Are AI photos safe?

The Hidden Risks of AI-Generated Photos

Fast, cheap, and everywhere — AI-generated photos feel like a free win for anyone creating content on a deadline. But the safety picture behind that convenience is genuinely alarming, and the evidence suggests the tool you pick matters as much as the image you make.

Consider a peer-reviewed study that tested ten leading text-to-image generators with health-misinformation prompts. Overall, 52% of outputs were classified as potentially harmful to public health. More striking is the spread between platforms: ChatGPT's image generator produced harmful content at a rate of 0%, while Reve Image 2.0 hit 98%. In other words, some platforms have essentially no effective content moderation, and choosing a generator is itself a safety decision.

The risks extend well beyond misleading health images. Unit 42, Palo Alto Networks' threat intelligence team, has documented dozens of deepfake scam campaigns impersonating public figures in seven languages across at least nine countries, likely run by a single threat actor group. A "Deepfakes as a Service" ecosystem now sells tooling on forums and messaging platforms, with video creation typically costing $60 to $500. One company lost $25 million in early 2024 after a deepfake impersonation of its CFO on a video call.

Regulators see the same trajectory. The UK government assesses that generative AI "will almost certainly act as a force multiplier for safety and security risks" by proliferating and enhancing threat actor capabilities — and warns that watermarking and similar authentication solutions are currently unreliable, requiring updates as the technology evolves.

For business owners and teams adopting AI imagery, the practical takeaway is that safeguards vary dramatically, so vet before you adopt:

• Prefer generators with demonstrated content moderation for any public-facing content — the 0% to 98% spread in harmful-output rates shows moderation is not a given.
• Treat video- and image-based identity verification with suspicion, especially for financial requests, and confirm through a separate channel.
• Disclose AI use and adopt emerging content credentials — hiding AI's role can backfire if discovered.
• Add substantial human modification to AI images; purely AI-generated images likely carry no U.S. copyright protection following the Supreme Court's denial of cert in Thaler v. Perlmutter.

The same scrutiny applies to the workflows behind the images. If you use AI agents to generate, edit, or vet visuals, the skills powering those agents deserve vetting too — published research found roughly 26% of agent skills carry at least one vulnerability. SkillzBench is building a free directory and free MCP connector that security-scans, validates, and benchmarks agent skills, so you can equip your agents with vetted skills only. Join the free waitlist to be first in line when it launches — everything, from scanning to benchmarking, stays free.

Deepfake Fraud and the Erosion of Trust

A single video call. A familiar face. A $25 million wire transfer. In early 2024, a company's finance team sat through an entire video conference with what they believed was their chief financial officer and several colleagues — every one of them a deepfake. The employees transferred funds on the strength of what they saw, and the company lost $25 million, according to threat researchers at Unit 42.

That case sits at the extreme end of a rapidly industrializing problem. Unit 42 documented dozens of scam campaigns impersonating public figures in seven languages across at least nine countries — likely run by a single threat actor group — with a "Deepfakes as a Service" ecosystem now selling ready-made tooling on forums and messaging platforms. The economics are alarming: a convincing deepfake video typically costs just $60 to $500 to create, while face-swapping tools range from free to over $249 per month.

The scale of distribution is equally sobering. Deepfake scam domains tracked by Unit 42 averaged 114,000 global accesses each and stayed live for an average of 142 days — long enough to reach enormous audiences before takedown. The UK government's own assessment warns that generative AI will "almost certainly act as a force multiplier" for security risks, projecting that synthetic media could comprise a large proportion of online content by 2026.

Authentication technology hasn't kept pace. The same UK government assessment concludes that watermarking and similar authentication solutions are "currently unreliable, requiring updates as generative AI evolves." In other words, you cannot yet depend on a technical stamp to prove an image is real.

This creates a genuine dilemma for anyone publishing AI-generated photos. Research cited in a journalism ethics study found audiences are skeptical of content labeled as AI-generated — even when it's true, or even human-made — because they assume full automation. Yet practitioners point out the opposite risk: hiding AI involvement "can backfire if discovered," eroding trust precisely when you need it most.

For organizations, the practical response looks like this:

• Treat video- or image-based identity verification with suspicion, and confirm financial requests through an out-of-band channel.
• Disclose AI use in your own visual content rather than waiting to be caught — transparency beats exposure.
• Adopt emerging content credentials, such as CAI's embedded "nutrition labels," even though watermarking alone isn't yet dependable.
• Vet the AI tools and agent workflows in your pipeline, not just their outputs — published research found roughly 26% of agent skills carry at least one vulnerability, which is why SkillzBench is building a free, security-scanned directory of vetted skills.

The deeper cost of deepfake fraud isn't measured only in stolen millions. It's the slow erosion of a default we once enjoyed: believing that seeing is believing.

The Copyright Minefield: Can You Legally Use AI Photos?

If you generate an image with AI and a competitor uses it, can you stop them? Probably not — and that's only half the problem. The other half is whether the image itself infringes someone else's rights.

Start with ownership. Purely AI-generated images likely have no U.S. copyright protection at all, according to an attorney-curated legal guide. Human authorship is settled law as of March 2026, after the Supreme Court denied cert in Thaler v. Perlmutter. In practical terms, the AI photo you paid for and published can be freely copied by anyone, because copyright law requires a human creator.

Then look at what your tool actually grants you. The terms of service contain traps most business users never read:

• Free-tier Midjourney images carry no commercial rights — users are bound by a CC BY-NC 4.0 non-commercial license.
• Companies grossing over $1,000,000 per year must be on a Pro ($60/month) or Mega ($120/month) plan to use images commercially.
• Midjourney provides no IP indemnification, meaning if an image triggers a lawsuit, the legal bill is yours alone.
• Prompting copyrighted characters — Mickey Mouse, Darth Vader, Shrek — is explicitly high-risk behavior.

That last point isn't theoretical. Gary Marcus and Reid Southen's experiments, published in IEEE Spectrum, showed Midjourney and DALL-E 3 can produce images that appear to display copyrighted material — what the authors call a "visual plagiarism problem" and a "copyright minefield." Disney and Universal sued Midjourney in June 2025 over generation of their characters, with DreamWorks later joining the case. The outcome could reshape commercial-use norms for every AI image tool.

The money at stake is enormous. The Bartz v. Anthropic settlement in August 2025 totaled at least $1.5 billion, compensating roughly 500,000 eligible works — an expected payment above $3,000 per work. That settlement concerned training data, but it signals how courts and litigants now value copyrighted works in the AI pipeline. Output infringement claims could follow similar math.

So what can a business actually do? Add substantial human modification to AI images — human editing strengthens any copyright claim you have and reduces infringement risk. Verify licensing tiers and revenue thresholds before commercial use, and never claim you personally photographed or drew an AI-generated image. And scrutinize the whole AI workflow, not just the images: published research found roughly 26% of agent skills carry at least one vulnerability, which is why SkillzBench is building a free directory that security-scans, license-checks, and benchmarks agent skills before you connect them to your agents.

The legal landscape here is genuinely unresolved, and today's safe practice is tomorrow's court case. Treat AI photos as legally fragile assets — useful, but handle with documented care.

Ready to vet the AI skills behind your image workflows? The waitlist is live at SkillzBench — join free and be first in line when the directory, security scans, and benchmarks launch.

Best Practices for Using AI Photos Safely

The goal is not to avoid AI photos entirely — the UK government's own risk assessment warns against disproportionate anxiety that causes organizations to miss AI's benefits. It's to use them with managed care. The evidence points to a handful of practices that make the difference.

Start with tool choice, because it is itself a safety decision. A peer-reviewed study of 10 leading image generators found harmful-output rates ranging from 0% (ChatGPT) to 98% (Reve Image 2.0) when prompted with health misinformation themes. Prefer generators with demonstrated content moderation for anything public-facing.

Before commercial use, verify the legal fine print. Midjourney's terms bind free users to non-commercial rights, require Pro or Mega plans for companies grossing over $1M/year, and provide no IP indemnification. Confirm licensing tiers and revenue thresholds on your actual plan before shipping AI images in marketing or products.

On copyright, purely AI-generated images likely have no U.S. copyright protection, so add substantial human modification to strengthen your claims — and never prompt copyrighted characters, a behavior that drew Disney and Universal lawsuits against Midjourney in 2025.

• Choose generators with proven safeguards for public-facing content.
• Verify licensing terms and revenue thresholds before commercial use.
• Add meaningful human editing; avoid copyrighted characters in prompts.
• Disclose AI use with CAI content credentials — hiding it can backfire if discovered.
• Treat image- and video-based identity verification with suspicion.

That last point deserves emphasis. Unit 42 researchers documented a $25 million corporate loss from a single deepfake video-call impersonation, and deepfake creation costs as little as $60–$500. Train staff to distrust visual identity alone and use out-of-band confirmation for any financial request.

Finally, apply the same scrutiny to the AI workflows behind the images. Research on agent skills found roughly 26% carry at least one vulnerability — SkillzBench is building a free directory and MCP connector that security-scans and benchmarks skills so you can equip agents with vetted ones only. Managed use, applied consistently, turns AI photos from a gamble into a dependable creative asset.

Vet the AI Agents Behind Your AI Images

Vetting the tool that makes your AI images is only half the job. If you use AI agents to generate, edit, or review those images, the skills powering those agents need the same scrutiny — because a compromised skill can quietly undermine every safeguard you've put in place.

Agent skills are instruction packages that extend what an agent can do, from calling image APIs to post-processing outputs. They're easy to install and even easier to trust blindly. That's a problem: published research found that roughly 26% of agent skills carry at least one vulnerability, meaning nearly one in three skills you might install could expose your workflow to risk.

The stakes mirror the image-safety problem itself. A peer-reviewed study found harmful output rates ranging from 0% to 98% across leading image generators — proof that tool choice is a safety decision. The same logic applies one layer down: which skills your agent runs determines what it can access, what it sends out, and how it handles your images and data.

Before equipping an agent with any skill, check a few basics:

• Does the skill's source publish a pinned, verifiable version you can audit?
• Has it been security-scanned, not just eyeballed by other users?
• Is the license compatible with your commercial use of the outputs?
• Is it actively maintained, or abandoned code waiting to drift out of date?

Doing this manually for every skill is unrealistic, which is why SkillzBench is building a free directory that continuously discovers agent skills from GitHub, pins each version to a commit SHA, and runs license checks and security scans before a skill earns a 0–100 trust score. A free MCP connector will let you toggle exactly which vetted skills each agent can see — and nothing more. Benchmarking is also in development, testing each skill in a sandbox with the same model, with and without the skill, so you can see the measured Lift in percentage points rather than taking anyone's word for it.

The lesson from image safety applies here too: transparency and verification beat blind trust. Deepfake threat researchers note that as malicious usage grows, defenders' proactive detection efforts should grow with it. The same proactive posture belongs in your agent stack. Vet the images, and vet the skills behind them — join the SkillzBench waitlist to get notified when the free directory and MCP connector go live.