Multimodal AI Risks, Limits and Governance Rules to Set First

Common uses include describing and tagging product images, extracting data from screenshots of reports, generating alt text, reviewing ad creatives against brand guidelines, transcribing and summarising video interviews and generating images or short clips for campaigns.

The appeal is speed on tasks that used to need a person looking at every file. A catalogue of thousands of product photos can get draft alt text and attribute tags in an afternoon.

Multimodal models misread more often than text-only tasks suggest. Small text in images, dense charts, handwriting and unusual layouts cause errors. Counting objects and reading exact numbers from charts are known weak points. The model may describe what it expects to see rather than what is there.

• Fine detail: small labels, fine print and low-resolution images are often misread.
• Charts: models may estimate values incorrectly or invent a trend.
• Context: a model does not know a photo was staged or that a screenshot is outdated.
• Audio: accents, cross-talk and background noise reduce transcription accuracy.
• Generation: hands, text inside images and brand logos are common failure points in generated visuals.

Personal data is the first. Photos from events, screenshots of CRM screens and recorded calls often contain faces, names and contact details. Uploading them to an AI tool is processing personal data, with all the legal obligations that brings in regimes like GDPR or Turkey’s KVKK.

Rights are the second. Generated images may resemble existing work, and terms on commercial use differ by tool. Uploading a photographer’s images for “inspiration” may breach the licence you hold.

Prompt injection through images

Text hidden in an image, such as small or low-contrast instructions, can be read by the model and treated as a command. If an assistant processes uploaded files from customers or the web and can take actions, this becomes a real security issue. Treat any image or document from outside as untrusted input.

• Upload rules: list media types that may not be uploaded, such as identity documents, customer screenshots with personal data and unlicensed images.
• Approved tools: name the tools and plans with acceptable data terms.
• Human review: any extracted figure used in a report and any generated asset published externally is checked by a person.
• Labelling: decide when generated or edited images and videos must be disclosed, in line with platform rules and local law.
• Consent: get consent before recording or processing people’s images and voices.
• Storage: define where prompts, inputs and outputs are kept and for how long.
• Action limits: assistants that read untrusted files must not trigger actions without approval.

Major platforms have introduced labelling for AI-generated or significantly altered content, particularly for realistic images and political or social issue ads. Meta, YouTube and TikTok all have disclosure requirements of some kind; check each platform’s current policies before publishing. The EU AI Act includes transparency obligations for certain AI-generated content, including deepfakes.

Content credentials, based on the C2PA standard, attach provenance information to media files. Some generators and cameras add them automatically. They are worth preserving in your workflow, since stripping metadata removes a useful trust signal.

Sample regularly. For alt text, review a random set each month for accuracy and usefulness. For data extracted from images, compare a sample with the original files. For generated visuals, keep a log of prompts, tool, plan and approval, so you can answer questions about origin later.

Track error types, not just error counts. If most errors come from low-resolution inputs, the fix is better source files, not a different model.

Ownership usually sits best with whoever already owns data protection or brand standards, working with marketing operations. A single named person keeps the policy current, approves new tools and handles exceptions. Without an owner, rules written at launch drift out of date as tools add features such as video generation or screen reading.

Keep the policy short enough to read in five minutes. A long document that nobody opens protects no one. Review it every quarter, and update it whenever a tool your team uses adds a new media type or changes its data terms.

Alt text as a quick win

Generated alt text is one of the safest first uses, because the stakes are low and review is quick. Ask for concise descriptions of what the image shows and its purpose on the page, not keyword lists.