supplied the thoughts and spoken material, and the software had merely helped with their arrangement. AI had participated, but whether that makes the resulting post meaningfully “AI-written” depends on what we are trying to tell the reader.

In August, writer Andy Betts described deliberately producing an article without AI assistance and then testing it. Different detectors returned 100%, 78% and 42% AI, whilst another classified it as human. He also submitted his published 2019 work, which received AI scores of 35% and 16% from two different tools.

In September, The Verge reported renewed Instagram labelling complaints. Content strategist Jess Bruno said Canva had acknowledged that some assistive tools were being tagged as generative. The cosmetics brand About Face also disputed a label on an iPhone photograph that its social manager said had received only Photos-app edits. Those precise edits remained unconfirmed. The report described both over-labelling complaints and generated images escaping labels: a useful reminder that sensitivity and reliability are very different things.

A recent paper by Victor Angelier provides a more extensively archived example. Tests conducted on 30 August returned sharply different results for a manuscript described as human-authored. Winston AI reported 0% human. Copyleaks returned 80.4% AI, and Originality.ai 81% likely AI. GPTZero gave a human-generated category alongside a 42% AI score. A fifth tool leaned towards human authorship.

How can we truly use these tools regularly with such little consistency?

These disputes can become very public. Three finalists in the 2026 Commonwealth Short Story Prize faced detector-backed allegations. The Commonwealth Foundation subsequently reviewed drafts, time-stamped documents and notes, and spoke with the regional winners. It said it was satisfied that AI had not written the winning stories. Jamir Nazir went on to win the overall prize. In this instance, an automated assessment circulated as an accusation, whilst the organisers reached a different conclusion after examining the writing process. The reputation of a real person sat between those competing judgements.

Nor does the problem disappear if we accept more false positives in exchange for catching everything else. The systems miss material too. Reuters reported back in July that Meta’s preview image detector recognised the original images in its test but missed 55% after cropping. Meta acknowledged that substantial cropping could weaken the watermark. Not exactly a robust, fool-proof tool.

“AI detection” has become a catch-all description for several different activities.

Meta documents the use of industry signals such as C2PA and IPTC metadata, alongside user disclosures. It also embeds markers in its own generated images and has described developing classifiers for content without those markers. A metadata signal can record that a tool participated in an image’s production. It does not automatically explain how important that participation was.

LinkedIn’s Content Credentials badge exposes signed provenance information about images and videos, potentially including the camera, application and AI involvement. The badge alone does not mean the image was invented. Separately, LinkedIn uses its own systems to identify generic, repetitive content and limit its distribution. It's widely repeated 94% figure refers to identifying generic content in initial testing, rather than proving AI authorship with 94% accuracy.

LinkedIn also says member reports that something seems like AI slop inform its classifiers. I can see the appeal of asking people what they find valuable. I am less comfortable with the possibility that familiar formatting, polished English or an unpopular writing style becomes evidence against someone’s authenticity.

The EU’s approach offers a useful comparison.

Article 50’s transparency requirements began applying on 2 August 2026, with a transitional provision for marking by some existing systems. The framework separates machine-readable marking by AI providers from disclosure obligations on those using the systems.

It also recognises distinctions that broad labels can lose. Standard editing assistance, or processing that does not substantially alter the input or its meaning, can fall outside the provider marking obligation. Public-interest text has a disclosure exception where human review or editorial control takes place and a person or organisation assumes editorial responsibility. Artistic and fictional deepfakes have appropriately adapted disclosure requirements, rather than an unrestricted exemption.

The EU’s optional icons distinguish fully generated material from partially modified content. Their use is optional; the applicable legal duties are not. This is not a universal requirement to attach the same warning to everything an AI tool has touched.

That does not make every broader platform label unlawful. Nor would it be fair to claim companies like Meta has simply ignored the framework: it announced in July that it was signing the EU transparency code. The question is how well those distinctions survive implementation, and how clearly they reach the person scrolling past.

For me, useful transparency should explain enough of the process to support a judgement. A photograph with an assisted background edit, a dictated article organised by software and a fabricated news image have different histories and consequences. We should expect the information attached to them to reflect that.

These examples do not establish that every detector is useless. They establish limits on the confidence we should place in a result, especially when that result becomes a public accusation or affects someone’s livelihood. A flag can justify a closer look. It should come with uncertainty, meaningful context and a way to challenge it. Drafts, original files and editing records deserve more weight than an unexplained percentage.

Sources & Further Reading

Did you enjoy this article?

Recommend it — Standard Reader surfaces well-loved writing to more readers across the network.

OFF TOPICS - BLOQDIGITAL @bloqdigital.pckt.cafe