We Can’t Uninvent AI. So Why Aren’t We Building Better Defences?

Recently, I watched a demonstration of Griffin, a new AI model from a company called Tavus.

It was extraordinary yet unsettling. A convincing human face, natural conversation, expressions, interruptions. The sort of thing that would have been science fiction only a few years ago.

Some reactions were predictably alarmed. Technology this convincing should be banned said one guy on X, another even suggested the creators of Tavus should be sued already.

I understand the fear. I don’t entirely agree with the proposed solution.

I don’t think we’re looking at it right.

Rather than asking how we stop AI from becoming increasingly convincing, why aren't we asking how technology could make it increasingly difficult to deceive us? Don’t try to ban it because that won’t work but develop the defence against it.

Imagine being able to point your phone at an interview, a political speech, or a video call and immediately receive a credible assessment of its authenticity.

Not a guess based on whether somebody has six fingers. A serious verification system, built with the same sophistication as the technology it’s designed to scrutinise.

And before anyone misunderstands me, I think what Tavus has achieved is impressive even if deep down I suspect it will disrupt yet again entry level jobs.

There are potentially enormous benefits to having AI systems capable of interacting with people in ways that feel genuinely natural. Education, healthcare, and customer service in developing countries are just a few areas where AI can drive real value.

But consider another scenario.

Imagine an employee receiving a video call from their CEO. The face is familiar. The voice is unmistakable. The conversation feels completely natural.

The CEO asks them to authorise an urgent payment or share confidential information.

Except it isn't the CEO.

When a familiar face and voice can be convincingly recreated, the authority we normally attach to them becomes a vulnerability.

But stronger verification procedures, clearer authorisation processes, and employees who feel comfortable questioning instructions, even when those instructions apparently come from the most senior person in the company should be present.

And that raises another uncomfortable question. In organisations where challenging a senior executive is already considered risky, how many employees would dare question an instruction delivered by a perfect imitation of one?

An urgent request shouldn’t become legitimate simply because it arrives through a familiar face or voice.

Nor should an organisation’s credibility depend entirely on its ability to react after a convincing fabrication has already circulated.

Protecting communication, information, and the integrity of decision-making is becoming as important as protecting the systems that store them.

This is something I believe should be part of every serious organisational discussion about adopting AI.

Not simply how much more productive it can make us, but how we protect ourselves against the new vulnerabilities it creates.

Building our defences

I’m not suggesting that AI should develop without regulation.

There are applications that clearly require legal boundaries, accountability, and, in some cases, outright prohibition (don’t get me started on the bot farms that have all but destroyed our ability to have a healthy debate on a social media platform).

But restrictions alone cannot guarantee protection against a capability that is already spreading around the world.

We need technological countermeasures, organisational preparedness, and a much more serious investment in verifying what is real. There are entrepreneurial opportunities here, to build the solutions required to counteract the bad stuff.

However, the most sustainable path lies in trusted provenance and independent identity verification, not just detection.

As synthetic media capabilities advance, these tools will struggle to keep pace, and the arms race will only widen. Detection alone is not a long-term solution.

Instead, we should focus on building systems that make deception harder, not just easier to detect. This means embedding trust infrastructure into the core of communication, decision-making, and identity.

We need systems that can provide reliable evidence of where information came from, whether it has been altered and who is responsible for it.

There’s very little political will to slow this stuff down, so we need to ramp up protections.

We've become very good at building machines capable of imitating reality.

Perhaps it's time we showed the same ambition in building technologies that can help us establish what's real.

And for leaders, that means preparing their organisations to question what they see and hear, rather than simply trusting the familiar face delivering the message.

If AI is becoming more convincing by the month, our ability to verify it needs to become more sophisticated by the day.