ASA publishes results of 'cutting-edge' trial using AI to monitor alcohol ads

The question

How will AI-driven monitoring by the ASA change the advertising landscape?

The key takeaway

AI assessments enhance the speed and scale at which the ASA can identify non-compliant advertising. This development is likely to lead to an increase in proactive enforcement and less reliance on complaint-driven investigations. Staying ahead requires rigorous CAP Code adherence by advertisers and proactive risk checks to avoid being flagged by automated systems.

The background

The Advertising Standards Authority (ASA) has published the findings of its recent trial integrating cutting-edge AI in the assessment of online advertising for alcohol.

Using its internal Active Ad Monitoring system, the trial captured nearly 6,000 paid ads served to UK consumers over a one month period through social media, search, and display channels in early 2025. The ads were then evaluated by large language models (LLMs) capable of interpreting and applying the rules set out in Section 18 (‘Alcohol’) of the CAP Code. As part of its review, the LLMs (also referred to as ‘AI agents’) highlighted ‘probable non-compliant content’ warranting further human review.

The ASA report states that: “overall compliance was very encouraging: 96% of ads were likely compliant with the CAP Code, with just 1-3% (101 total) assessed as likely non-compliant, with the remainder being ambiguous requiring further investigation". The rate of non-compliance was significantly higher within the alcohol-free category (also monitored under Section 18 regulations). In this category, 48% of ads were assessed as having compliance issues.

Though largely a success, the trial highlighted the LLMs’ limitations. “While they can spot clear-cut rule breaches with impressive speed and consistency, they do not possess the same depth of judgement or contextual reasoning as human experts. As a result, the tool tended to flag many potential breaches - a significant proportion of which were ultimately judged not to be problematic upon closer (human) inspection.”

The development

The trial showed that the majority of ads within the sector are compliant, but the significant win was its illustration of AI’s ability to identify potential non-compliance with the Code at scale. The ASA’s aim of discovering whether AI could effectively flag a range of potential breaches was demonstrated through the LLM’s ability to systematically assess thousands of ads in minutes – significantly surpassing the capacity of a human team.

Why is this important?

The success of this trial has prompted further investment in the ASA’s Active Ad Monitoring system. Widespread implementation of tech-assisted assessment is expected to result in increased speed in issue identification, increased ASA responsiveness to noncompliance, and the ability to more swiftly recognise critical areas of advertising that may require increased enforcement, clarity, or guidance.

The automated boost in the scale and speed at which the ASA assesses ads for compliance will reduce the ASA’s reliance on complaint-driven investigations and spur the ASA’s ability to proactively identify breaches that require further investigation.

Any practical tips?

Though this trial highlighted the potential for AI review, it also identified its limitations. Preliminary LLM review easily spotted clear-cut breaches of the CAP code, but was less accurate in assessing the compliance of ‘borderline’ advertisements. To ensure compliance at the preliminary assessment, advertisers should strictly adhere to the CAP Code(s) regulating advertising in their sector. This means:

• staying alert and adaptable to CAP Code revisions, formal guidance, and ASA rulings;
• stress testing ads against recent ASA rulings to attempt to determine whether ads might survive ASA scrutiny; and
• surpassing minimum requirements by prioritising clarity, transparency, and full disclosure of material consumer information.

Winter 2025