AI-Assisted Campaign Content: Human Review Checklist

The eight checks are sequential, not parallel. Each check gates the next. Content that fails Check 1 (factual accuracy) should not proceed to Check 2 — not because the other checks are less important, but because factual errors change the legal and reputational risk calculus for every subsequent check. A content item that does not pass all eight checks should not be published. It should be returned for revision or, if revision is not possible in the time available, held.

Check 1, factual accuracy, is the most time-consuming check and the most frequently skipped. AI language models produce plausible-sounding text — they are optimised to do so. That optimisation does not include a commitment to accuracy. Statistics may be wrong, dates may be hallucinated, names may be transposed, and quotations may be invented. Every factual claim in AI-generated content must be verified against a primary source — a published report, a verified news source, an official government or institutional document — before the content moves forward. If a claim cannot be verified, it must be removed or rewritten.

Check 3, AI disclosure, is where many campaigns currently have a gap. The default position should be to disclose. If your content was drafted with AI assistance, label it. The EU AI Act requires disclosure in any case where the content is "likely to be perceived as authentic" — which, for polished campaign copy, is almost always the case. The disclosure does not need to be prominent or to undermine the content's impact; it can be a small-text statement at the foot of a page, a metadata tag on an image, or a platform-standard label on an ad. What it cannot be is absent.

Check 5, targeting lawfulness, is specific to content that will be delivered to individuals using personal data for targeting. Under GDPR, processing personal data to deliver political content to individuals requires a lawful basis. In most EU member states, political opinions are special category data under GDPR Article 9, attracting the stricter conditions of Article 9(2). Inferred political opinion — even where the individual has not disclosed their views — is increasingly treated as special category data by EU data protection authorities. If your targeting parameters involve any inference about political preference, interest, or behaviour, you need to confirm the lawful basis and data minimisation approach with a qualified data protection professional before the content goes live.

Check 7, human approval, requires a named accountable human to sign off on content before publication. "Named" and "accountable" are both essential. An anonymous approval step, or an approval by an automated system, does not meet the standard. The person signing off must be identifiable, must have the authority to halt publication, and must understand that they are taking responsibility for the content at the time of approval.

For most campaign structures, the most appropriate sign-off hierarchy depends on the type of content and its risk profile. Routine social media posts drafted with AI assistance might require sign-off from a digital manager. Paid advertising must require sign-off from a compliance lead and the campaign manager or director. Any content referencing regulatory matters, legal proceedings, named individuals, or policy positions should require legal review — even a short legal screen — before the communications sign-off. Any synthetic media must require the most senior communications authority the campaign has, because synthetic media carries the highest legal and reputational risk.

The sign-off record should be created at the time of approval, not reconstructed after the fact. A simple log — content item, date, reviewer name, what was reviewed, and the outcome — is sufficient for most campaigns. Larger campaigns with high-volume content production should consider a content management system that builds the audit trail automatically as part of the approval workflow. What cannot be acceptable is a system where approved content cannot be traced back to a named approver, because that system provides no accountability and no compliance defence.

The deadline override is the most dangerous failure mode. A campaign needs content for a time-sensitive moment — a breaking news event, a conference appearance, a response to an opponent — and the review process is shortened or skipped to meet the deadline. This is exactly the scenario in which AI content errors are most likely to cause damage, because the speed pressure that created the override also means less time for human fact-checking and correction after publication. Build the review process so that it can move quickly — but build it so that it cannot be bypassed entirely.

The distributed responsibility failure happens when no one person owns the review process. If factual accuracy is "everyone's responsibility," it is nobody's responsibility. Assign a named individual to each of the eight checks for each piece of content. That individual is accountable for their check. If they are not available, a named deputy must cover the check. Content does not proceed until the check is complete and recorded.

The synthetic media blind spot occurs when campaign teams treat AI-generated images, audio, and video as equivalent to stock photography or licensed footage. They are not equivalent. Synthetic media carries specific disclosure obligations under the AI Act, specific risks around the representation of real people (who may have legal recourse if their likeness is used without consent), and specific risks around misidentification — AI image generators routinely produce images of real people who are unrecognisably altered, potentially creating defamation exposure. Every piece of synthetic media must be identified as such at the point of production, before it enters the review queue, and must be subject to the synthetic media check.

Finally, the audit trail gap is a failure mode that only becomes visible when something goes wrong. Campaigns that cannot demonstrate, with a contemporaneous record, what was reviewed, by whom, and when, have no compliance defence if a piece of AI-generated content is challenged by a data protection authority, an electoral commission, or a court. The audit trail is not bureaucracy — it is the evidence of the human oversight that the AI Act requires.