For generations, sight has been treated as a reliable form of evidence. If we saw an event happen, watched it on video or heard someone speak, we felt we had a reasonable basis for believing it.
That assumption is now under serious pressure.
AI can generate convincing images, videos, voices and written messages that never existed in the physical world. It can also alter genuine material so subtly that even trained experts may struggle to identify what has changed.
We are moving from an age in which evidence could be seen to an age in which evidence must be verified.
The end of seeing is believing
Deepfakes are AI-generated or manipulated images, video and audio designed to appear authentic. Synthetic media is the wider category, also including AI-generated text, voices, documents and entirely artificial people.
The scale is increasing quickly. According to the UK Government, an estimated 8 million deepfakes were shared in the UK during 2025, compared with around 500,000 in 2023.
That does not mean every example was used for fraud. Some were created for entertainment, satire or experimentation. However, the same tools can be used to:
- Clone a person’s voice and request an urgent payment
- Create a video of a public figure appearing to endorse a false claim
- Alter a genuine recording to change its meaning
- Produce convincing identity documents or profile photographs
- Generate professional-looking emails that imitate a known colleague
- Fabricate evidence that appears to show an event that never happened
The technology does not need to be perfect. It only needs to be convincing for long enough to influence a decision.
The UK Government’s Technology Secretary, Liz Kendall, described the wider problem clearly:
“Deepfakes are being weaponised by criminals to defraud the public, exploit women and girls, and undermine trust in what we see and hear.”
That final phrase matters. The damage is not limited to people who believe a fake. It also affects people who begin to doubt genuine information.
Why our instincts fail
Most of us believe we would notice if a video or voice recording had been manipulated. We look for unnatural movements, odd lighting, distorted hands, unusual blinking or a voice that sounds slightly wrong.
Sometimes those clues are present. Increasingly, they are not.
Research from Veriff’s Deepfakes Report 2026 UK found that 74% of Britons were familiar with the term “deepfake”, yet their ability to distinguish genuine content from manipulated content was close to random guessing.
The same report found that:
- 81% were concerned about AI-related fraud and impersonation
- 78% were concerned about deepfakes eroding trust online
- 44% felt confident in their ability to identify a deepfake
- 22% did not try to verify suspicious content they encountered online
This creates a dangerous mismatch between confidence and capability. People may recognise the risk in theory while still trusting their own visual judgement in a real situation.
The problem is not a lack of intelligence. Human perception evolved to interpret faces, voices and body language in everyday social settings. It did not evolve to assess media produced by systems that can reproduce those signals at scale.
A convincing fake can also exploit our expectations. If a message appears to come from a senior manager, arrives during a busy afternoon and refers to a genuine project, our brain tends to process it as familiar rather than suspicious.
The more realistic the content becomes, the less useful instinct alone becomes.
The trust spillover effect
One of the most unsettling consequences of synthetic media is that realistic fakes can make us distrust genuine material.
Research published at ACM CHI 2026 found that participants struggled to distinguish authentic videos from AI-manipulated footage and rarely used detection tools.
A separate study found that exposure to realistic AI videos increased doubt and reduced confidence when participants later viewed authentic videos. This is sometimes described as a trust spillover effect.
The effect works in both directions:
- A fake video may be accepted as real
- A real video may be dismissed as fake
- Genuine photographs may be treated as uncertain
- Authentic recordings may be rejected without investigation
- People may retreat into believing only what supports their existing opinions
This connects to the idea of the liar’s dividend. Once deepfakes become widely known, someone caught on genuine video can simply claim that the footage was generated by AI.
That claim does not need to be true to create doubt. It only needs to introduce enough uncertainty to delay accountability or divide public opinion.
The result is not just deception. It is epistemic confusion, where the usual methods we use to establish what happened become less dependable.
The business reality
For organisations, this is not only a social media or political problem. Synthetic deception is already relevant to routine business activity.
Consider a few ordinary scenarios:
- A finance employee receives a voice message apparently from a director asking for an urgent transfer
- A supplier’s email says its bank details have changed
- A video call appears to show a colleague approving a payment
- A customer submits a synthetic identity during an account application
- An AI-generated email imitates the tone and formatting of a trusted contact
- A criminal uses information from a company website to make an impersonation attempt sound credible
The latest UK Finance Annual Fraud Report recorded £1.28 billion in gross fraud losses in 2025, alongside 4.1 million cases. Within that total, Authorised Push Payment, APP, fraud accounted for £576.4 million. Its data also shows that businesses remain exposed to invoice and mandate scams, where criminals redirect a legitimate payment to an account they control.
The report found that 66% of APP fraud cases originated through online sources, with a further 17% enabled through telecommunications. These figures cover more than deepfakes, but they demonstrate why online identity, email and telephone communications must be treated as connected parts of the same risk environment.
The appropriate response is not to distrust every message. It is to make important actions depend on repeatable verification processes rather than visual confidence.
Verification should become a business discipline
A practical verification process does not need to be complicated. It does need to be consistent.
Verify through a known channel
Do not use the telephone number, email address or link contained in an unexpected message. Contact the person or organisation through a number already held in your records or published on an official website.
A cloned voice can sound familiar. A callback to a trusted number creates a separate point of confirmation.
Introduce callback protocols
For payments, access requests and changes to supplier details, establish a rule that someone must confirm the request independently.
For example:
- Pause the request
- Find the contact details in an existing system
- Call the person using a known number
- Confirm the amount, purpose and destination
- Record who approved the action
This is particularly important when a message creates urgency or secrecy.
Separate approval from execution
Where practical, the person requesting a payment should not be the only person approving and sending it. A second person should review the transaction using information obtained independently.
Good controls do not depend on one person spotting a fake. They reduce the consequences when a fake looks convincing.
Check provenance where available
Content provenance systems such as C2PA can provide information about where media came from and how it has been edited. These systems are useful, but they are not magic proof. Missing provenance does not automatically mean content is false, and the presence of metadata does not remove the need for judgement.
Detection tools should be treated as supporting evidence, not as a final verdict.
Slow down emotionally charged requests
Fraud attempts often rely on pressure. The request may involve an urgent payment, a confidential acquisition or an apparent emergency involving a colleague or family member.
A short delay can be protective. Ask:
- Is this request unusual for the person involved?
- Why must it happen immediately?
- Can the information be confirmed elsewhere?
- What would happen if we waited 15 minutes?
- Is the message asking us to bypass an established process?
A genuine urgent request should withstand a sensible verification check.
Human judgement still matters
It may seem that the answer to synthetic media is simply better AI detection. Technical tools will help, particularly when analysing large volumes of content or identifying patterns that people cannot see.
However, technical detection has limits. Generative systems change quickly, detection tools can disagree and a piece of content may be partly authentic and partly manipulated.
Human judgement remains essential because verification involves context.
A person can ask whether:
- The request fits the organisation’s normal behaviour
- The sender would usually use that channel
- The timing makes sense
- The financial details match existing records
- The content conflicts with other reliable information
- Someone may be trying to exploit trust between colleagues
At ABC Service, our local IT and security work includes the practical side of this challenge, from monitoring and security awareness to keeping people involved in important decisions. Our managed IT services reflect a wider principle that technology works best when people understand how to use it safely.
The goal is not to remove human beings from the process. It is to give them clear procedures, sensible safeguards and enough time to make a considered decision.
A new relationship with evidence
Which? recently tested whether members of the public could distinguish AI-generated videos from genuine footage. Its May 2026 investigation found that 70% of participants failed to correctly identify all the real and fake videos shown to them.
The lesson is not that people are helpless. It is that the old shortcut of “I can tell when something looks wrong” is no longer reliable enough.
We need a more mature relationship with digital evidence:
- Seeing something should prompt questions, not automatic belief
- Not seeing obvious flaws should not be treated as proof of authenticity
- A familiar voice should not replace a callback
- A professional-looking email should not replace an approval process
- A detection tool should inform judgement, not replace it
- Genuine content should be supported by independent context and corroboration
The takeaway
AI is not only changing what can be fabricated. It is changing how we interpret everything that appears on a screen.
Deepfakes, synthetic voices and AI-generated text make deception easier to scale. They also create a second problem, because people may begin to reject genuine evidence simply because fake evidence exists.
The future of trust will depend less on what we can see and more on how carefully we verify it.
For individuals and organisations, the most valuable habit is simple: pause, use a known channel, confirm important requests independently and keep human judgement involved.
If seeing is no longer enough to believe, what verification habit should we introduce before we trust the next message, image, video or voice recording we receive?