What happens when seeing someone’s face, hearing their voice, or knowing their personal information is no longer enough to prove who they are?

That is quickly becoming one of the defining security challenges of the AI era.

In this episode of Edly Spotlight, Burak Sahin; an independent advisor and global biometric expert with more than 25 years of experience in identity verification explains how AI is changing the rules of digital trust.

The biggest takeaway is simple: businesses can no longer treat identity verification as a single check.

A face can be generated. A voice can be cloned. A video feed can be manipulated. Personal information can be stolen. And increasingly, an AI agent may be acting on behalf of a real person without any obvious way to prove that it has permission to do so.

For organizations building financial products, digital services, AI agents, authentication systems, or customer onboarding flows, that creates a new question:

How do you establish trust when almost every traditional signal of identity can be replicated?

Sahin’s answer is not a single technology. It is a layered approach that combines biometrics, liveness detection, device security, provenance, transaction risk, and stronger authentication.

Here is what businesses can learn from his experience.

AI Has Changed What It Means to Verify Someone’s Identity

Biometric verification was once associated primarily with government, law enforcement, and high-security environments. Today, it is part of everyday digital life.

We unlock phones with our faces. Banks verify customers remotely. Companies onboard employees digitally. Airports compare travelers against identity documents. Financial institutions authenticate transactions without requiring people to visit a physical branch. That convenience has created enormous value But it has also created a much larger attack surface.

As Sahin explains, identity systems no longer need to answer only:

“Does this face match the person we have on record?”

They increasingly have to answer:

“Is this even a real person in front of the camera?”

Detecting whether what we’re seeing is actually a human, as opposed to a spoof or mimicking some other human’s characteristics, has become also part of the equation. That distinction matters because AI has made it dramatically easier to reproduce the signals organizations traditionally relied on for trust.

The implication for businesses is important: matching identity is no longer enough. You also need to verify authenticity.

Better AI Improved Biometrics And Made Attacks More Powerful

One of the most interesting parts of Sahin’s perspective is that the same technological progress responsible for stronger biometric systems also helped create their newest threats.

Between roughly 2014 and 2018, advances in deep learning dramatically improved facial recognition.

Face recognition moved closer to biometric modalities that had historically been considered more accurate, including fingerprints and iris recognition. That made facial verification far more practical for consumer products because practically every smartphone already had the necessary sensor: a camera.

But there was a tradeoff.

The machine learning techniques improving facial recognition were also making synthetic images, cloned voices, and manipulated video more convincing.

As Sahin puts it:

“The same technological underlying frameworks that pretty much propelled this advancement is, on the other side of the coin, also helping fraudsters pretty much produce viable material as easily.”

That creates an uncomfortable reality for security teams.

AI is improving both sides of the identity arms race. Organizations benefit from better recognition models, but attackers benefit from better generative models.

The competitive advantage therefore no longer comes simply from having an accurate face-matching algorithm. It comes from building a broader identity architecture around it.

Live Deepfakes Turn Video Calls Into an Identity Problem

For many people, “deepfake” still means a manipulated video uploaded to social media.

That mental model is already outdated. One of the biggest risks Sahin highlights is the ability to manipulate identity during a live interaction.

A person can join a video call while software alters their appearance to resemble somebody else. Their voice can also be modified or cloned.

The person on the other side may therefore believe they are speaking with a colleague, executive, customer, or authorized individual when they are actually communicating with someone entirely different.

Sahin explains:

“I can talk to you, but you would think I’m somebody else.”

For businesses, this means video itself can no longer automatically be treated as proof.

That matters for situations such as:

• approving financial transactions,

• confirming sensitive business instructions,

• remote employee onboarding,

• executive communications,

• customer support,

• account recovery,

• remote KYC,

• and high-value purchases.

The lesson is not that businesses should stop using video.

It is that visual familiarity is no longer an authentication method.

A recognizable face and voice may create confidence, but they should not create authorization.

The Bigger Threat Is Not Always the Deepfake Itself

Another useful distinction Sahin makes is between a deepfake and the mechanism used to deliver it.

Attackers do not necessarily have to fool a camera by physically presenting something fake in front of it. They can attack the technology stack itself.

One example is an injection attack.

Instead of allowing a phone or computer camera to provide the genuine live feed, an attacker can attempt to hijack the capture process and inject a different video stream.

That changes how organizations need to think about identity security.

It is not enough to ask:

“Does this image look real?”

Security teams also need to ask:

“Where did this image come from?”

That includes signals such as device integrity, capture method, source provenance, metadata, application permissions, and the path the media took before reaching the verification system.

Sahin summarizes the idea well:

“Identity is a big composite picture. We can’t be good at just one part of it. We have to be good at all parts of it.”

For organizations designing identity systems, this may be the most actionable lesson from the conversation. Trust should come from multiple independent signals not one highly accurate algorithm.

How Deepfake Detection Actually Works

If generative AI can create increasingly convincing fake media, how can organizations detect it?

One approach Sahin describes is surprisingly intuitive. To identify deepfakes, detection systems can be trained on both authentic content and artificially generated content.

Organizations do not have to wait for attackers to produce examples. They can generate deepfakes themselves, analyze the artifacts they create, and continuously improve their detection models.

The goal is similar to biometric recognition: teach the system to distinguish between categories with increasing reliability.

But visual analysis is only one part of the process.

A stronger system can combine:

• characteristics inside the video,

• signs of manipulation,

• device-level security information,

• media provenance,

• capture metadata,

• application behavior,

• and other contextual signals.

This means businesses should think beyond deepfake detection software as a standalone tool.

The better question is:

How many independent reasons does our system have to trust this interaction?

The more trustworthy signals agree with each other, the harder it becomes for an attacker to successfully fake the entire identity chain.

Liveness Detection Is Becoming Essential for Remote Identity Verification

Deepfake detection sits within a broader challenge: determining whether the person interacting with a system is genuinely present. This is where liveness detection, also known in standards terminology as presentation attack detection, becomes important.

A biometric system may correctly determine that an image resembles a customer but organizations increasingly need another layer that asks whether the biometric sample was captured from a real, present person rather than from manipulated or replayed media.

That distinction is particularly important for non-proctored digital experiences.

Think about remote banking. There may be no employee physically watching the customer. The entire identity decision may happen through a phone.

If the system is going to approve an account, transaction, application, or identity claim automatically, it needs greater confidence that the biometric data it receives is genuine.

For product and security teams, the benefit of strong liveness detection is therefore not simply “better fraud prevention.”

It enables organizations to safely move more identity processes online without requiring expensive manual verification.

The Next Identity Challenge Is Not Human; It Is AI Agents

The most forward-looking part of the conversation comes when Sahin turns from humans pretending to be other humans to AI agents acting on behalf of humans.

AI agents are increasingly capable of performing tasks such as:

• booking flights,

• scheduling appointments,

• making purchases,

• interacting with customer service,

• comparing financial products,

• and completing multi-step workflows.

That creates a new kind of identity problem.

An AI agent can know plenty about you. It may know your name, address, birthday, travel preferences, payment information, calendar, or account details. But knowing information about you does not prove it has permission to act for you.

Identity has traditionally been built around three categories:

What you know? What you have? What you are?

An AI agent may have access to the first two.

The difficult part is establishing a trustworthy connection to the third.

That means the future of AI-agent security may depend heavily on delegated authorization.

An organization needs confidence not only that an agent knows who you are, but that you explicitly authorized that agent to perform a particular action.

Risk-Based Authentication Could Make AI Agents Safer

Not every action performed by an AI agent requires the same level of verification.

That gives organizations an opportunity to apply risk-based authentication rather than making every interaction equally difficult.

If an AI agent books a free event ticket, additional identity verification may provide little value. But imagine that same agent is about to purchase a $5,000 non-refundable airline ticket.

That transaction carries far greater risk. Before completing it, the system could send a notification to the user asking for biometric confirmation.

The AI agent handles the workflow. The human authorizes the consequential action. That model could become increasingly important as businesses deploy autonomous agents. The goal is not to force humans back into every step.

It is to put human verification at the moments where mistakes or unauthorized actions would be costly.

For companies building agentic AI products, this creates a useful design principle:

Automate the workflow. Authenticate the risk.

Why Security Questions Are Becoming Less Useful

AI is not the only reason identity systems need stronger authentication. Traditional knowledge-based authentication has been weakening for years.

Security questions such as your birth date, former address, family information, or other personal details rely on the assumption that those facts are secret.

Increasingly, they are not.

Large data breaches, social media, public databases, and stolen identity records have made personal information far easier to obtain.

Sahin notes that standards organizations have been cautioning against excessive reliance on knowledge-based authentication for years.

That matters even more in the AI era. An AI-powered attacker can potentially aggregate publicly available and leaked information far faster than a human attacker manually researching a target.

So if an organization’s identity strategy still depends heavily on “something only the real person should know,” it may be relying on an assumption that no longer holds.

What National Digital Identity Systems Teach Us About Scale

Sahin’s perspective is shaped by work on identity systems that operate at an unusually large scale. He has been involved with biometric identity initiatives connected to countries including Mexico, India, and Indonesia.

These systems illustrate why digital identity is not simply a cybersecurity feature.

It can also become infrastructure. A robust identity system can help governments deliver services to citizens who previously struggled to prove who they were. Biometrics can reduce duplication, simplify access to benefits, and help services reach people living far from physical government offices. The same principles are increasingly relevant to private companies.

Whether an identity platform needs to support ten thousand users or hundreds of millions, organizations eventually confront similar questions:

• How accurate must the system be?
• How should false matches be handled?
• How much friction will users tolerate?
• Which biometric modalities make sense?
• What happens when verification happens remotely?
• How should liveness be incorporated?
• And how does the system respond as fraud techniques evolve?

There is rarely one universally correct answer. The architecture has to reflect the risk of the transaction and the consequences of getting identity wrong.

What Businesses Should Take Away

AI is not making digital identity irrelevant. It is making strong digital identity more valuable.

For organizations building AI products, financial platforms, authentication systems, marketplaces, government services, or digital onboarding experiences, several lessons stand out:

• Do not rely on a face or voice alone. Generative AI makes both increasingly reproducible.

• Treat identity as a collection of signals. Biometrics, liveness, device integrity, provenance, behavioral signals, and transaction context work better together.

• Design authentication around risk. A low-value action should not require the same friction as a high-value, irreversible transaction.

• Prepare for AI agents as identity actors. Organizations will increasingly need to verify not only who a customer is, but whether software has permission to act for them.

• Move beyond knowledge-based authentication. Personal information is becoming easier for both human attackers and AI systems to obtain.

• Build security that can evolve. Deepfake techniques will improve, so identity architecture needs to accommodate new detection and authentication layers rather than depend on one permanent defense.

The bigger lesson is that the question is changing.

For years, digital identity systems asked:

“Are you the person you claim to be?”

Now they increasingly need to ask:

“Are you real, are you authorized, can we trust the channel you are using and if an AI agent is acting for you, did you actually give it permission?”

That is a much harder problem. But it is also an opportunity.

Organizations that build strong identity infrastructure now will be better positioned to automate more processes, deploy AI agents more confidently, reduce fraud, and create digital experiences that customers can actually trust.

Have a Story to Share?

If you have insights, experiences, or a story worth sharing, send us an email to be considered as our next Edly Spotlight guest.

And if you’re looking to build or scale your eLearning ecosystem, talk to our experts at Edly. We know how to create secure, scalable learning environments that keep your data protected and your training ready to grow.

Share Your Story | Talk to an Edly Expert

Watch the Full Interview

Watch the full conversation with Burak Sahin to explore how deepfakes, biometric authentication, digital identity verification, liveness detection, and AI-agent security are converging and what organizations need to do to prepare.