In July 2026 the European Commission published a series of documents that directly affect the way professionals assess AI output. Where the discussion long revolved around the question of whether a model answers 'well enough', attention is now shifting to something more fundamental: can you demonstrate where a text comes from, who generated it and how it was checked? For everyone who works with confidential information, that is a relevant shift.
What the European Commission published in July 2026
The Commission issued the Guidelines on transparency obligations for providers and deployers of AI systems, together with an accompanying Code of Practice on Transparency of AI-Generated Content. According to these documents, the transparency obligations from Article 50 apply from 2 August 2026, with a transitional date of 2 December 2026 mentioned for some already existing generative AI systems.
The Code of Practice describes concrete measures: machine-readable marking, watermarks, detection and verification of AI-generated audio, image, video and text. The core of this approach is that transparency is not just a matter of a label at the top of a document, but of signals that travel with content throughout the entire chain. We call that provenance: the origin of content must be traceable.
Logging and traceability as the second pillar
The second pillar comes from the broader AI Regulation. On the European Commission's AI Act page it is repeated that logging of activity is part of the obligations for high-risk AI systems. Article 12 requires such systems to enable the automatic recording of events over the lifetime of the system, in order to ensure traceability.
For professionals this means that validation does not depend solely on the certainty a model radiates about its own answer. It is precisely the auditable recording of what happened — which source, which step, which moment — that forms the basis for verifiability. Model confidence is not proof; a traceable trail comes closer.
What NIST adds in concrete technique
Where the EU documents outline the obligations, NIST's Artificial Intelligence Risk Management Framework offers more concrete guidance. The Generative AI profile recommends assessing accuracy, quality, reliability and authenticity of generated output against known ground truth, using multiple evaluation methods. NIST also recommends reviewing and verifying sources and citations in the output, both before deployment and during ongoing monitoring.
These recommendations align with the transparency principle: source review and verification are not a one-off action, but a recurring process. The responsibility lies with the person who deploys the output.
Research shows how far automatic detection reaches
That automatic verification is developing is evident from a paper in the Findings of the Association for Computational Linguistics from 2026: Hallucination Detection in Long-Form Text Generated by LLMs. The authors introduce a benchmark for detecting hallucinations in long texts and propose a black-box, zero-resource method that checks factual consistency using hyper-relational knowledge graphs and multi-hop reasoning.
The importance of this work is twofold. On the one hand, it shows that ever stronger detection methods are being evaluated in a peer-reviewed manner. On the other hand, it confirms that hallucination detection is an active field of research with open questions — not a solved problem. Anyone who counts on a single automatic check to catch all errors overestimates the current state of the technology.
The shift: from 'the truth' to a verifiable chain
If you place the EU guidelines, the NIST framework and the ACL research side by side, one line emerges. AI output validation in 2026 is not about one model that determines what is true, but about a chain of marking, source control and auditable logs. Provenance, logging and claim-by-claim checking complement each other; none of the three is sufficient on its own.
For professionals with sensitive information — lawyers, notaries, company doctors, journalists, researchers and compliance teams — that is a workable perspective. The question shifts from "do I trust this answer?" to "can I check and account for this answer step by step?".
Where Vera fits in this chain
Vera is not a language model and not a chatbot, but a verification layer for those who work with confidential information. That position aligns with the shift described. In the workflow, pre-processing and anonymisation take place on EU infrastructure via the Semantic Privacy Shield; the process is designed to send only anonymised content to the selected AI models. If the privacy check does not succeed, nothing is forwarded.
By putting answers through multiple models and making the verification steps visible, Vera can help with checking AI output — in the spirit of the source control NIST recommends. It gives more insight into differences between models and into the substantiation of claims. Documents can be viewed and edited in the same secure environment via Vera Office.
What Vera does not do is equally important. Vera does not promise that an answer is correct and does not remove the risk of hallucinations; it makes the check possible and makes steps visible. The professional final judgement always remains with the user. That fits the new rules: transparency and traceability are aids, not a replacement for craftsmanship.