Image Authentication Should be a Condition of AI Decisions

Mounir Ibrahim is chief communications officer and head of public affairs for Truepic and a former Foreign Service officer with the United States Department of State.

Eight years ago, I wrote that emerging technology used to authenticate images could help hold war criminals like Bashar al-Assad, then president of Syria, accountable. As a United States diplomat on the ground in Syria, I saw why verifying the authenticity of images mattered: people making consequential decisions needed evidence grounded in reality. 

Nearly a decade later, it is clear that AI is playing a growing role in decisions once made solely by humans. Whether we like it or not, we are living in a world in which AI systems make, or significantly influence, decisions ranging from drone strikes in a war zone to insurance claims about a car accident. Many of those decisions depend on visual evidence.

Helen of Troy showed us that a face could launch a thousand ships. Today, I worry the person behind the face may not exist at all.

Decisions of consequence must be grounded in reality. We cannot accept a “garbage in, garbage out” approach to AI decisioning. Such systems need to operate on authentic and verifiable visuals. Otherwise, we risk scaling fraud and chaos without any tether back to reality.

A growing number of companies — including Truepic — are making image authentication technology available across smartphones, digital cameras, software, and editing platforms, with an open standard supporting interoperability. Authentication is not a silver bullet, but when built into the capture process, it verifies origin, critical metadata and whether an image has been altered, giving AI a firmer basis for judgment. Though it will take time for authenticity to be ubiquitous, we need to start discussing why AI must run on authentic visual inputs.

Let's start with digitization of processes that were once human-run. This space is where visual fraud, deception and miscalculation live. Automation systems need to confidently identify authentic visuals if businesses and even governments are to rely on them. This is not the case today.

Meanwhile, bad actors leverage easy access to hyper-realistic visual creation and flood the very systems that need authenticity with deception. Highlighting this dynamic, a technologist, Nikita Bier, recently shared a satirical post instructing an AI agent to conduct widespread return fraud against hundreds of companies with a single prompt. This fraud is happening today.

Consider the insurance world, where a Verisk report found that 98% of insurers agree that AI-powered editing tools are fueling a rise in digital media fraud. In the same report, 99% of insurers encountered manipulated or AI-altered documentation, yet only 32% feel very confident in detecting deepfakes. Similar flags are being raised across financial industries and even OpenAI’s own Sam Altman warned of the coming AI-fraud crisis.

Governments are raising alarm bells too. The government of South Korea just warned that AI is driving increased fraud, particularly in insurance. In the US, a recent RFI from the Consumer Product Safety Commission (CPSC) began inquiring about the presence of “altered, fabricated, or artificially generated images,” which can undermine governmental efforts to remove hazardous products from circulation.

Despite these warning signs, AI deployments in critical systems and services are increasingly being used for decision-making. Gartner found that by 2028, 32% of organizations expect to use AI to support decisions, and 27% expect AI to make decisions primarily without human intervention. Agentic capabilities will likely fuel the autonomous nature of business with less human oversight in the loop. McKinsey's State of AI survey found that the largest organizations have scaled use of agentic AI from 27% to 40% in less than one year.

Some may suggest that AI will be smart enough to identify visual deception, but I remain skeptical. Interoperable markers, like C2PA content credentials, are now being applied to some AI outputs as companies work to respond to transparency requirements under laws like the EU AI Act and California’s AI Transparency Act. Bad actors, however, can use models that leave no such signals, or strip out safeguards entirely with free tools such as “Heretic,” whose “decensored” models have been downloaded more than 13 million times. Further, so-called ‘cheapfakes’ — basic non-AI image edits, time, date or location changes — remain a major problem that AI cannot reliably detect. If AI is going to play a greater role in consequential decisions, maintaining an auditable and verifiable connection to reality should become a basic condition of its deployment.

As we move toward AI-driven decision-making, companies, industries, and governments setting the rules must ensure the evidence informing those decisions is verifiably authentic.

This is not hypothetical. The technologies described already exist on consumer devices and software platforms used by businesses and governments today. The next step is to ensure authenticated inputs for all decisions of consequence, especially when AI is the decider.