From Regulation to Reassurance: The EU AI Act on the Signal Degradation Curve.

Where it starts

At the source, the claim is precise. The amendment is Regulation (EU) 2026/1744, published in the Official Journal on 24 July 2026.^[2] It moves the rules for high-risk uses such as hiring and education to 2 December 2027, and the rules for AI built into products to 2 August 2028.^[4] It leaves the transparency rules and the Commission's enforcement powers on their August 2026 schedule.^[5] There's one narrow exception: generative AI systems already on the market get until 2 December 2026 for the content-marking part.^[6]

That's a careful, conditional claim. It's also almost impossible to fit in a headline.

The first squeeze

This is the surprising part: the squeeze starts at the source. The Commission's own summary says the regulation became applicable on 2 August 2026, “with some exceptions.” ^[4] Those exceptions include the high-risk rules most buyers care about. It's accurate, but the most important information sits in a qualifier.

Commentary

Next comes the layer of guides and compliance summaries, where the conditions start dropping off.

One guide, marked as updated in July 2026, still tells readers that high-risk obligations become directly applicable on 2 August 2026.^[7] That was two months after the postponement was agreed.

Another, from a compliance vendor, says organisations that signed the transparency Code by the July deadline secured a presumption of conformity.^[8] The regulation reserves that phrase for harmonised standards once they're cited in the Official Journal.^[3] The Commission and the AI Board describe the Code as an adequate voluntary tool to demonstrate compliance, which is useful but a different thing.^[9]

The headline

By the time it reached general coverage, the whole amendment had been reduced to five words: the AI Act was delayed. As one analysis put it, that summary is wrong, and for a compliance team it's actively dangerous.^[10]

The sales call

At the end of the curve, it reaches the buyer as a single sentence. We haven't recorded a sales call; these are illustrations of the kind of line buyers hear.

The dates are gone. So is the split between who builds and who uses. So are the conditions. What's left is confidence.

The twist: degradation may have a direction

Here's what the AI Act suggests about the framework. The curve doesn't only exaggerate. It seems to bend each claim toward the interest of whoever is telling it. If you're selling compliance, you tend to hear urgency. If you're selling a product, you tend to hear delay. Both simplify, in opposite directions, and both end up at a confident sentence. The examples above fit that reading. We'll keep testing it.

We watched the curve run in real time with the first enforcement step. The Commissioner didn't name who received the requests for information, and several major labs hadn't responded to reporters' questions.^[11][12] A few days later, one guide said the recipients reportedly included OpenAI, Anthropic and Google.^[13] Then a Spanish-language outlet, citing that guide, dropped “reportedly” and named them outright.^[14] From unnamed, to “reportedly,” to named, in three retellings.

What it means for a buyer

The version of a claim that reaches the person making the decision is the one furthest from where it started. And the voices closest to that buyer, especially the small vendors serving SMEs and healthcare, often have the least reason, and the least capacity, to put back what got lost along the way.

So how does an organisation trace a vendor's “we're compliant” back to what the regulation actually requires, before it acts on it? That's a separate conversation.

References