Contents:
This page lets you look at two advertisements side by side and ask simple questions about them. Where is the scientific claim? Is it scientifically reutable? Can we find out how the drug was tested? Where is the safety warning? And, perhaps less obviously, is the advert creating an impression through imagery which the words themselves never state?
The analysis answers such questions region by region. A region might be a heading, a block of body text, a warning, an image, or some other reasonably coherent part of the advertisement. For each region, the system gives you a score. You can see where the system thinks something — possibly something bad — is happening, how strongly it thinks it is happening, and why.
At the top of the page is a set of advertisement thumbnails. There can be several collections: each collection is called a project, and selected by one of the options on the Project dropdown above the thumbnails. A search box can then be used to narrow the collection, using part of an advertisement's name or part of its filename.
A few rows below the thumbnails is the main display. This has two sides, A and B. Choose A or B under Analysis controls, then click a thumbnail. That advertisement will be put into the corresponding side of the comparison. You can therefore compare two different advertisements, or two versions of the same advertisement.
The dates above the thumbnails are there mainly to help while we develop and test the system. They record when the advertisement image and its analysis were created or last changed.
The buttons immediately above the comparison control the analysis, selecting what is displayed. Original shows the advertisements without a semantic heat map. The other six buttons show six different semantic fields.
SD — Scientific Dodginess. This asks whether the material invokes science which is dubious, implausible or pseudo-scientific. Claims about impossible mechanisms, scientifically dubious entities, or supposed scientific principles which do not stand up would belong here.
ED — Episodic Dodginess. This is rather different. The underlying science may be perfectly respectable, but the evidence for this particular product may still be inadequate. ED therefore looks for weakly supported claims about what the product itself has been shown to do.
CF — Chardin Field. I name this after Peter Medawar's famous criticism of Teilhard de Chardin. Medawar complained about words such as energy, force, dimension and vibration being borrowed from science and then used metaphorically while retaining the authority of their scientific meanings. An advert talking about, for example, recharging one's immune system or moving the body to a higher plane might score on this field. Energy and plane have precise well-accepted meanings in science; but these are not being used here.
VPF — Vitality Projection Field. Not every claim is made in words. Pictures of sunlight, greenery, sport, movement, smiling people, mountains and fresh air can all suggest health, vitality and renewal. VPF is intended to detect this kind of projected impression.
SCF — Safety Concern Field. This concerns actual safety issues: possible adverse effects, unsafe combinations, contraindications. Even if an advert is perfectly reputable, these may matter to somebody deciding whether it needs amendment.
RSF — Risk Suppression Field. A safety issue can exist without being adequately communicated. RSF therefore asks whether risk appears to be hidden, minimised, made visually inconspicuous, or otherwise downplayed.
Once you select one of the six fields, the analyser will draw coloured regions over the two advertisements. These correspond to scores between 0 and 1. A score near 0 means that the selected feature is hardly present in that region; a score near 1 means that the analysis regards it as strongly present.
The colour strip is a key to these scores. It is worth remembering that the colours belong to particular regions. Even if there's a strongly coloured warning at the bottom of an advert, it does not mean that the whole advert has received that score.
Move your mouse over a coloured region and a small explanation appears. This gives the region identifier, its score, and the reason produced during the analysis. This is often more useful than the number by itself. A score of 0.8 tells you that the system has found something fairly strongly; the accompanying sentence tells you what it thinks it has found.
There is a Claims / regions section below the heatmaps and structure diagrams. This gives another view of the currently selected field. Instead of drawing the result on the advertisement, it lists the relevant regions in a table, together with their text or description, score and reason.
At present this table concentrates on regions with non-zero scores. Thus an empty table does not mean that the analysis has failed. It may simply mean that the current field has scored zero throughout that advertisement.
An advertisement is not treated as an indivisible rectangle. The analysis first divides it into regions and records how those regions fit together. A heading may contain text; an information panel may contain several smaller regions; an image may occupy another part of the page.
The Analysis structure section exposes this representation. It is collapsed by default because most of the time you will not need it. It is useful, however, when you want to know exactly what the analyser thought the parts of the advertisement were, or when a heat-map result looks surprising and you want to investigate.
The final section takes the semantic scores a stage further. These figures are not simply another set of scores. They are calculated from the six fields by a fuzzy-logic rule system.
Fuzzy logic is useful here because regulatory concern is rarely an all-or-nothing matter. A claim does not suddenly jump from harmless to seriously misleading because one numerical score passes an arbitrary boundary. Instead, rules can say things such as:
If episodic dodginess is high and Chardin field is high, then misleading-claim concern is very high.
The system currently calculates three kinds of result: misleading claim concern, risk minimisation concern, and review priority.
Under each result you can also see the rules which actually fired, together with their firing strengths. This is important. Rather than merely presenting a number, the system can show something of the route by which it arrived there.
For example, if ED is high while SD is low, that has a fairly natural interpretation. The scientific background may be plausible, but the evidence offered for the particular product may still be weak. The fuzzy rules can respond to that combination without requiring us to pretend that every kind of scientific problem is the same.
Some projects also contain supporting material. This may include background information, source documents or other material which helps in interpreting the advertisements. It is kept in a collapsed section near the top of the page so that it is available when needed without getting in the way of the comparison itself.
This is still a demonstrator rather than a finished regulatory system. In particular, the semantic analysis is being produced by a large language model. Repeating an analysis does not necessarily produce exactly the same numbers every time. Indeed, part of the purpose of the present work is to investigate how stable and useful the results are.
The regulatory-concern figures should likewise be read as indicators for review, not as legal judgements that an advertisement is or is not compliant. The interesting question at this stage is whether the system helps a human reviewer find potentially important material, understand why it has been flagged, and compare one advertisement with another.
That is also why the explanations matter. A coloured rectangle or a number may draw attention to something. The reason underneath tells us whether it has drawn attention to the right thing.