Deepfakes: Protecting Business Identity and Communications

A deepfake is audio, video or visual material created or modified with AI techniques to produce a synthetic representation. For a business, the important problem arises when that representation is used to imitate someone, attribute a message to management or obtain an unauthorised operation. Verification needs to address both the material and the requested action.

A familiar voice is not approval within a system. A convincing video does not establish authority to change a record. Equally, synthetic material used legitimately in a presentation is not automatically fraud. Separate how content was produced, the identity of the communicator and their authority to request a specific action. Each calls for different evidence.

This guide proposes a verification protocol for teams and applications; it does not describe a WebHat customer's incident. Diagrams are original. Photographs illustrate security and infrastructure; decorative screens in the device photograph do not represent an actual deepfake detector or a tool we have deployed. The proposed workflow is separate from the photographic context.

What is a deepfake and why does it matter for business identity?

Imitating a representation does not confer a person's identity or permissions

The definition used in the NSA's 2023 information release concerns multimedia generated or manipulated through machine learning or deep learning. That release also discusses impersonated leadership and fraudulent communications targeting organisations. A simple editing cut and a synthesised voice are different mechanisms, even though either may be used to support a misleading claim. The business must examine the intended effect.

Within the types of artificial intelligence, content generation is a technical function. Its place in a commercial workflow depends on purpose, usage rights and presentation. Assessing a message goes beyond whether pixels or sound are synthetic: establish what the recipient is being asked to do and whether the requester has authority for that operation.

In our teaching scenario, someone requests a change to an important field in a supplier record. The request may arrive through voice, video or text. Company policy requires confirmation by the responsible person and approval of the exact operation, however familiar the message appears. This separation also makes the procedure relevant to other forms of impersonation.

How are synthetic voices, images and videos produced?

Modality and the proposed alteration define the analysis required

Generative models can produce or alter visual and audio representations. Material may change a person's appearance, mouth movements, voice or apparent speaking context. Other assets are generated entirely. Analysis should specify which component is being checked instead of applying the same technical label to every edited file. Different forms of manipulation require different evidence and evaluation conditions.

Our guide to generative AI and classification models explains the difference between producing material and assessing it. A generator creates content while a detector classifies it under its method. Those functions have their own data and criteria; the availability of a generation tool does not imply perfect recognition of all its outputs by another system.

A file may combine authentic and synthetic elements. A real background may accompany fabricated speech, while a genuine recording may be taken out of context. Our proposed assessment records the exact component observed and its source. Failure to detect an alteration is not transformed into confirmation of every statement within the material or of the sender's identity.

Which business requests need a separate verification step?

Prioritise the requested effect, not just image or voice quality

Company procedure defines operations affecting data, access, publication or responsibility. Replacing a contact, activating access or publishing a statement in management's name has a concrete effect. The verification rule follows that effect and the role entitled to request it. It should not depend on whether a recipient personally finds the recording realistic or persuasive.

Pressure to act urgently, conceal a request from its responsible owner or use a new contact is a reason to check, not technical proof of a deepfake. In our proposed exercise, the team identifies the operation and pauses its effect until the required confirmation. Employees do not first need to prove the precise method used to fabricate a message.

For employee AI skills, this case is more useful than memorising visual defects. Operators need to know what to verify, which channel to use and where to refer the case. Decision ownership does not transfer to a colleague merely because that colleague recognises the requester's apparent voice or has previously seen similar media.

How do we verify a request through an independent channel?

Confirmation uses previously validated details, not an address supplied in the request

In its voice-impersonation guidance, the FTC recommends contacting the person through an independently known number. Apply that principle to a business process: reach the responsible owner using the company's validated records. Replying to the original message or calling a new number supplied within it does not create independent verification, even if the subsequent conversation appears consistent.

Confirmation specifies what changes, which record is affected and what effect will follow. Do not ask only whether the person sent a message. In our teaching scenario, the operator states the proposed field value and checks the owner's role. The application connects approval to exact values so that a later modification cannot silently reuse an earlier confirmation.

If the owner is unavailable, the procedure defines escalation and a pending state. Do not invent a substitute or exception on the spot. The team needs a practical route for genuine urgency; otherwise pressure shifts decisions outside the system. The exercise tests that path as well as the ideal situation in which every responsible person responds immediately.

The contact directory used for confirmation has its own owner and update procedure. A change requested through the suspicious message does not immediately alter that record. Our proposed scenario separately checks who may replace a contact and what evidence supports the change, preventing the supposedly independent verification from returning to the same requester through a newly supplied address.

How should a business use a deepfake detector?

Read a score within the evaluated tool's conditions and scope

NIST AI 100-4, published in 2024, distinguishes provenance, marking and synthetic-content detection. Detector evaluation depends on modality, test material and file processing. Specify those conditions and error types in the project rather than treating a score as a general certificate of personal identity. A useful technical result needs a defined relationship to the business decision.

For a particular tool, establish what material it accepts, what output it returns and which situations were evaluated. An image test does not automatically become an audio-call test. Checking an original file is not equivalent to checking it after resizing, recompression or transmission through another channel. The evaluation should reflect the material actually received within the proposed workflow.

A detector result can assist assessment, but the operational decision remains separately defined. A positive signal goes for review; a negative one does not remove authorisation requirements. Record false alarms and missed cases in the exercise using authorised demonstrative material. Do not publish people labelled as fraudsters solely because a classifier produced a particular score.

What do provenance and Content Credentials establish?

A signed file history differs from the truth of its claims

C2PA describes Content Credentials as cryptographically signed information about an asset's provenance. Validation concerns declared data, integrity and connection to the file under the applicable trust model. That differs from verifying every real-world statement or approving an operation in the business application. The meaning of a technical verification should remain clear to the person acting on it.

In our proposed workflow, retain the authorised original and available provenance information separately from distribution copies. If a platform or export modifies the file, check which evidence survives. A missing manifest does not itself establish synthetic origin; a present one requires actual validation rather than reliance on an isolated icon. The system must record what was checked.

For its own communications, a company can document approval, version and official publication channel. That record helps staff answer verification requests. Do not confuse technical signing with automatically obtaining rights to someone's image or voice; rights and purpose belong to the production process. Internal approval records and publicly verifiable provenance also serve different audiences and should be designed accordingly.

How do we prevent a persuasive request from becoming an unauthorised operation?

Controls are enforced in the system that changes the data

The application checks the user, role and affected record. For operations defined in the project, approval is tied to actual fields and values. If the proposal changes after verification, repeat the corresponding control. A video attachment or a statement that the director approved something does not automatically create technical permission within the destination system.

This is a concrete application of AI alignment in business applications: desired outcomes and action boundaries are explicit. Even an assistant that accurately summarises a request may use only permitted functions. Recognising content and executing a change are separate stages supported by different evidence. The interface should make that distinction visible to its operator.

For data protection in AI integration, also define which material may be sent to an analysis service. A suspicious recording may contain personal or internal information. Do not automatically distribute it to the entire team or a public tool merely because analysis might be interesting. Access and retention follow the verification purpose and the authorised workflow.

How can a business legitimately use synthetic voices and content?

Authorised production has a defined purpose, rights and clear presentation

Generation can support demonstrative material, creative variants or voice interfaces with rights and permissions appropriate to the use. The company defines who approves material, how it is presented and where it is distributed. Do not attribute an unapproved statement to someone or present a simulation as an actual customer result. Production decisions should retain that distinction through editing and publication.

An AI voice assistant for bookings and support has a defined operational role. Design service identity, information supplied and permitted actions together. Voice quality should not conceal interaction with a system or change responsibility for a commercial confirmation. The application still needs the controls appropriate to the action it offers.

Before publication, check text, asset rights, the approved version and channel context. An editing change can alter a statement's meaning without AI. The same editorial discipline supports synthetic content, authentic photographs and recordings. Document what was approved, not merely the software used to create the file, so that the team can later verify the published version.

What should we prepare for an attempt to impersonate the business?

Separate evidence retention, stopping the effect and official communications

The procedure designates an owner who receives reports and can pause the affected operation. Retain necessary information: the material or its reference, time, channel and observed request. Avoid unnecessary redistribution of a suspicious file, particularly when it contains internal data or could amplify a message wrongly attributed to the company. Evidence access should remain connected to the response task.

The owner checks whether an effect has already occurred: a changed record, granted access or published statement. External communication uses only information the company can confirm. Notify the relevant channel through its procedure; specific actions and responsibilities follow the verified situation rather than the initial assumption about who produced the material. Keep operational correction separate from unsupported attribution.

Internal records distinguish observation, interpretation and action. For example, an account sending a request is an observation; its sender's real identity requires verification. The owner records what was confirmed, which effect was stopped and who continues assessment, without turning a suspicion into a public allegation. This also makes handover between responsible teams more precise.

Within online promotion and business communications, consistent official channels and contact details help audiences find a verified explanation. A clarification should not unnecessarily repeat false material. The team knows where to publish, who approves the response and how to retain the version communicated. The communication plan supports the operational process rather than substituting for it.

How do we test the procedure without exposing people or customers?

Use a declared scenario, demonstrative data and participants informed under the exercise rules. An unusual request can be simulated without cloning a real person's voice or fabricating material about a customer. The purpose is to test decisions and confirmation paths, not to surprise colleagues with offensive content. Select materials for the control being evaluated.

Assess whether the operation remained paused until the appropriate check, whether an independent channel was used and whether approval covered exact values. Record cases where procedure is cumbersome or an owner is unavailable. A rule that cannot be applied in normal work needs revision; do not conceal that issue in the training report while declaring the process successful.

Repeat relevant cases after contact, role or integration changes. Proposed measures include confirmation time, recorded decisions and operations executed under approval, with definitions established for the exercise. Do not present an internal test as the general detection rate for internet deepfakes or proof that every future attempt will be identified. Measure the particular workflow that was actually tested.

How does WebHat approach verification within digital business workflows?

Start with operations that matter to the company: data changes, access and publication. Define roles, confirmation and necessary evidence, then examine how rules are enforced in existing systems. An authorisation problem is corrected in the component executing the operation rather than through an additional message displayed beside a button. The control belongs at the point where the effect is created.

Through custom business digitalisation, we can design proposal interfaces, approvals, checks and effect records. The specification separates implemented integration from functions needing further assessment. A detector, where relevant, has its own operating scope and test rather than replacing the entire process. Its result must be usable within the company's defined decision path.

For a practical discussion, prepare sensitive operations, approving roles, validated contact channels and the applications involved. Initial conversations do not require confidential recordings or real customer data. We can describe a demonstrative scenario and acceptance criteria before selecting technical components, keeping the first design conversation focused on the process and the evidence it needs.

Frequently asked questions about deepfakes and request verification

Is every AI-generated asset an attack against the business?

No. Purpose, rights and presentation matter. Declared and authorised synthetic material may have a legitimate use. Impersonation and unauthorised requests are assessed separately from the technology used to produce a file. The same business control can also be relevant when misleading material was created without AI.

Can a familiar voice replace approval?

Apparent voice recognition does not replace company procedure. Defined operations require confirmation by an authorised person through a validated channel and approval of exact values. Check the requested effect, not simply how convincing the message sounds or how familiar its language appears to the recipient.

May we execute a request if the detector raises no warning?

A detector result grants no application permissions. The request still passes identity, role and appropriate approval checks. Interpret its score within the tool's scope and test conditions, separately from the operational decision. This keeps technical assessment from silently becoming authorisation to alter records.

Do Content Credentials confirm a video's statement is true?

They supply verifiable information about provenance and integrity within the technical model used. Check the statement, context and personal authority separately. Missing credentials do not themselves prove synthetic origin, just as a present icon is not a substitute for validation of the underlying record.

Must we clone a voice to train the team?

The workflow can be tested using demonstrative messages and data without imitating a real person. Assess pausing the operation, independent confirmation, approval and outcome records. Materials and participants follow the agreed exercise rules, with the scenario chosen to test the required control rather than visual or audio spectacle.

What belongs in an initial internal procedure?

Operations requiring verification, authorised roles, validated contacts, confirmation steps, escalation and retained information. The team also needs to know what happens when an owner cannot be reached, before urgency pushes a decision outside procedure. Test that exception path alongside the standard process.