agentage.org / Articles / Explainer

A0, A1, A2: How AI Involvement Labels on News Work

A practical guide to the three-level labeling scheme that replaces binary detection with transparent provenance.

By Zoltán Varga2026-11-055 min read

Why a Binary AI Label Is Not Enough

A simple yes-or-no label fails to capture the reality of modern news production. Most content exists on a spectrum between fully human creation and autonomous generation. Software detectors cannot reliably distinguish between these states, especially when human editors modify AI output. Relying on detection tools creates a false sense of security and ignores the actual workflow used to produce the text.

The European Union recognizes this complexity in its regulatory framework. The AI Act distinguishes between raw AI output and content that undergoes editorial control. A binary label cannot reflect this legal distinction. It forces publishers to choose between hiding their tools or mislabeling hybrid content. This ambiguity undermines trust rather than building it.

Readers need to know who is responsible for the information they consume. They care less about the specific tool used and more about the accountability of the author. A nuanced labeling system provides this clarity. It shifts the focus from technical detection to provenance verification. This aligns with the goal of creating a 'non-slop zone' where the origin of every fact is known and trusted.

Current estimates suggest that a significant portion of online content involves some level of AI assistance. However, the exact percentage varies widely depending on the sector. Advertising and search results show higher rates of automation than editorial journalism. Treating all automated content as identical ignores these critical differences in quality and intent.

• Binary labels obscure the difference between assisted drafting and full automation.
• Software detectors fail on hybrid content edited by humans.
• Regulatory frameworks require disclosure based on editorial responsibility, not just generation.

A0: Human-Written Content

The A0 label applies to content created entirely by human authors without AI assistance in drafting or research. This category represents the traditional model of journalism and academic writing. The author takes full responsibility for every fact, argument, and stylistic choice. No generative models were used to generate text, structure ideas, or summarize sources.

Publishers often reserve this label for opinion pieces, investigative reports, and deep-dive analysis. These formats rely heavily on the unique perspective and expertise of the writer. The value lies in the human judgment applied to the subject matter. Readers expect this level of scrutiny and personal accountability from A0 content.

Maintaining the integrity of the A0 label requires strict internal guidelines. Authors must declare if they used any AI tools during the process. Even using an AI for grammar checks or basic fact-checking might shift the classification to A1. Clear documentation of the workflow ensures the label remains accurate and defensible.

This category serves as the baseline for trust in the news ecosystem. It demonstrates that human intellect and ethical oversight drive the publication. While AI-free content is becoming rarer, it remains a premium standard for high-stakes reporting. The label signals to readers that the information comes from a verified human source.

• No AI tools used for drafting, research, or summarization.
• Full human accountability for all facts and arguments.
• Reserved for opinion, investigation, and expert analysis.

A1: AI-Assisted Research and Transcription

The A1 label covers content where humans write the final text but use AI for specific support tasks. Common uses include transcribing interviews, summarizing background documents, or brainstorming initial ideas. The human author retains full control over the narrative and the final wording. The AI acts as a tool to increase efficiency, not as a co-author.

This category reflects the most common workflow in modern newsrooms. Reporters use AI to process large volumes of data quickly. They verify the output against primary sources before incorporating it into the story. The final article is a product of human synthesis and judgment. The AI contribution is limited to the preparatory phase.

Transparency is key for A1 content. Publishers should specify which stages involved AI assistance. Did the AI transcribe the interview? Did it summarize the court documents? These details help readers understand the depth of human involvement. It prevents the misconception that the entire piece was machine-generated.

The distinction between A1 and A2 hinges on who writes the prose. In A1, the human writes the sentences. In A2, the AI generates the draft. This difference matters because it changes the nature of the editorial review required. A1 content still demands fact-checking, but the structural risks associated with AI hallucination are lower.

• Human authors write the final text.
• AI used for transcription, summarization, or brainstorming.
• Human retains full control over narrative and facts.

A2: AI-Drafted, Human-Reviewed

The A2 label applies when an AI system generates the initial draft of a news item. A human editor then reviews, verifies, and edits the content before publication. This process is common for routine updates, financial summaries, or sports results. The AI handles the bulk of the writing, while the human ensures accuracy and context.

Under EU regulations, this category is exempt from mandatory disclosure if the content undergoes rigorous editorial control. The law assumes that a responsible editor can catch errors and bias. However, voluntary labeling builds greater trust. It informs readers that the text originated from a machine, even if a human signed off on it.

The review process for A2 content must be thorough. Editors cannot simply proofread for typos. They must check facts, verify sources, and ensure the tone matches the publication's standards. This step is crucial to prevent the spread of misinformation. The human editor becomes the final gatekeeper of truth.

Publishers should document the review process for A2 items. This record proves that the content met editorial standards. It also protects the organization in case of disputes. The label signals that while the AI did the heavy lifting, a human expert validated the result.

• AI generates the initial draft.
• Human editor performs rigorous fact-checking and editing.
• Exempt from mandatory disclosure if editorial control is proven, but voluntary labeling is recommended.

Boundary Cases and Relabelling

Content workflows often change during the editing process. A story might start as A2 but evolve into A1 if the editor rewrites significant portions. Conversely, an A1 draft might be expanded by AI to meet word counts. The label must reflect the final state of the article, not the initial intent. Static labels become inaccurate as the work progresses.

Translation presents a specific challenge. If an AI translates a human-written article, does it become A2? The answer depends on the extent of the translation. Machine translation alone usually keeps the content in the A1 category if the original human authorship is preserved. However, if the AI rewrites the text to fit a new cultural context, it may cross into A2 territory.

Relabelling requires a clear protocol. Editors must update the label whenever the level of AI involvement changes. This ensures the metadata remains truthful. Automated systems can track these changes, but human oversight is necessary to make the final call. Consistency across the publication is essential for credibility.

The goal is to maintain a 'provenance chain' that tracks the origin of every element. This chain survives edits and revisions. It allows readers to trace the history of the content. By updating labels dynamically, publishers demonstrate their commitment to transparency. This practice strengthens the overall trust architecture of the news agency.

• Labels must reflect the final state of the article, not the draft stage.
• Translation workflows require careful assessment of AI rewriting versus direct conversion.
• Dynamic updating of labels ensures accuracy throughout the editing process.

AI-assisted: drafted with a language model from the author's research notes, then edited and checked before publication. Author: Zoltán Varga, AI visibility researcher and AI ecosystem architect, Budapest.