Bugs per developer under high AI adoption rose 9% last year and 54% this year. What the deterioration reveals about where the bottleneck in your pipeline actually moved.
Most engineering teams tracking EU AI regulation spent the first half of 2026 preparing for August 2.
That was the date high-risk obligations under Annex III were due to become enforceable: conformity assessments, CE marking, EU database registration. A heavy lift for any organisation in scope.
On June 29, 2026, the Council of the European Union gave final approval to the Digital Omnibus package, moving the Annex III deadline to December 2, 2027, with product-embedded high-risk systems under Annex I moving to August 2, 2028.
Many teams read the headline and filed the entire regulation under next year.
August 2, 2026 did not become an empty date. According to the European Commission's own guidance on Article 50, the transparency obligations applied from that date as scheduled. The Commission published its final Article 50 guidelines on July 20, 2026.
Those duties have now been in force for four weeks.
The default assumption in most planning meetings is that AI regulation targets model builders. Article 50 does not work that way.
It splits duties between the provider, which develops an AI system, and the deployer, which uses one under its own authority. An organisation with no high-risk AI anywhere in its stack can still carry obligations simply by operating a customer-facing chatbot or publishing model-generated content.
The obligations also reach beyond the EU. A provider or deployer established anywhere is in scope when serving EU users, following the same extraterritorial pattern GDPR established.
Penalties reach 15 million euros or 3% of total worldwide annual turnover, whichever is higher. For SMEs and startups, the lower figure applies.
| Scenario | Who owes the duty | What is required |
|---|---|---|
| AI system interacts directly with people | Provider (system design) | Design so the person knows they are dealing with a machine |
| AI generates or manipulates synthetic content | Provider | Mark output in machine-readable form as artificially generated |
| Emotion recognition or biometric categorisation | Deployer | Inform the people exposed to the system |
| Deepfakes and AI-generated public-interest text | Deployer | Disclose that content is artificially generated or manipulated |
Per Addleshaw Goddard's analysis of the transparency rules, the deepfake category has practical edges worth understanding: routine touch-ups to an image are generally exempt, while altering a product to look better than it actually is likely is not.
One carve-out matters more than any other for teams with products already live.
For AI systems placed on the market before August 2, 2026, the marking and detection obligation specifically, Article 50(2), applies from December 2, 2026 rather than immediately. Content generated before August 2 does not require retroactive labelling. The date of generation is what counts.
The disclosure duties, meaning informing users they are interacting with AI, were not covered by that grace period.
If your product was live before August, your machine-readable marking obligation is 93 days out at the time of writing.
Our read: the Annex III delay produced a reasonable but expensive misreading. The AI Act calendar split into separate tracks, and the track most small SaaS companies occupy is the one that did not move.
This is where compliance stops being a memo and becomes a sprint.
The Commission confirmed the Code of Practice on Transparency of AI-Generated Content, published June 10, 2026, as an adequate route to demonstrating compliance. It points toward established watermarking, metadata and provenance standards rather than a bespoke EU format.
In practice, two families of approach dominate.
The IPTC DigitalSourceType property is the established vocabulary for declaring how media was produced. For fully AI-generated output:
# Mark an image as fully AI-generated (IPTC DigitalSourceType)
exiftool \
-XMP-Iptc4xmpExt:DigitalSourceType="https://cv.iptc.org/newscodes/digitalsourcetype/trainedAlgorithmicMedia" \
-XMP-dc:Creator="YourProduct" \
-overwrite_original \
generated-image.jpg
# Verify it wrote correctly
exiftool -XMP-Iptc4xmpExt:DigitalSourceType generated-image.jpg
For content where a model assisted but a human meaningfully edited, the vocabulary distinguishes compositeWithTrainedAlgorithmicMedia and related values. That distinction maps onto the fully-generated versus AI-assisted taxonomy the Code of Practice describes, so choosing the wrong value is a substantive error rather than a cosmetic one.
For stronger, tamper-evident provenance, C2PA Content Credentials attach a cryptographically signed manifest rather than a plain metadata field. That is the more durable option where content is likely to be redistributed.
This is where most implementations fail, and it is entirely an engineering problem rather than a legal one. Image processing libraries strip metadata by default in many configurations.
# Common failure: Pillow drops XMP/IPTC on save unless explicitly carried over
from PIL import Image
img = Image.open("generated-image.jpg")
resized = img.resize((800, 600))
# This silently discards provenance metadata
resized.save("resized-BROKEN.jpg")
# Carry the original bytes through instead
exif = img.info.get("exif")
xmp = img.info.get("xmp")
resized.save(
"resized-OK.jpg",
exif=exif,
xmp=xmp,
)
Run the same check against every transformation stage: upload handlers, thumbnail generation, CDN image optimisation, and any third-party media service. A CDN configured to strip metadata for payload size will quietly undo a correct implementation upstream.
Work through this with an engineer who knows the media pipeline, not only with counsel:
Compliance implementation has an awkward shape for a small engineering team. The work is genuinely technical, clearly bounded, externally dated, and belongs to nobody.
It also tends to land on whoever has capacity rather than whoever understands the media pipeline. That is precisely how a team ends up writing provenance metadata correctly at generation and stripping it three transformations later, then discovering the gap during an audit rather than during a sprint.
This is the surface area problem we described in our March analysis of AI-expanded scope. AI did not only add capability to your product. It added obligations attached to that capability, arriving on a schedule you do not control. It compounds the pipeline strain documented in our analysis of delivery bottlenecks, where downstream stages were already absorbing more than they were sized for.
The engineers we embed at Percime Technologies take work of exactly this shape: bounded, technical, deadline-driven, and not worth a permanent hire. Media pipelines, metadata, logging, and the parts of a system that must be correct by a specific date and then continue working quietly afterward.
The high-risk deadline moved. Article 50 did not. December 2 is 93 days out.
© 2026 Percime Technologies. All rights reserved.