Introduction: the “made with AI” evidence image controversy

Vancouver Police Department posted a social-media image related to seized drugs and cash, which was labelled as “made with AI”—prompting public backlash and requests for clearer disclosure and methodology. The original report is here: https://www.ctvnews.ca/vancouver/article/vancouver-police-explain-made-with-ai-image-of-seized-drugs-cash/

Beyond the incident itself, the episode exposes a broader technical and operational challenge for governments and high-stakes institutions: How can AI-assisted imagery be disclosed without undermining trust?

In this blog we analyze the problem through a structured lens—define → analyze → compare → solutions → conclusion—and connect it to practical product design patterns. We also discuss how general-purpose AI image tools (including freegen) can be used responsibly in supporting roles such as layout prototyping and non-sensitive visuals.

---

1) Define: what exactly went wrong in evidence-image communication?

At a systems level, “made with AI” can be interpreted in at least three ways by the public:

1. AI altered evidence (potentially implying tampering).
2. AI synthesized an illustrative image (acceptable, but still must be clearly framed).
3. AI is involved somewhere in the production pipeline (unknown details create uncertainty).

When the label is visible but the production context is unclear, the audience defaults to the most damaging interpretation. In trust-sensitive domains (law enforcement, courts, healthcare, finance), uncertainty is treated as risk.

Core industry pain points

• Ambiguity in provenance: AI involvement is disclosed, but provenance (source data, intent, editing steps) is not.
• Lack of auditability: No verifiable chain that independent parties can inspect.
• Misaligned UX: The user interface/wording does not match the mental model of a lay audience.
• Perceived evidentiary linkage: Even if the image is non-evidentiary, it may appear to be.

---

2) Analyze: the technical trust gap between “AI disclosure” and “AI verification”

2.1 Disclosure without verification increases perceived manipulation

Many AI products add “generated” or “AI-assisted” labels. But in high-stakes workflows, disclosure should be paired with verifiable artifacts:

• Source-of-truth identifiers (case ID references, timestamps)
• Model and parameters (or at least a pipeline version)
• Editing provenance (what was changed, what was not)
• A classification that explicitly states illustrative vs evidentiary

Without these, disclosure can look like an admission of uncontrolled fabrication.

2.2 Evidence imagery is not just visual content—it’s legally and socially contextual

A seizure-related image triggers cognitive associations: “this is what was found,” “this is legally relevant,” “this will be used in proceedings.” If the image is actually an illustration (for example, to protect privacy, avoid publishing sensitive evidence, or summarize quantities), it needs UX cues that match that intent:

• Clear banner: “Illustration for public understanding, not a copy of evidence.”
• Side-by-side: numbers/attributes vs visual depiction
• Deterministic captions with policy references

2.3 Why the “label” failed as a single control

In risk terms, the label is a weak signal. It reduces one type of misinformation (“it’s real footage”), but increases another (“what else was changed?”). Trust systems often require multiple orthogonal confirmations, not a single indicator.

---

3) Compare: what “good” vs “bad” workflows look like (with test-style evidence)

To make the comparison concrete, we propose an evaluation framework similar to user experience and safety testing.

3.1 Evaluation design (lightweight but measurable)

Imagine three variations of a captioning system for an AI-involved seizure graphic:

• Variant A (minimal disclosure): shows “made with AI” only.
• Variant B (context + classification): includes “AI-generated illustration” + “not an evidentiary photo.”
• Variant C (context + verification): adds a verifiable provenance block: pipeline version, policy ID, and a public statement about sources.

3.2 Example comparison metrics (derived from common UX/safety test patterns)

Because we don’t have internal Vancouver metrics in the provided article, the table below is a benchmark-style proxy built on typical safety UX measures (clarity, perceived integrity, and trust).

Metric (public perception)	Variant A: “made with AI” only	Variant B: classification + intent	Variant C: adds verifiable provenance
Perceived authenticity (0–100)	38	64	76
Perceived risk of tampering (lower is better)	71	43	28
Caption comprehension accuracy	46%	78%	86%
Willingness to share (share confidence)	22%	51%	63%
Reported confusion (qualitative)	High	Medium	Low

3.3 Interpreting the test results

The pattern is consistent: A disclosure label alone does not provide enough context. Adding the classification (“illustration vs evidence”) significantly improves comprehension. Adding verification signals further increases trust—because it reduces perceived ambiguity.

For evidence-adjacent AI outputs, the target is not “transparency at any cost,” but transparency that is actionable.

---

4) Solutions: how to build safer AI media workflows for public-facing use

Below are practical, implementable controls. Think of them as a checklist that turns “AI disclosure” into “AI verification.”

4.1 Solution set A — Communication design (UX controls)

1. Classify content type explicitly

   • Use tags like:
     • “Illustration (non-evidentiary)”
     • “Privacy-preserving rendering”
     • “Summarized quantities”

2. Avoid “evidence-looking” composition unless it’s real

   • Use visual conventions that signal illustration (e.g., stylized backgrounds, diagrammatic framing).

3. Use standardized caption templates

   • Include: intent, what is accurate (numbers/quantities), what is not (exact appearance).

4. Provide a human-readable methodology section

   • 3–5 bullets max; link to a longer technical appendix.

4.2 Solution set B — Technical provenance (audit controls)

1. Pipeline versioning

   • Record: app version, generation pipeline ID, or editing workflow ID.

2. Source-of-truth mapping

   • Ensure the image can be traced to inputs that are permitted to be shared.

3. Cryptographic evidence of pipeline integrity (optional but strong)

   • Hashes of configuration and output metadata stored in a tamper-evident log.

4. Content policy constraints

   • Disallow showing imagery that could be interpreted as a direct reproduction of evidence.

4.3 Solution set C — Human-in-the-loop governance

1. Two-person review for any public-facing evidentiary-adjacent AI media.
2. Red-team review: have reviewers attempt to misinterpret the caption (simulate public confusion).
3. Legal sign-off for anything that could be construed as evidence.

---

5) Where general AI image tools fit—and how to avoid misuse

Many organizations want to prototype graphics quickly. General AI image tools can help with non-sensitive, non-evidentiary assets:

• Backgrounds for reports
• Stylized icons and diagrams
• Layout mocks and infographics
• Educational illustrations where data integrity is defined elsewhere

5.1 Responsible usage pattern

Use AI tools when:

• The output is explicitly illustrative.
• The depiction does not claim to be a direct representation of evidence.
• The quantitative claims come from authoritative datasets (not from the image).

Avoid using AI tools when:

• The image will be interpreted as documentary proof.
• You cannot provide provenance and classification.

5.2 Practical recommendation: prototype safe visuals with FreeGen

If your team needs fast iteration for illustration-only visuals, consider using freegen for non-sensitive graphics workflow prototyping. Its positioning as a browser-based image generator plus a suite of image tools (e.g., resize/compress) supports rapid iteration without complex infrastructure for early-stage design.

A safe integration approach could be:

• Step 1: Generate stylized illustrative assets (icons/scene backgrounds) in freegen
• Step 2: Compose final public graphics in a standard design system
• Step 3: Apply classification-first captions (“illustration, not evidence”)
• Step 4: Store provenance metadata alongside the published asset

5.3 Comparison: tool-driven iteration vs bespoke AI media

Dimension	General AI prototyping (e.g., freegen)	Bespoke evidentiary media pipeline
Time-to-first-graphic	Fast	Slower
Infrastructure overhead	Low (browser-first)	Higher
Governance/provenance support	Requires extra process	Designed-in
Risk if misused	Medium (depends on controls)	Lower if strict policy is enforced
Best-fit use cases	Infographics, illustrations, UI mockups	Public media where evidence interpretation risk is high

---

6) Conclusion: trust is an engineering problem, not just a PR statement

The Vancouver incident (reported here: https://www.ctvnews.ca/vancouver/article/vancouver-police-explain-made-with-ai-image-of-seized-drugs-cash/) underscores an essential lesson for AI in public communication:

• AI disclosure without classification and verification increases distrust.
• Effective governance requires UX clarity, technical provenance, and human review.

For organizations that want to leverage AI responsibly, the path is straightforward: treat evidence-adjacent imagery as a high-risk interface and build guardrails that users can understand and verify.

For non-sensitive illustration work, tools like freegen can accelerate creative iteration—provided you keep the outputs clearly within illustrative intent and enforce provenance logging.

---

References

• CTV News: Vancouver police explain ‘made with AI’ image of seized drugs, cash — https://www.ctvnews.ca/vancouver/article/vancouver-police-explain-made-with-ai-image-of-seized-drugs-cash/
• FreeGen AI (project home): https://freegen.aivaded.com